Layer Selectivity: When American States Court Artificial-Intelligence Semiconductor Fabs While Rationing Hyperscale Datacenter Incentives — How Arizona’s $265 Billion TSMC Expansion, Governor Katie Hobbs’ Three-Year Tax-Exemption Pause, Rival State Regimes, and the Democratic Congress of 2027 Could Redefine Industrial Investment Across the Five-Layer AI Economy Through 2030

Introduction: Two Announcements, One State, Two Economic Judgments

Imagine the closing hours of an Arizona tax-incentive application window in the last two weeks of June 2026. For datacenter developers weighing sites across metropolitan Phoenix, timing had suddenly become an economic variable of the first importance. On June 13, Governor Katie Hobbs had signed a bipartisan budget whose taxation omnibus, HB 4168, enacted as Laws 2026, Chapter 140, instructed the Arizona Commerce Authority to stop accepting applications from new computer datacenters for the state’s longstanding sales-and-use-tax exemption from July 1, 2026, through June 30, 2029. The clock would stop at the end of June. According to figures the Commerce Authority subsequently released and that Arizona newsrooms reported in July and again in early November, 113 applications arrived between June 15 and June 30, a volume almost equal to the 123 applications the agency had received in the entire thirteen years since the program’s creation in August 2013; by November the agency confirmed that 107 of those late applications had been approved.^[1][2][3] An application, of course, did not guarantee certification, and certification did not prove that a developer would ultimately pour concrete, energize a substation, or fill a hall with accelerators. But the rush demonstrated, with unusual clarity, how much a change in fiscal eligibility could matter to decisions about sites, servers, and construction schedules — and how a policy intended to save the state money could, in its first fortnight, bring forward the very behavior it was designed to restrain.

Less than a month later, the camera turned to a different kind of industrial announcement. On July 16, 2026, during its second-quarter earnings conference, Taiwan Semiconductor Manufacturing Company disclosed plans for another $100 billion of investment in Arizona, including four additional fabs using 2-nanometer or more advanced processes, together with additional advanced-packaging capacity. The City of Phoenix and the Arizona Commerce Authority characterized the combined plan as $265 billion across ten fabs, two advanced-packaging facilities, and a research and development center, spread over a campus of more than 1,100 acres in north Phoenix.^[4][5][6] Governor Hobbs publicly welcomed the expansion in the strongest terms available to a sitting governor.

“TSMC’s historic investment firmly establishes Arizona as the nation’s epicenter for advanced semiconductor manufacturing and innovation. With America’s most advanced chip-making technologies, world-class education partners, and a fast-growing workforce, Arizona is poised to lead the future of AI and next-generation technologies.”

— Governor Katie Hobbs, Office of the Governor of Arizona, July 16, 2026 ^[4]

The distinction in chronology matters and should be kept intact throughout this paper. TSMC’s first Arizona commitment, in 2020, was described as a roughly $12 billion single-fab project; subsequent revisions brought the three-fab plan to $65 billion; in March 2025 the company announced an additional $100 billion, bringing its stated total to $165 billion; and in July 2026 the company announced another $100 billion, bringing the stated total to $265 billion.^[7][8] These announced totals are future investment commitments and plans, not a claim that $265 billion has already been spent. Indeed, TSMC’s own capital-expenditure disclosures put the entire company’s 2026 budget at $60 billion to $64 billion across all of its sites in Taiwan, Japan, Germany, and Arizona, with $40.64 billion actually spent through September.^[9] The Arizona figure is therefore a multi-decade intention, not an annual flow, and a disciplined reader should treat it accordingly.

For a casual observer, the juxtaposition looks like a contradiction. One side of the emerging AI system received public congratulations and a powerful political welcome; another lost access to a longstanding state tax exemption for new applicants. Was Arizona enthusiastic about artificial intelligence or newly skeptical of it? Did officials want the chips but not the facilities that run them? Did a state advertise global technological ambition in July after imposing a local restraint in June? The questions appear paradoxical only if artificial intelligence is treated as one undifferentiated industry. For public finance, the semiconductor factory and the hyperscale computing campus are not the same transaction, and they have never been the same transaction; what is new is that the scale of both has grown so large, so quickly, that the difference between them has become a matter of state budgets, household electricity bills, and regional water security rather than a technicality for economic-development professionals.

A fab is an extraordinarily complex manufacturing installation, employing thousands of technicians and engineers, buying specialty materials and equipment, creating supplier relationships, and generating products that move through a national or global market. A hyperscale datacenter may be equally expensive to equip, and its construction can support substantial employment, but its recurring local labor requirements, electrical load, tax treatment, and relationship to the utility network differ. Both activities consume large quantities of water and electricity. Both may require infrastructure that strains nearby communities. Their public benefits and risks must therefore be measured rather than inferred from the prestige of the words AI, cloud, or advanced manufacturing. The economic contrast is sharper than the marketing language, and the policy response that is emerging across the United States is beginning to reflect it.

November added a democratic judgment to the fiscal one. Katie Hobbs won reelection on November 3, 2026, defeating Representative Andy Biggs by roughly thirteen points, and Arizona Democrats captured both chambers of the state legislature for the first time since 1966.^[10][11] A newly elected Congress will convene in January 2027 with Democratic majorities in both the House of Representatives and the Senate — the Senate flipped by at least seven seats, including Texas, Ohio, Iowa, and Kansas, and the House by the largest Democratic margin in nearly two decades — while a Republican president remains in the White House.^[12][13] Six governorships changed party, and candidates who campaigned on datacenter moratoria or on making datacenters “pay their own way” won in Texas, Ohio, and Georgia.^[14][15] That combination is unlikely to produce a single coherent national answer to every AI infrastructure question. Congressional appropriations, federal export controls, state incentives, utility regulation, and municipal siting decisions operate through different legal authorities and political incentives. A federal executive order can accelerate certain federal reviews without automatically overriding a city zoning decision or guaranteeing a state tax exemption.^[16] The practical geography of AI investment is about to become more intricate, not less.

I choose the title Layer Selectivity because it names the apparent Arizona paradox as a systematic policy choice rather than a rhetorical contradiction. The state is not deciding whether it supports AI in the abstract. It is assessing — and changing the public terms available to — different activities along a connected production system. The term is meant to capture a new industrial-policy question: which layer produces sufficient public value to justify a particular financial or regulatory concession, and under which measurable conditions?

I also choose Layer Selectivity because the Five-Layer AI Economy makes that decision visible across the entire chain, not merely inside a semiconductor or datacenter debate. Energy determines physical carrying capacity; chips embody industrial capability; datacenters convert power and silicon into compute; models transform compute into intelligence; and applications and agents translate intelligence into work. A government can court one layer, scrutinize another, and still depend on all five. The title consequently supports a forward-looking inquiry into how governors, Congress, corporations, and communities may allocate investment and obligations through 2030, and it supplies a vocabulary that is portable from Phoenix to Richmond, Austin, Harrisburg, Columbus, Sacramento, and Atlanta.

The remainder of the paper proceeds in six sections. Section 1 defines the Five-Layer AI Economy and the concept of Layer Selectivity, and proposes a public-value test. Section 2 examines Arizona in detail as the paradigm case. Section 3 compares the divergent regimes emerging in Virginia, Texas, Pennsylvania, Ohio, Indiana, Michigan, California, Georgia, and New York. Section 4 turns to the corporations — TSMC, NVIDIA, AMD, Intel, Micron, Amazon, Google, Microsoft, Meta, OpenAI, and Anthropic — and reads their third-quarter 2026 disclosures through the lens of the framework. Section 5 analyzes the new Congress, the White House, and three scenarios for 2027–2030, and proposes a Five-Layer Public Investment Scorecard. Section 6 distills eight pillars of what has been learned, and the Conclusion explains why the term is likely to describe the next phase of American AI industrial policy.

Section 1: From One AI Industry to Five Distinct Public Bargains

1.1. The Five-Layer AI Economy as an industrial map

The Five-Layer AI Economy begins with energy, proceeds through chips and datacenters, reaches model development, and culminates in applications and autonomous agents. It is useful because it asks a physical and financial question before an ideological one: what must each stage obtain from the economy, what does it deliver, and where do its benefits and costs fall? The layers are interdependent but not interchangeable. A kilowatt-hour consumed in a fabrication plant helps make a chip; the same kilowatt-hour consumed in a computing facility helps operate a chip. Those services can be economically complementary while competing for power-system capacity, for skilled electricians, for industrial water, and for the patience of the communities that host them.

Layer 1 — Energy — includes generation, fuel, transmission, substations, distribution, water for power operations, and the institutional arrangements determining who bears construction and reliability costs. It is geographically rigid. A utility cannot simply move its transmission network to a lower-tax state, and a regional transmission organization such as PJM or ERCOT cannot relocate its interconnection queue. Layer 2 — Chips — includes semiconductor design, foundry fabrication, memory, advanced packaging, testing, equipment, and material supply. Some intellectual-property assets are relatively mobile, but a fab is an immovable, multiyear industrial commitment whose economics depend on process yields that are themselves the product of years of accumulated engineering learning. Layer 3 — Datacenters — combines land, buildings, cooling, backup power, networking, accelerators, storage, and high-density operations. It is partly mobile: a model training job might be assigned to another region, but latency, customer proximity, data-residency requirements, power availability, and sunk construction costs constrain that flexibility.

Layer 4 — Models — is organized around frontier training, post-training, inference architecture, evaluation, safety, and intellectual property. Much of its output is digital, but it depends on large installations in Layer 3 and scarce chip and networking capacity in Layer 2. Stanford’s 2026 AI Index reports that global AI compute capacity has grown roughly 3.3-fold per year since 2022, reaching about 17.1 million H100-equivalents, with NVIDIA accounting for more than 60 percent of the total — a statistic that belongs to Layer 2 and Layer 3 simultaneously and that would be meaningless without Layer 1.^[17] Layer 5 — Applications and Agents — comprises enterprise workflows, consumer products, robotics, software, decision assistance, and increasingly persistent agentic systems. Layer 5 is where public gains from productivity, accessibility, healthcare, science, and new businesses become apparent. It is also where privacy, employment, product safety, and accountability questions become concrete. A model company might employ workers in California, rent accelerator capacity in Texas, and serve a manufacturer in Michigan. The tax base, electricity cost, and social return are not always in the same state, and they are rarely in the same county.

This separation is foundational to Layer Selectivity. A state that builds a fab is not automatically guaranteed the datacenter load, model research, or application employment that may rely on its chips. The economy has increasingly separable locations for design, manufacture, deployment, and use. The political bargain must therefore be assessed at the site of the relevant obligation, not assumed to follow the path of a semiconductor through the entire national system. The Brookings Institution’s work on the geography of the AI economy makes the same point from the other direction: AI talent and startup formation remain concentrated in a small number of “superstar” metropolitan areas, and the places that host the physical infrastructure are frequently not the places that capture the high-wage design and application employment.^[18]

“Geography matters. There is significant dispersion of AI activity into new and different metros across the country. Our report allows for an important new look at where we are on AI.”

— Mark Muro, Senior Fellow, Brookings Metro ^[18]

1.2. What Layer Selectivity is — and what it is not

Layer Selectivity is a jurisdiction’s deliberate preference for investment in some parts of the AI production system while using different fiscal, environmental, or regulatory standards for others. It may operate through tax credits, exemptions, grants, procurement, energy tariffs, siting requirements, labor standards, water rules, permitting timetables, or publicly financed infrastructure. It may be explicit — a budget denying new datacenter exemptions while preserving semiconductor recruitment — or implicit, produced by agencies with different mandates acting on different calendars. The defining characteristic is unequal treatment of different layers, accompanied by an explanation of why their public returns or external costs differ.

It is not synonymous with an anti-datacenter policy. Arizona’s 2026 measure restricts eligibility for a particular tax benefit for new datacenters during a specified period; it does not ban datacenter construction, revoke every existing certification, or prohibit a privately financed facility that meets other applicable requirements. The Arizona Commerce Authority itself emphasized that many of the late-June applications were for additional facilities or future phases at already-certified campuses, and that existing certifications — which under the 2013 statute can run for ten years at new sites and twenty years at qualifying redevelopment sites, through 2033 — were not terminated.^[3][19] Nor does a warm official reception to TSMC mean that its fab expansion is exempt from environmental regulation, municipal permitting, utility charges, or federally imposed conditions. A proper analysis must distinguish fiscal preference, regulatory permission, infrastructure access, and political endorsement. Conflating these categories produces misleading headlines and unsound investment analysis.

The concept also differs from a policy of choosing corporate winners. In principle, Layer Selectivity can be technology-neutral within each layer. A state might apply the same large-load tariff to a Google datacenter and an independent colocation provider, or the same measurable water-reuse standard to rival semiconductor makers. In practice, large firms possess political relationships, technical staff, financial guarantees, and site-development capabilities that smaller firms lack. Policy design must ask whether differentiated requirements reflect real public costs or simply become a preference for incumbents. The Harvard Electricity Law Initiative’s warning about a regulatory “race to the bottom” — in which utilities compete for hyperscale load by offering discounted contracts whose costs are shifted to captive ratepayers — is precisely a warning about selectivity gone wrong, where the layer that is favored is favored because of its bargaining power rather than its public value.^[20]

1.3. Why the same public dollar has a different meaning in each layer

The first economic reason for differentiation is the structure of locally retained value. A fab’s capital spending can support construction jobs, recurring manufacturing positions, process-engineering careers, specialty suppliers, and a cluster of material and equipment services. A datacenter can bring property taxes, construction activity, specialized operations positions, digital-service access, and demand that finances new power infrastructure. But a large proportion of its capital expenditure flows toward accelerators, networking systems, and other equipment produced elsewhere — much of it, ironically, fabricated by TSMC and packaged in Taiwan — so that the local multiplier of a datacenter dollar is structurally different from the local multiplier of a fab dollar. The job profile is not identical, and the permanent local staffing of a completed facility may be much smaller than the temporary workforce that constructed it.

Virginia’s nonpartisan Joint Legislative Audit and Review Commission provides the most carefully documented caution in the American literature. Its December 2024 study estimated that the datacenter industry supported 74,000 jobs, $5.5 billion in labor income, and $9.1 billion in state GDP annually — but it broke those figures down with unusual honesty: roughly 59,000 of the jobs, or 80 percent, were construction-phase positions, and only about 15,000 were ongoing operations jobs, of which about 4,400 were direct datacenter employees. It also described a representative 250,000-square-foot datacenter as employing around 50 full-time workers, about half of them contractors, while peak construction at an individual building could involve roughly 1,500 workers.^[21][22] These are Virginia-specific analytical estimates and illustrative industry descriptions — not employment multipliers that can be mechanically transferred to every proposed Arizona project. They prove that datacenters create real value while underscoring the difference between temporary building activity and permanent operational employment. Good Jobs First, a subsidy watchdog that has tracked datacenter deals for more than a decade, draws the sharper conclusion that per-job subsidy costs in datacenter “megadeals” have averaged roughly $2 million and in the most extreme recent case — a Stream Data Centers project at New York’s STAMP site — approached $11 million per permanent position.^[23][24]

“The data center industry does not need any public financial support.”

— Kasia Tarczynska, Senior Research Analyst, Good Jobs First ^[1]

A second reason is the counterfactual. A state should not award a tax preference simply because a corporation announces that it will spend billions. It should ask what would happen without the preference. Would the facility relocate, shrink, delay, or proceed anyway? Would the state gain a genuinely scarce capability? Would the investor capture the fiscal concession as an increase in land price or project margin? Does the site impose a new cost on households or existing manufacturers that exceeds the local tax receipts? Different answers are plausible for fabs, datacenters, model laboratories, and application startups. Good Jobs First’s own site-selection research concludes that for datacenters “it’s not the subsidies, but availability of land and cost of power that remain the most important factors,” which, if correct, implies that much of the exemption is non-additional — a transfer rather than an inducement.^[24] The semiconductor counterfactual is different in kind: Chris Miller of Tufts University’s Fletcher School, the author of Chip War, has argued that TSMC’s international competitiveness was never built on cost efficiency, and that the strategic rationale for a domestic leading-edge fab rests on the catastrophic consequences of losing access to Taiwanese production rather than on any ordinary return calculation.^[25][26]

“Unlike in the Cold War, where the government was trying to make missiles fly more accurately, the civilian use of chips has become so big and the economic impact is so large that the loss of chip production in Taiwan would create such a massive manufacturing depression in the US that it would have strategic ramifications similar to wartime.”

— Chris Miller, Associate Professor of International History, The Fletcher School, Tufts University ^[26]

The third reason is duration. Semiconductor manufacturing can anchor equipment and workforce relationships over several technology generations, but it also carries enormous obsolescence, process-yield, and market risks. A datacenter can operate for decades, but accelerators within it may be replaced several times, and load could migrate if efficiency or customer demand changes. Energy infrastructure may last even longer. A rational public agreement should match the term of the benefit to the useful life of the induced local investment, the risk of stranded facilities, and the government’s ability to recover costs. Arizona’s original datacenter statute, which granted exemptions for up to ten years on new sites and twenty on redevelopment, was in effect a duration bet made in 2013 on a technology whose power density and capital intensity have since changed by an order of magnitude; the 2026 pause is, among other things, a recognition that the original term structure no longer matched the asset.

1.4. Scarcity is not a single variable

Power dominates many AI infrastructure debates, but megawatts are not equivalent to water rights, skilled electricians, land-use entitlements, or political consent. A site might have sufficient electrical energy on an annual basis but inadequate transmission capacity during summer peaks. Another site might have an available substation but uncertain long-term groundwater. A semiconductor fab may need highly reliable power, ultrapure process water, chemicals, cleanroom maintenance, and specific technical skills. A datacenter may emphasize high-density power delivery, liquid cooling, fiber connectivity, backup systems, and accelerated commissioning. Both compete for scarce construction trades, and in Phoenix in particular both compete for the same pool of pipefitters and electricians.

The Lawrence Berkeley National Laboratory’s United States Data Center Energy Usage Report: 2025 Update, published for the Department of Energy in June 2026, projects a central estimate of 11.8 percent of U.S. electricity consumption in 2030, or about 649 terawatt-hours, with a scenario range of 9.5 percent to 15.3 percent, up from 192 terawatt-hours and 4.7 percent in 2024.^[27][28] These are national modeled outcomes, not a prediction that every locality will experience the same growth; indeed, the same report found that 2024 consumption came in at the low end of the laboratory’s prior expectations, and that datacenters account for only about a third of projected national load growth through 2030. The International Energy Agency’s global outlook moves in the same direction: datacenter electricity demand grew 17 percent in 2025 while AI-focused facilities grew about 50 percent, and the agency projects total datacenter consumption roughly doubling from about 485 terawatt-hours to about 950 terawatt-hours by 2030 — approximately 3 percent of global demand — with AI-specific demand tripling.^[29][30] The significance of these projections is the scale of planning uncertainty: utility investments cannot be sized responsibly on optimistic corporate load requests alone, and an individual state’s incentives may import obligations that emerge years after the tax decision.

“Global electricity demand from data centres is set to more than double over the next five years, consuming as much electricity by 2030 as the whole of Japan does today. In the United States, data centres are on course to account for almost half of the growth in electricity demand.”

— International Energy Agency, Energy and AI, World Energy Outlook Special Report ^[31]

Energy planning therefore requires at least four distinct quantities: connected or requested capacity in megawatts, actual coincident peak demand, annual consumption in megawatt-hours, and the incremental cost of firm reliability. A 100-megawatt nameplate load running at 85 percent average utilization would consume approximately 744,600 megawatt-hours annually — not the same as drawing 100 megawatts at every instant. Cooling technology, maintenance, power-management controls, and backup configurations all influence the result. Counting announced datacenter gigawatts as if they were simultaneous actual consumption exaggerates the immediate system requirement; ignoring their prospective coincidence understates future risk. ERCOT’s experience in 2026 is the cautionary tale: by midsummer the Texas grid operator was carrying more than 1,800 large-load interconnection requests totaling roughly 474 to 498 gigawatts — approximately 90 percent of them datacenters, and several times the entire system’s peak demand — of which only about 65 gigawatts across 150 requests were preliminarily eligible as “base load” under the new Batch Zero screening.^[32][33] A queue is not a forecast, and a forecast built on a queue is a recipe for stranded transmission.

1.5. A proposed public-value test

Layer Selectivity should be disciplined by a five-part public-value test, rather than a ranking that automatically places fabs above datacenters. First, durable employment and skills: how much ongoing local income and transferable worker capability does the project create, net of the workers it draws away from other essential projects? Second, fiscal return: what are discounted future taxes and fees after incentives, government services, and public infrastructure costs? Third, resource burden: what power, water, land, air, and reliability obligations remain with other users, and are they measured rather than asserted? Fourth, strategic capability: how much does the project improve domestic manufacturing, research, resiliency, or competitive access, and is that value national rather than local? Fifth, reversibility and accountability: what occurs if the project shrinks or fails, and what enforceable protections survive?

The public-value test can be expressed conceptually rather than disguised as a universal formula:

Net public value = discounted local economic and strategic benefits − discounted subsidies, resource externalities, infrastructure costs, and failure risk.

Every term requires assumptions and distributional analysis. The strategic value of domestic leading-edge chip capacity is real but hard to price. Some datacenter benefits are national or global rather than local. A government that counts national AI productivity in its benefit column while leaving local ratepayers to fund the grid may exaggerate its return. A government that ignores the benefits of computing, regional tax revenue, and construction may understate it. The objective is to illuminate tradeoffs, not to manufacture a false numerical precision.

Layer Selectivity becomes defensible only when different treatment can withstand these tests. Otherwise the concept risks degenerating into politically fashionable support for one class of investor and arbitrary hostility toward another — a risk that is not hypothetical when, as Gallup found in March 2026, seven in ten Americans oppose the construction of an AI datacenter in their own community, a level of opposition that exceeds even the 53 percent who oppose a nearby nuclear power plant.^[34] This is the paper’s first core argument: a connected technological stack does not imply an identical public bargain at every layer, and the variation in bargains must be justified by evidence rather than by sentiment in either direction.

Section 2: Arizona’s Deliberate Asymmetry — A $265 Billion Chip Vision and a Three-Year Fiscal Pause

2.1. How Arizona became an advanced-manufacturing anchor

The Phoenix semiconductor story did not begin in July 2026. TSMC’s first Arizona commitment in 2020 was initially described as a $12 billion factory project. Subsequent revisions brought the first three-fab plan to $65 billion. In March 2025, the company announced an additional $100 billion in intended U.S. investment, encompassing three more fabrication facilities, two advanced-packaging plants, and a research center, raising the total to $165 billion; TSMC’s own release described the expansion as “the largest single foreign direct investment in U.S. history” and projected 40,000 construction jobs over four years and tens of thousands of permanent high-technology positions.^[7] Then, at its July 16, 2026, earnings event, TSMC announced another $100 billion, including four additional advanced fabs. Chairman and Chief Executive C.C. Wei was explicit about what the money would build.

“This is to build several or more semiconductor logic wafer fabs for 2-nanometer and below technologies as well as advanced packaging fabs.”

— C.C. Wei, Chairman and CEO, TSMC, second-quarter 2026 earnings conference, July 16, 2026 ^[35]

Arizona’s economic-development agency accordingly described a cumulative plan of $265 billion, ten fabs, two packaging facilities, and one research center, and the City of Phoenix counted twelve facilities in all.^[4] It is particularly important to keep this chronological accounting intact: the two $100 billion expansions are separate announcements, but all components belong within the $265 billion cumulative total. The July 2026 announcement also arrived roughly two weeks after a U.S.–Taiwan trade agreement that reduced tariffs on Taiwanese goods to 15 percent in exchange for about $250 billion in planned Taiwanese investment in the United States, which means that the Arizona figure is as much an artifact of bilateral trade diplomacy as of Phoenix’s local attractions.^[6]

One fab is already a substantial technological accomplishment. TSMC’s first Arizona fab began high-volume production of N4-family chips in late 2024, with process yields the company describes as comparable to its Taiwanese operations; Apple announced in February 2026 that it would purchase more than 100 million chips manufactured at TSMC Arizona during 2026; and by the time of the company’s third-quarter 2026 earnings call on October 15, TSMC had completed construction of its second Phoenix fab and begun early construction on a fourth fab and on its first Arizona advanced-packaging facility.^[8][36] These details do not mean that the ten-fab vision will become ten simultaneously operating installations by 2030. They mean that a credible leading-edge manufacturer has chosen Arizona as a multigeneration investment location, with actual output dependent on construction, equipment, engineering, customer orders, permits, labor and visa availability, and commercial returns.

Arizona’s achievement extends beyond one company’s corporate announcement. Equipment manufacturers, specialized maintenance providers, materials suppliers, universities, and community colleges have developed around the cluster. The Arizona Commerce Authority reported a $35.5 million expansion connected to the University of Arizona’s semiconductor nano-fabrication and training capabilities in February 2026, and it operates Future48 workforce accelerators, including semiconductor technician training developed with local colleges and industry participants.^[37][38] The SEMICON West trade show, the industry’s principal North American gathering, will move to a spring schedule in Phoenix beginning in 2027, a symbolic relocation of the industry’s center of gravity.^[39] Local economists have begun to speak of the cluster in macroeconomic rather than project terms: with Arizona’s nominal gross state product at roughly $620 billion, a single company’s $265 billion plan together with Intel’s roughly $32 billion Chandler expansion represents committed private investment equal to about half of one year’s state output.^[40] These programs reveal why states may judge fab investments as catalysts for a broader industrial base rather than as isolated structures with large price tags.

The national implications are equally significant. TSMC is the critical production partner for the companies designing the most advanced processors and accelerators; its third-quarter 2026 results showed high-performance computing, overwhelmingly AI-related, accounting for 68 percent of revenue, and 2-nanometer products already contributing 8 percent of wafer revenue only three quarters after volume production began.^[41] A domestic manufacturing location does not eliminate reliance on foreign equipment, specialty chemicals, packaging capacity, or global supply networks. But it can diversify fabrication geography, develop domestic operational expertise, and create additional options for selected products. These national-security benefits extend beyond Arizona; they help explain why federal support may be justified even if a particular state’s narrow fiscal return is less impressive than the advertised capital commitment. The crucial analytical distinction is between national strategic value and state-level budget value. A project may provide both, but they must be evaluated separately, and the party that captures each should be the party that pays for it.

2.2. The legal anatomy of the 2026 datacenter pause

Arizona’s June 2026 omnibus tax legislation, HB 4168 (enacted as Laws 2026, Chapter 140, and signed June 13), contains the decisive provision. From July 1, 2026, through June 30, 2029, the Arizona Commerce Authority may not accept applications for any new computer datacenter under the relevant certification statute, and new facilities do not qualify for the associated tax relief.^[42][43] The preexisting Computer Data Center Program, created in 2013, offers qualifying owners, operators, and certain tenants exemptions from Arizona transaction privilege tax and use tax on qualifying equipment, including applicable state, county, and local components, and the certification process determines whether the statutory exemption applies.^[44] The fiscal stakes had grown quickly: according to the Arizona Department of Revenue’s tax-expenditure reporting, the exemption cost the state $19.4 million in fiscal 2024 and $38.5 million in fiscal 2025 — a 98 percent increase in a single year, and a trajectory that, if the 107 newly approved applications mature into equipment purchases, could continue steeply upward despite the pause.^[43]

The phrase tax-exemption moratorium is accurate; construction moratorium would not be. A datacenter may still be proposed, financed, permitted, and constructed in Arizona during the three-year period, subject to otherwise applicable laws. Previously issued certifications and applications received before the cutoff require separate legal assessment; the statute’s prospective language does not support an assumption that every already-certified project suddenly loses its existing benefit, and the Commerce Authority’s approval of 107 late applications confirms that the pre-July pipeline remains eligible.^[1] Nor is a state tax exemption the same as a negotiated utility rate, zoning clearance, building permit, groundwater authorization, or federal environmental determination.

The same legislative session produced a second, less-noticed instrument that belongs to a different layer. HB 2756, signed June 4, 2026, requires large public power entities and electric public service corporations to report periodically on new “extra high load factor customers” — the tariff category under which Arizona Public Service serves datacenters and which APS proposed in its 2025 rate case to raise by roughly 45 percent — and authorizes cost-of-service studies showing how the costs of serving such customers are allocated.^[45][46] Where HB 4168 operates on Layer 3 through fiscal eligibility, HB 2756 operates on Layer 1 through disclosure and cost-causation. Together they illustrate that Layer Selectivity in practice is rarely a single statute; it is a bundle of instruments adopted by different committees for different reasons, which happen to converge on the same class of facility.

The distinction matters for financial modeling. Consider two otherwise comparable projects, one with a valid certification obtained before the statutory deadline and one beginning its application process afterward. Their equipment acquisition plans may yield materially different state-and-local tax outcomes, even if both ultimately build and use similar electricity. The later project’s sponsors may alter procurement timing, choose a different campus, negotiate lower land costs, raise additional equity, or accept a lower return. But none of those responses is inevitable. Demand, sunk costs, energy contracts, customer geography, tax interpretation, and the availability of alternative states will determine the actual result. That is Layer Selectivity operating through fiscal eligibility rather than physical prohibition.

Governor Hobbs did not present the budget as an attempt to end Arizona’s technology industry. Her administration framed the pause as part of a responsible path that protects water and household utility costs, and it sits alongside a broader energy strategy. In September 2025 the governor had created, by executive order, the Arizona Energy Promise Taskforce, chaired by the director of the Governor’s Office of Resiliency, with an explicit mandate to “facilitate data center and other large load customer growth” while minimizing impacts on ratepayers; when asked whether the order was about datacenters, Hobbs replied, “This is not just about data centers. This is about the high energy demand that we’re seeing across industry that is growing our economy, creating jobs for Arizonans.”^[47] The taskforce’s April 2026 report issued 31 recommendations to meet a projected 40 percent increase in peak demand, prioritized virtual power plants and distributed solar, and warned that “large-load growth is occurring across utility service territories and impacting infrastructure systems, creating a need for greater coordination.”^[48][49] State policy therefore placed the datacenter preference inside a broader contest over fiscal resources, grid capacity, and local impacts rather than treating the exemption as an entitlement that should expand indefinitely.

2.3. The last-minute application surge is evidence — but of what?

The reporting that 113 applications were submitted during the final sixteen days of the June window, and that 107 were approved, is one of the most revealing pieces of the story. The flow is consistent with a strong incentive to preserve prospective tax eligibility before rules changed. It also contains, by the Commerce Authority’s own account, a large share of applications for expansions of existing facilities or future phases of development rather than wholly new projects, together with some that are speculative, early-stage, or subject to later material revision — one applicant, a land-consulting firm, described plans for a $25 billion development near Tonopah that exists, for now, as a plan.^[50][51] The applications should not be multiplied by an assumed gigawatt size and reported as a verified new load pipeline. They are expressions of interest in a fiscal position, not certified operating megawatts or proof of completed construction.

The episode illustrates a subtle policy-design risk that the economics literature on anticipated regulation has long recognized. An announced future restriction can bring forward the very activity it intends to scrutinize. Firms may file in advance, accumulate contingent approvals, or accelerate equipment decisions to preserve options. In this case the statute controlled applications for a particular incentive, not demand for physical interconnection. Still, the general lesson applies: implementation dates can matter as much as policy rates. A lobbyist for the national Data Center Coalition argued that the number of applications “isn’t as significant as it sounds,” because most were extensions of existing projects; a critic in the legislature countered that the state had received 136 percent more applications in two weeks than in the prior thirteen years combined.^[2][50] Both observations can be true at once, which is exactly why the raw count is a poor policy metric. If a future governor wants to tighten large-load conditions, a transitional rule should account for strategic applications, incomplete plans, information quality, and good-faith investors that already incurred substantial costs under the prior regime.

It would be tempting to describe the rush as proof that Arizona’s datacenter policy was previously too generous. It is not conclusive evidence of that proposition. A rational developer will seek a lawful tax benefit even if the benefit merely offsets extraordinary construction costs. The surge does, however, reveal that fiscal treatment was salient enough to change behavior on a short timetable, and the doubling of the exemption’s annual cost between fiscal 2024 and fiscal 2025 reveals that the benefit was no longer marginal to the state’s finances. Joseph Palomino of the Arizona Center for Economic Progress put the policy question plainly: the issue is whether the projects would happen without the tax break.^[1] For researchers and lawmakers, that is a reason to publish application, certification, projected investment, actual investment, and energization data in clearly separated categories, so that the 2029 review can be conducted on evidence rather than anecdote.

2.4. Semiconductor water is still water

Arizona’s preference for semiconductor manufacturing must not be romanticized. A fab uses substantial water in cleaning, chemical processing, and facility operations. It requires reliable power and a dense industrial supply chain. Phoenix faces long-term groundwater constraints, and the Arizona Department of Water Resources maintains an extensive management system in the Phoenix Active Management Area, whose groundwater modeling it updated in 2026.^[52] A state cannot rigorously defend manufacturing incentives on water-security grounds while declining to measure industrial water consequences simply because the investment has strategic prestige.

TSMC has outlined a meaningful mitigation strategy, and it has begun to build it. The company’s Arizona operation currently recycles about 65 percent of its water through an in-house water resource center that supplies cooling towers and air scrubbers; in late summer 2025 it broke ground on a 15-acre Industrial Reclamation Water Plant, scheduled to operate in 2028, designed to return industrial wastewater to the ultrapure standard required for chipmaking, to reach an 85 percent recycling rate at startup and 90 percent or better thereafter, and to approach “near-zero liquid discharge.”^[53][54] TSMC’s environmental review disclosed a development agreement with the City of Phoenix for access to 11.4 million gallons of water per day, and the company has acknowledged that a revised agreement would need to account for regional water demands as the campus expands.^[53]

“One of the many reasons we chose Phoenix for our US operation was the state’s and city’s thoughtful and thorough planning, including a water supply plan for 100 years. We know that water is a constant concern in the region. We will be a responsible corporate neighbor in our use of natural resources, including water.”

— Rose Castanares, President, TSMC Arizona ^[54]

Those figures are the company’s reported operating configuration and future engineering objectives, respectively; they should not be presented as third-party-verified results from a plant that will not operate until 2028. The useful research question is net freshwater withdrawal, quality of reclaimed water, seasonal supply security, brine and chemical management, and enforceable performance — not simply the percentage appearing in a corporate presentation. The United Nations University’s June 2026 report on the environmental cost of AI’s energy use is a reminder that water accounting must extend beyond the fence line: it estimated that the water footprint associated with the projected electricity consumption of global datacenters in 2030 could reach 9.3 trillion liters, most of it embedded in power generation rather than drawn at the site.^[55]

Datacenters also exhibit substantial variation. Some rely on water-intensive evaporative cooling; others adopt closed-loop liquid systems or designs that trade lower onsite water consumption for different energy requirements. A location’s climate and the utility’s generation mix influence total water footprints. Policymakers should distinguish onsite consumption from offsite water used to generate purchased electricity, and withdrawal from consumption. California’s September 2026 legislation, discussed in Section 3, is the first state statute to require datacenter water-supply assessments and scarcity plans as a condition of permitting; Arizona’s 2029 review could borrow the instrument. A fab’s ultrapure process-water requirement is not equivalent to a datacenter’s cooling-water profile, but neither can be evaluated by slogans. Transparent measurement is a condition of credible Layer Selectivity.

2.5. The power-system bargain beneath the tax bargain

The Salt River Project, one of metropolitan Phoenix’s principal electricity providers, illustrates why energy charges can be more important than an upfront equipment exemption. SRP says it plans for datacenter growth within an integrated system framework and seeks to prevent new infrastructure costs from being shifted to residential or other customers; its published pricing descriptions say large commercial and industrial customers, including datacenters, must pay associated infrastructure costs, and its 2027 Integrated System Plan process is already underway.^[56][57] Arizona Public Service, the state’s largest investor-owned utility, has gone further in public: it says it developed a high-load-factor rate in 2017 precisely so that datacenters’ rates would reflect their cost of service, and in its June 2025 rate request it proposed a roughly 45 percent increase for the extra-high-load-factor class “so that high load factor users will cover their increasing costs and to ensure residential customers are not subsidizing data centers.”^[46]

This yields a constructive way to separate issues. A state could decide that every new datacenter should pay ordinary transaction taxes on equipment and also cover genuinely incremental transmission, substation, and reliability costs. Alternatively, it could offer narrowly conditioned exemptions to projects with verified grid benefits or unusually strong local employment. Those are distinct policy choices. A blanket exemption is easy to administer but can over-subsidize projects that would arrive anyway. A complicated set of discretionary awards risks delays, favoritism, and uncertainty. Good design balances transparency, administrability, investment predictability, and fairness.

Consider an illustrative facility with 100 MW of demand operating at an average 85 percent load factor. Its annual electricity consumption would be about 744,600 MWh. At a hypothetical incremental infrastructure-related charge of one cent per kilowatt-hour, annual collections would be approximately $7.45 million. This is not an Arizona tariff, a projected obligation of any named company, or a claim about what the grid upgrade costs. It demonstrates the scale: over a decade, recurring electricity-related obligations can approach or exceed the economic importance of equipment-purchase taxes. Policymakers should avoid comparing a one-time exemption with recurring system burdens without discounting both over a consistent horizon.

The empirical literature on who actually pays is maturing quickly, and it cuts in more than one direction. A 2025 analysis by Carnegie Mellon University and North Carolina State University projected that datacenter and cryptocurrency load growth could raise average U.S. electricity bills by about 8 percent by 2030 and by as much as 25 percent in Virginia; a May 2026 update by the same North Carolina State team, published in Environmental Research Letters, raised the regional upper bound to 57 percent.^[58][59] A CEPR discussion paper using utility-territory data found that each additional datacenter raised average residential prices by about 0.182 cents per kilowatt-hour, with a ten-facility increase associated with roughly 13 percent higher residential prices relative to the post-2022 mean, and with households bearing increases 30 to 40 percent larger than commercial customers.^[60] Against this, a Rutgers policy-lab study of the 2014–2024 period found no robust evidence that datacenter openings had raised local residential bills before the hyperscale era, and Berkeley Lab’s own 2026 analysis notes that some utilities project that datacenters could eventually lower residential bills by spreading fixed costs across more customers.^[61][27] The reconciliation is institutional rather than physical: as the Rutgers authors put it, “electricity bills are shaped by institutions, not just physics.” Whether a datacenter raises or lowers a neighbor’s bill depends on the tariff, the contract, and the regulator — which is exactly why HB 2756’s cost-of-service studies are the right companion to HB 4168’s fiscal pause.

A further possibility, advanced by energy scholars at the University of Southern California among others, is that the Layer 1–Layer 3 bargain changes character altogether if large computing campuses begin to arrive with their own firm generation. Shon Hiatt of the USC Marshall School of Business, whose research documents the unprecedented increase in electricity and cooling-water consumption produced by GPU-driven datacenters, has argued that the most plausible durable solution after 2030 is for datacenters to bring clean, reliable power with them — in his view most likely through small nuclear reactors — rather than to compete with households for the existing grid.^[62] If that vision materializes, the public-value test does not disappear; it migrates from the utility tariff to the siting, safety, and water terms of the generation itself. FERC’s 2026 show-cause orders on co-location, discussed in Section 3, are the first systematic attempt to write rules for exactly that configuration.

“I see a future in which data centers bring their own clean and reliable energy with them, through reactors.” (translated from Portuguese)

— Shon Hiatt, Professor, USC Marshall School of Business, São Paulo School of Advanced Science in Systems Change and Sustainability ^[62]

2.6. Workforce infrastructure as a form of selectivity

Tax treatment receives the most attention, but workforce investment may be the more durable industrial-policy instrument. A state can help a chip manufacturer recruit by subsidizing building costs; it can also improve the labor supply shared by the manufacturer and dozens of other firms. The latter can create spillovers if training credentials are portable and programs are not limited to one employer. Arizona’s Future48 initiatives and university facilities therefore deserve as much analytical attention as the eye-catching investment figure. They are investments in an ecosystem, not just a construction site. The Arizona Building and Construction Trades Council has estimated that roughly 12,000 construction-trade workers could be needed to build out the TSMC fabs alone, a number that is itself larger than the permanent operating workforce of several dozen hyperscale datacenters combined.^[40]

The other side of the bargain is local opportunity cost. Electricians, pipefitters, mechanical contractors, and project managers are required for fabs, datacenters, substations, housing, hospitals, and public infrastructure. When several megaprojects start together, skilled labor can become scarce and wages or scheduling costs rise. Training programs are valuable if they expand the workforce rather than merely recruit workers away from other essential projects. Policymakers should monitor apprenticeship completion, wages after placement, employer retention, and housing affordability, not just enrollment in promotional training initiatives. Tom’s Hardware’s assessment of the July 2026 TSMC announcement was blunt on this point: execution in Arizona “still faces labor, water, and visa constraints, which will do as much as demand to determine how quickly four more fabs are built in Phoenix.”^[6]

By the end of this decade Arizona may be judged on more than whether it attracted 2-nanometer and sub-2-nanometer fabrication. The stronger benchmark is whether a diversified cluster of manufacturing jobs, maintenance businesses, research institutions, equipment companies, and viable energy services remains productive across market cycles. If the state succeeds, its selective treatment will be defensible as an effort to maximize long-lived public capabilities. If benefits are overstated and the public absorbs infrastructure or water liabilities, the policy will need revision regardless of how advanced the chips become.

2.7. The reelection result does not erase institutional constraints

Hobbs’s November reelection — and the Democratic legislative trifecta that accompanied it — gives policy continuity a plausible political foundation, but it does not place the entire regulatory system under one official’s personal control. Arizona’s legislature writes statutes, agencies administer certifications, local governments make many land-use decisions, water rules follow state law, the Arizona Corporation Commission regulates investor-owned utilities while SRP is governed by its own elected board, and public and investor-owned utilities operate through different governance structures.^[10][63] Congress controls national laws and appropriations, while TSMC’s board and customers control corporate deployment decisions. Governors can set priorities, negotiate, appoint officials when vacancies arise, and propose legislation; they cannot guarantee global demand or command every local approval. Hobbs’s own posture is instructive: she has simultaneously welcomed TSMC, signed the datacenter pause, created a taskforce whose charge includes facilitating large-load growth, and told reporters that the state “needs to work together to address the energy challenges raised by both high-tech manufacturing and data centers.”^[64] That is not incoherence. It is the ordinary posture of a chief executive managing five layers with different instruments.

Consequently, Arizona should be read as a case study in multi-institutional selectivity, not a simplistic story about one politician favoring one company. The 2026 juxtaposition emerged from a legislature, a budget negotiation, an economic-development agency, a global manufacturer, two very different utilities, local communities, and voters. The research task for 2027–2030 is to assess whether those institutions can maintain coherent criteria as hundreds of billions of planned private investment move through construction and operation. The three-year datacenter pause ends on June 30, 2029, under current law. Its expiration is not an automatic policy endorsement; it creates a decision point at which lawmakers — now a Democratic legislature working with a second-term Democratic governor — can preserve, revise, or replace the regime with the benefit of three years of data that did not exist in June 2026.

Section 3: A Patchwork Republic — How States Are Rewriting the AI Infrastructure Bargain

3.1. Virginia: taxing electricity consumption rather than rejecting computing

Virginia offers the most direct fiscal counterpoint to Arizona, and it is the natural place to begin because the Commonwealth hosts the largest concentration of datacenter capacity on earth — roughly 11.3 gigawatts of operational capacity in Northern Virginia alone, about 13 percent of the global total, according to the state’s own auditors.^[22] Rather than suspend its sales-and-use-tax exemption for datacenter equipment, which the state Senate had sought to end in 2027, Governor Abigail Spanberger and the General Assembly reached a compromise in the biennial budget signed on June 30, 2026: the exemption survives, but the budget establishes a datacenter electricity-consumption tax of $0.011 per kilowatt-hour on all electricity consumed at each Virginia datacenter — whether purchased from a utility, a competitive supplier, or generated on site — for the period beginning July 1, 2026, and ending before July 1, 2028, with annual general-fund collections capped at $600 million and amounts above the cap returned to operators.^[65][66][67] The measure is the first of its kind in the United States, and the governor described it as positioning Virginia to be “a national leader on data centers.”^[68]

The measure is consequential because it changes the denominator for public revenue. An equipment exemption depends on the timing and classification of purchased hardware, which a sophisticated operator can manage. An electricity-consumption tax follows ongoing use, which it cannot. Under an illustrative, constant 100-MW load drawing power at an 85 percent annual load factor, 744.6 million kilowatt-hours multiplied by $0.011 equals about $8.19 million a year. A larger campus, different load factor, or legal classification would produce a different result. The point is not a forecast of Virginia revenue — the legislature’s own estimate is roughly $1.2 billion over the biennium — but that consumption-based policy connects a portion of the public payment to the continuing use of energy resources rather than to a one-time purchase.^[69]

The design entails tradeoffs that are worth stating carefully. An electricity tax is administratively observable because power is metered. It can also raise the cost of inference, hosting, and model training even when the customer has already paid the full cost of new electric infrastructure through a utility tariff. If an operator buys clean electricity at a premium, uses load-shifting to reduce peak system stress, or installs on-site generation, an undifferentiated kilowatt-hour tax — which Virginia deliberately extended to self-generated power — fails to reward those improvements. Clean Virginia’s executive director complained that the industry “still received the better end of this deal,” because a two-year capped consumption tax does not offset an exemption he valued at nearly $2 billion a year.^[70] A sophisticated policy would distinguish revenue collection, peak-capacity cost causation, and environmental damages rather than expecting one rate to accomplish all three.

Virginia’s earlier JLARC findings show why categorical judgments are inadequate. The industry supports substantial construction and supply-chain activity, but much of that contribution is concentrated in capital-intensive building cycles. The commission’s companion incentive evaluation found that the exemption had cost roughly $2.7 billion over fiscal 2015–2024 and returned about 48 cents in other state revenue for each dollar exempted — a net cost, but far from a total loss, and a performance JLARC judged typical rather than poor by comparison with Virginia’s other incentives.^[22] This does not automatically make the exemption unjustified; the public may value jobs, private investment, or long-term industry development. It means that the benefit should be explained and compared with alternatives. In the days after the budget signing, Spanberger also signed an “Affordable Virginia” package addressing datacenter backup-generator emissions and giving localities new assessment tools, so that Virginia’s regime now combines a preserved exemption, a use-based tax, and local conditions.^[71] In that sense, Layer Selectivity is already operating as an intergovernmental learning process. Arizona’s approach gives Virginia a comparison; Virginia’s use-based tax gives Arizona a different instrument to examine when its own three-year pause approaches expiration.

3.2. Texas: an energy-system discipline that predates the political transition

Texas demonstrates that skepticism about speculative datacenter demand is not confined to one political party. Under Republican Governor Greg Abbott, the legislature enacted Senate Bill 6, signed June 20, 2025, directing the Public Utility Commission of Texas to establish new standards for large-load interconnections in the ERCOT region, including requirements related to transmission planning, cost allocation, financial commitments, site control, and the disclosure of on-site generation.^[72] On June 18, 2026, the commission approved ERCOT’s “Batch Zero” study process for loads of 75 megawatts and above. On June 10, Abbott directed the commission to ensure that datacenter interconnections “will result in reduced residential electrical bills” and to “require data centers to pay for all of their electric infrastructure costs.” On August 3, 2026, he went further, ordering a comprehensive verification and audit of every datacenter project in ERCOT’s interconnection queue — some 250 to 300 Batch Zero projects — before any additional project could move forward, with findings due December 10, 2026.^[73][74][32]

“The PUCT and ERCOT must complete this audit before any data center project moves forward. Any project that fails to comply with the requirements set forth by the PUCT and ERCOT will be denied connection to the Texas grid.”

— Office of Governor Greg Abbott, August 3, 2026 ^[73]

This chronology matters, because Texas regulators were already implementing technical and financial requirements for large-load applications before voters elected Democrat Gina Hinojosa on November 3 — by a 57 percent margin, ending a three-decade Republican hold on the office, in an election in which polling showed 56 percent of Texas voters opposed datacenters.^[75] Hinojosa made the issue central; in an October 22 video she promised, “When I am governor I will issue a moratorium — a ban — on all data center approvals until laws can be passed by the Legislature to protect Texans,” and after her victory she described Abbott’s audit pause as inadequate and called for a longer moratorium followed by legislation requiring union labor, water protections, and an end to the sales-tax exemptions that Good Jobs First estimates now cost Texas more than $1 billion a year, up from a 2023 projection of $130 million.^[76][77][78]

“The data centers connect economic development to everyday concerns.”

— Brandon Rottinghaus, Professor of Political Science, University of Houston ^[79]

A new governor does not inherit a blank slate. Hinojosa may pursue a different balance among utility affordability, growth, water, community consent, and industrial investment, but the law on the books, existing agreements, independent regulators, local officials, and ERCOT processes define the starting position; the Texas legislature meets in regular session only in odd-numbered years, so the earliest statutory change is 2027. It would be inaccurate to attribute every new Texas large-load rule to the November election or imply that a governor-elect has already changed a company’s permit. Investors should distinguish adopted rules, pending rules, announced political positions, and speculation about future legislative negotiations.

Texas also contains a formidable semiconductor-manufacturing opportunity. Samsung’s Taylor investment and the greater Austin semiconductor ecosystem create a Layer 2 interest that is related to but different from the enormous pipeline of Layer 3 projects in central, northern, and western Texas — a pipeline that includes OpenAI’s Stargate campus at Abilene, operating at roughly 0.3 gigawatts in 2026 with 1.2 gigawatts planned, a second Stargate site in Shackelford County planned at 2 gigawatts by late 2028, and Anthropic’s first custom sites under its $50 billion U.S. infrastructure program.^[80][81] TSMC itself confirmed in October 2026 that it had sent teams to evaluate potential Texas investments.^[82] The state’s electricity market, natural-gas resources, solar and wind development, and available land give it distinctive comparative advantages. Yet a megawatt requested in a queue is not a megawatt already secured, and a remote power plant is not valuable to a server campus unless transmission, reliability, or a legally workable supply arrangement connects them.

The potential Texas version of Layer Selectivity is therefore not simply “more fabs and fewer datacenters.” It could mean different proof requirements for different loads. Semiconductor manufacturing may justify a particular workforce and national-security package, while large AI campuses may face stronger commitments to pay for transmission upgrades, submit verified phased demand forecasts, post financial security, or demonstrate sufficient water availability. Conversely, a datacenter with a credible new generation project, measurable local tax return, and flexible load may merit favorable treatment. The category alone should not determine the outcome.

3.3. Pennsylvania: conditional approval and accelerated electricity supply

Pennsylvania illustrates a different strategy: condition large-load development while helping new energy projects move forward. Governor Josh Shapiro announced his Responsible Infrastructure Development (“GRID”) requirements in February 2026 as a voluntary framework, published detailed standards on May 27, and — after the Republican-controlled Senate declined to act on House Bill 2650, which would have codified them — signed Executive Order 2026-05, “Protecting Pennsylvania Consumers from Data Center Impacts,” on August 18, 2026.^[83][84] The order directs the Department of Environmental Protection to evaluate permit applications for datacenters with peak demand above 25 megawatts only if developers have signed a legally binding consent order committing to the GRID requirements and have obtained local approval; it removes all datacenter projects from the state’s Permit Fast Track Program and allows only GRID-certified projects to rejoin; it conditions the Commonwealth’s existing sales-and-use-tax exemption for datacenter equipment on GRID compliance; it bars nondisclosure agreements; and it directs the Special Counsel for Energy Affordability to press the Public Utility Commission to prevent cost-shifting to households.^[85][86][87] Shapiro said more than 100 datacenter proposals were under discussion in the state, many of them speculative, and that the Department of Environmental Protection had received permit applications tied to only about 20.^[88][89]

Pennsylvania’s instrument is notable because it is explicitly two-track. A developer that accepts enforceable conditions on energy affordability, community engagement, workforce, transparency, and environmental protection retains access to streamlined permitting and the tax exemption; a developer that declines loses both. In the vocabulary of this paper, Pennsylvania has made its Layer 3 fiscal preference conditional on Layer 1 and community obligations, rather than suspending it as Arizona did or converting it into a consumption tax as Virginia did.

The supply side is moving in parallel. Shapiro had sued PJM in 2024 over capacity-price spikes, negotiated a price cap in 2025 whose extension in February 2026 the administration valued at $45 billion in consumer savings across the region, signed Electricity Load Forecast Accountability provisions in November 2025 giving the PUC authority to validate the load forecasts utilities submit to PJM, and pressed the grid operator to create an expedited interconnection track.^[90] PJM’s Expedited Interconnection Track, approved by FERC in June 2026 and opened on July 31, accepted its first three projects in October 2026 — two Engie storage projects of 860 megawatts in Crawford County, Pennsylvania, and 800 megawatts in Morrow County, Ohio, and a 455-megawatt LS Power gas-fired uprate in Adams County, Pennsylvania — totaling about 2.1 gigawatts, with up to ten projects per year eligible in 2026 and 2027 for resources above 250 megawatts that can operate within three years.^[91][92] These are selected projects and projected delivery dates; they are not proof that the full output is operating today or that a particular datacenter owns it.

The real puzzle is sequencing. Suppose Pennsylvania adds generation capacity but no datacenter customers materialize when expected. Ratepayers, investors, or developers could face underutilized assets depending on contracts and market structure. Suppose computing demand arrives faster than power and transmission. The consequences could include congestion, scarcity prices, or delays to other industrial users. A state cannot solve either problem with a press release. It needs serious load forecasting, transparent interconnection milestones, and contracts that assign nonperformance risks to parties capable of managing them.

PJM spans thirteen states and the District of Columbia, which means Pennsylvania’s decisions also have regional consequences. An approval in one county can alter capacity and transmission requirements across a wider interconnected market. On June 18, 2026, FERC issued six simultaneous show-cause orders under Section 206 of the Federal Power Act — one to each regional transmission organization and independent system operator under its jurisdiction — directing each to justify its existing large-load tariff provisions or file reforms within 60 days, and proposing a working definition of a “large load” as a new commercial or industrial customer of 50 megawatts or more interconnecting above 69 kilovolts outside a co-location arrangement.^[93][94] The resulting framework is neither purely federal nor purely state: FERC regulates wholesale and transmission terms and co-location rules, while state and local authorities retain substantial authority over land use, distribution, and retail rates. Layer Selectivity must therefore be analyzed in regional market institutions as well as governors’ offices.

3.4. Ohio: semiconductor duration meets datacenter immediacy

Ohio offers a particularly useful comparison of industrial time horizons. Intel’s planned Ohio One campus near New Albany carries a stated first-phase investment of more than $28 billion for two leading-edge fabs, within a site the company once said could absorb $100 billion. Announced in January 2022 with production promised for 2025, the project was rescheduled in February 2025: the first module’s construction is now expected to be complete in 2030, with operations in 2030–2031, and the second module in 2031–2032.^[95] On October 21, 2026, Intel announced that crews had installed the first major structural truss at the first Ohio fab — a visible milestone that, as the company itself was careful to note, does not advance the operating timetable, because completing the shell, installing and qualifying tools, and producing wafers are separate stages.^[96][97] Intel has spent roughly $5.26 billion on the site to date, and its chief financial officer told investors on October 29 that 2027 would be an “investment year” with negative free cash flow as the company prepares its 14A process for production and builds out Ohio.^[98][99]

Meanwhile, the Columbus region faces rapid datacenter interest, alongside more immediate demands for land, electricity, labor, and utility service. Ohio’s November election elevated Amy Acton to governor-elect by about eight points over Vivek Ramaswamy, and Acton campaigned on a pledge to “take immediate action and support a conditional moratorium on data centers” until operators cover 100 percent of their utility costs and build with union labor.^[100][15] The responsible analytical statement is that the political bargaining environment has changed, not that a specific moratorium is already law; Acton will be inaugurated on January 11, 2027, and will face a legislature that has not yet acted. The state’s decisions will have to reconcile near-term power needs with the longer-term objective of turning a proposed semiconductor complex into operational capability.

Ohio also raises an uncomfortable question for industrial advocates. Should a state favor a fab despite a multiyear construction delay and high public cost, while requiring more near-term proof from a datacenter capable of operating sooner? There is a defensible argument for strategic patience when a fab fills a national supply-chain gap; there is also a requirement to reassess milestones and prevent open-ended subsidy dependence. An announcement is a promise contingent on execution, not a permanent claim on scarce public resources. Conversely, a datacenter’s faster delivery can yield near-term property revenue and construction demand that a future fab has not yet produced.

The Ohio case should therefore measure performance in calendar time, not only dollars. It should compare promised and actual manufacturing readiness, signed offtake or customer demand, technician training, utility construction, housing and transport capacity, and opportunity costs to other firms. By 2030, a regional government that manages both delayed fab development and immediate computing demand may possess a competitive advantage over one that promotes all projects as though construction schedules and economic multipliers were identical.

3.5. Indiana, Michigan, and New York: incentives, conditions, and prohibition

Indiana shows that many states continue to see datacenters as desirable economic-development assets. The Indiana Economic Development Corporation administers a Data Center Sales Tax Exemption program, and Indiana’s 2026 tax legislation refined provisions involving payments connected to qualified datacenter electricity purchases under particular awards.^[101] The scale of what Indiana has offered is instructive: Good Jobs First estimates that Amazon’s New Carlisle campus received a 50-year, roughly $4 billion local abatement and a separate $4 billion technology-and-property package over 35 years, for a combined $8.2 billion — among the largest datacenter subsidy packages ever documented — while Google has announced a $2 billion datacenter cluster in Fort Wayne.^[102][103] The point is not that Indiana has abolished its incentives, but that it is refining the terms on which some tax benefits and local arrangements operate. States can maintain a recruitment strategy while reconsidering exactly who receives which advantages.

This is relevant to AWS and other major operators considering Midwest expansion. A project may be attractive because of land, available electric capacity, fiber routes, and the possibility of long-term industrial partnerships. If Indiana requires a particular payment, certification, or local approval, companies must incorporate that rule into their investment model. A headline incentive amount tells only part of the story. Reliability requirements and local utility agreements can be economically decisive, particularly for high-density training clusters that cannot tolerate extended outages.

Michigan, where Jocelyn Benson won the governorship by nearly eleven points, offers another instructive contrast. The state has a deep automotive and industrial workforce, strong interest in AI-enabled manufacturing, a constrained energy transition, and at least one Stargate site under construction.^[14][80] It may attract demand for chips, industrial models, robotics, automotive applications, energy generation, and selected datacenters. A governor’s first task is not to choose a single winning layer. It is to understand how modernization of manufacturing and the creation of computing infrastructure interact with household rates, utility investment, old industrial sites, and worker retraining. Michigan’s potential strength lies in connecting Layer 5 industrial adoption to Layers 1–3, not simply copying Phoenix or Northern Virginia.

New York represents the far end of the spectrum. In July 2026 it became the first state to enact an outright prohibition — a one-year ban on new AI datacenters — even as it continued to finance Micron’s Clay megafab with a term sheet worth up to $5.5 billion over twenty years and as the STAMP datacenter subsidy controversy continued.^[104][105] No state illustrates Layer Selectivity more starkly: the same government that paused Layer 3 construction altogether is the most generous public partner in the country to a Layer 2 memory fab. Whether that combination is coherent depends entirely on whether the state can articulate the public-value difference between the two — and on whether the ban survives the inevitable argument that it simply exports load to Pennsylvania and Ohio.

At the same time, each state has different legal institutions. Michigan utility regulation cannot be assumed to match ERCOT or SRP, and a governor-elect’s campaign statements do not by themselves reprice a signed power contract. Credible forecasting requires an inventory of what is law, what is regulatory proceeding, what is corporate announcement, and what is political proposal. Treating those categories as interchangeable would undermine a serious 2027–2030 paper.

3.6. California and Georgia: more accountability without abandoning technology

California demonstrates selectivity extending beyond physical facilities. On September 21, 2026, Governor Gavin Newsom signed a package of seven datacenter bills — AB 1577, AB 2383, AB 2469, AB 2619, SB 886, SB 887, and SB 1168 — that his office described as “the most comprehensive data center laws in the nation.” Three of the laws shift electric-infrastructure costs from residential customers to datacenter operators and require operators to pay their share of grid-upgrade costs while complying with the state’s clean-energy procurement rules; three require disclosure of water and electricity use, workforce, and land use to local governments and water suppliers, with water-supply assessments and scarcity plans required from January 1, 2028; and SB 887 eliminates blanket exemptions from environmental review under CEQA, introducing a discretionary element into land-use review even where zoning permits datacenters by right.^{[106][107][108]} The package marked a reversal: Newsom had vetoed a water-disclosure bill in 2025 on the ground that regulation could stifle AI growth. CalMatters attributed the change to “intensifying public backlash,” citing the Gallup finding and a July Public Policy Institute of California poll showing similar opposition statewide, even as the Data Center Coalition reported that the industry supported 665,500 jobs and $14.1 billion in taxes in California in 2024.^[108]

“With these laws, we are ensuring that Californians remain in the driver’s seat — and that those profiting from data centers aren’t doing so at our expense.”

— Governor Gavin Newsom, State of California, September 21, 2026 ^[106]

California’s regulatory approach therefore spans Layer 3 physical infrastructure and, through its separate frontier-model statutes, Layer 4 model governance. The state can remain a leading home for software, frontier models, and applications while requiring greater disclosure and cost responsibility from the facilities hosting those systems. Xavier Becerra’s election as governor, by roughly 65 percent to 35 percent, means a new administration will inherit those enacted statutes rather than start from zero.^[109] In this paper’s time frame, Newsom is still governor; Becerra is governor-elect. Future executive priorities could change implementation, enforcement, permitting practice, or legislative negotiation, but the enacted legal baseline matters. For developers, the distinction between a new governor’s preferences and already effective statutory obligations is essential to discount-rate analysis and project scheduling.

Georgia’s election of Keisha Lance Bottoms — a Democratic flip of a Republican-held office — introduces another political transition in a state with substantial datacenter development and an existing high-technology datacenter equipment exemption administered by the Department of Revenue.^[110] The fiscal facts have shifted under the policy: Georgia’s own budget planners raised the projected fiscal 2026 cost of the exemption by 664 percent to about $2.5 billion, with fiscal 2027 approaching $3 billion and localities losing an estimated $1.1 billion in 2026 from state-awarded exemptions, figures publicized by Good Jobs First and reported by the Atlanta Journal-Constitution.^[78][111] Bottoms campaigned for “an immediate statewide freeze” on datacenters and told reporters, “I just want us to pause,” while her Republican opponent, Rick Jackson, proposed local hurdles and accused her of hypocrisy for backing a Midtown Atlanta datacenter deal as mayor.^[112][15] The existence of proposals does not prove their enactment. For 2027–2030, the core question is how the new administration, the legislature, the Public Service Commission, and local governments balance industrial recruitment against revenue transparency, generation planning, and household electricity bills.

These examples overturn a tempting binary. The country is not neatly separating into “pro-AI” and “anti-AI” states. One jurisdiction may welcome fab construction and scrutinize datacenter purchases; another may impose a per-kWh tax; another may use development standards or financial-security requirements; another may focus on water reporting or local authority; one has, for a year, said no altogether. Layer Selectivity names the variation in the public contract, not a universal direction of policy.

3.7. A comparative map of the emerging policy regimes

The cases can be summarized as follows. These are stylized analytical descriptions, not categorical statements that every facility in a state faces the same terms.

Jurisdiction	Verified 2026 policy or project anchor	Mechanism of selectivity	Principal unresolved 2027–2030 question
Arizona	TSMC’s additional $100B plan ($265B cumulative); HB 4168 three-year pause on new datacenter tax certifications (July 2026–June 2029); HB 2756 extra-high-load-factor reporting; 107 late applications approved	Semiconductor recruitment versus selective tax-exemption suspension plus utility cost-of-service disclosure	What replaces or follows the pause after June 2029, now under a Democratic trifecta?
Virginia	$0.011/kWh datacenter consumption tax, July 2026–June 2028, capped at $600M/yr; equipment exemption preserved	Use-based public payment and utility cost assignment	Does the Commonwealth extend, revise, or replace the charge — and revisit the exemption — in 2028?
Texas	SB 6 large-load rules; Batch Zero screening; Abbott’s August 2026 audit of 250–300 queued projects; Hinojosa elected on moratorium pledge	Interconnection credibility and infrastructure-cost allocation	How will the incoming administration and the 2027 legislature amend or extend the regime?
Pennsylvania	Executive Order 2026-05 GRID requirements (>25 MW); exemption conditioned on compliance; PJM Expedited Interconnection Track projects	Conditional development and coordinated electricity supply	Can new generation and approved load arrive in a fiscally safe sequence?
Ohio	Intel Ohio One first truss (Oct 2026), operations 2030–2031; Acton elected pledging conditional moratorium	Long-duration strategic manufacturing versus near-term large-load oversight	How will fab commitments and datacenter growth be reconciled?
Indiana	Existing datacenter exemption; $8.2B Amazon New Carlisle package; 2026 statutory refinements	Continued recruitment with eligibility and local-payment conditions	Can benefits be measured against utility and community costs?
New York	One-year ban on new AI datacenters (July 2026); up to $5.5B for Micron’s Clay fabs	Prohibition of Layer 3 alongside maximal support for Layer 2	Does the ban lapse, extend, or convert into conditions?
California	Seven datacenter laws (Sept 2026): cost allocation, water and energy disclosure, CEQA review; frontier-model oversight	Physical-infrastructure accountability alongside software leadership	How will the incoming Becerra administration implement inherited laws?
Georgia	Existing exemption now costing ~$2.5B/yr; Bottoms elected pledging a freeze	Incentive review and utility policy	Which fiscal preferences remain justified by measurable economic returns?

3.8. Why regional competition can still be healthy

There is a risk in treating policy variation itself as an obstacle. Competitive federalism can reveal information: some states may find that equipment exemptions attract economically valuable projects; others may discover that utilities or households bear costs that the project sponsor should pay. Different tariff models can be compared over time, and the 2026 cohort of experiments — a pause, a tax, an audit, a conditional order, a ban — is as close to a natural experiment as American industrial policy has produced in a generation. However, a race to offer larger subsidies without accounting for grid impacts can produce a transfer of fiscal resources to already wealthy firms, while a race to impose restrictions without understanding system-wide effects can redirect investment to less efficient or less transparent jurisdictions. Good Jobs First’s November 2025 survey found that at least 36 states offer datacenter-specific subsidies but only 11 disclose which companies receive them, which means that in most of the country the experiment is being run without a control group or a lab notebook.^[113]

“Taxpayer costs are rising far faster than policymakers anticipated, while disclosure and accountability continue to lag behind.”

— Kasia Tarczynska, Senior Research Analyst, Good Jobs First ^[111]

The appropriate comparison is not which state appears most enthusiastic in a single press release. It is which jurisdiction provides predictable rules, truthful resource accounting, enforceable obligations, and sufficient infrastructure capacity. A developer may rationally prefer a higher posted tax with an available substation and clear water rights to a zero-tax promise accompanied by years of uncertain interconnection. An advanced manufacturer may prefer a stable workforce program and secure equipment imports to a one-time subsidy without supplier depth. What governments sell to investors is increasingly a credible ability to execute.

Moreover, the federal system has an aggregation problem. If every state pursues only the most prestigious layer, the nation may underbuild supporting infrastructure or regional inference capacity. Conversely, if each locality accepts large power demands because another jurisdiction will capture the downstream AI benefits, grid and water costs can be misallocated. The federal response so far has been voluntary rather than statutory: the White House’s Ratepayer Protection Pledge, first signed on March 4, 2026, by Google, Microsoft, Meta, Oracle, xAI, OpenAI, and Amazon, and expanded on July 23, 2026, to 23 Republican governors, Duke Energy, NextEra, and roughly 200 stakeholders, commits signatories to build, bring, or procure the energy their datacenters need and to cover the full cost of infrastructure upgrades — while a bipartisan Ratepayer Protection Act sponsored by Representatives Gabe Evans and Kathy Castor, which would require state regulators to consider standards for facilities above 100 megawatts, cleared the House Energy and Commerce Committee unanimously the same week.^[114][104]

“We’re insisting that AI data centers and Big Tech companies pay their own way.”

— President Donald J. Trump, Ratepayer Protection Pledge roundtable, July 23, 2026 ^[115]

A former Energy Department official observed that the pledge “doesn’t actually implement” anything and depends on state legislatures and regulatory commissions to pass tariffs that embody it.^[115] That is the correct reading, and it points to the conclusion of this section: Layer Selectivity must be coordinated across jurisdictions through transparent data and consistent cost-causation principles, without pretending that a single nationwide siting rule can solve every local problem.

Section 4: Corporate Strategy Under Layer Selectivity — The Same Chips, Different Political Economies

4.1. TSMC, NVIDIA, and AMD: fabrication geography versus computing geography

For chip companies, the distinction between fabrication and deployment is fundamental, and the third-quarter 2026 reporting season made the scale of both unmistakable. NVIDIA and AMD design the processors and accelerator platforms; TSMC manufactures nearly all of their leading-edge products. TSMC’s third quarter, reported October 15, 2026, produced revenue of $46.72 billion, up 41.2 percent year over year and above the top of guidance, with net income of NT$763.06 billion up 68.7 percent, a gross margin of 66.4 percent, and fourth-quarter guidance of $52.3 billion to $54.3 billion; management reaffirmed 2026 capital spending of $60 billion to $64 billion and full-year revenue growth of about 44 percent in U.S. dollars.^{[116][9][117]} Chief Financial Officer Wendell Huang attributed the quarter to “strong demand for our leading-edge process technologies,” and C.C. Wei described the gap between demand and supply as “still quite big.”^[116][118]

A fab in Phoenix can serve chip customers whose final systems are deployed in Texas, Virginia, Ohio, Japan, the Middle East, or elsewhere. A state incentive directed at domestic fabrication therefore need not be accompanied by identical support for every datacenter that will ultimately use the chips. In fact, chip manufacturing diversification may be nationally valuable precisely because output can supply a broad geographic market.

There is direct evidence of the intended connection. NVIDIA announced in late 2025 that the first Blackwell wafer had been produced on U.S. soil at TSMC’s Arizona facility.^[119] That milestone is not the same as saying every finished Blackwell system was fabricated, packaged, and tested entirely in Arizona. Wafer production, advanced packaging, high-bandwidth memory integration, system assembly, and eventual deployment are distinct industrial steps. NVIDIA’s own second-quarter fiscal 2027 results, reported August 26, 2026, show how far the deployment layer has outrun the fabrication layer: revenue of $96.2 billion, up 106 percent year over year; datacenter revenue of $89.0 billion, up 117 percent; hyperscaler revenue of $48.7 billion and revenue from AI clouds, industrial, and enterprise customers of about $40 billion, up 138 percent; third-quarter guidance of $108 billion; a forecast of roughly 70 percent revenue growth in fiscal 2028, constrained by supply rather than demand; and shipments to China of less than 1 percent of datacenter revenue.^{[120][121][122]}

“AI has reached its inflection point. It’s doing useful work. Its tokens are productive and profitable. Now, compute is revenue.”

— Jensen Huang, Founder and CEO, NVIDIA, August 26, 2026 ^[120]

Huang also offered a number that belongs in every state’s incentive analysis: the cost of a gigawatt of datacenter, he said, has risen from about $30 billion five years ago to about $60 billion today.^[123] If that figure is even approximately right, the equipment inside a single gigawatt campus is worth more than two leading-edge fab modules, which means that the tax base at stake in a sales-and-use-tax exemption has grown far faster than the statutes that created the exemptions.

AMD’s third-quarter 2026 results, reported November 3, 2026, make the same point from a competitor’s vantage. Revenue reached a record $13.4 billion, up 45 percent; datacenter segment revenue was a record $8.5 billion, up 97 percent, and now represents 64 percent of the company’s revenue, up from 47 percent a year earlier; and the company began revenue shipments of its EPYC “Venice” processors, built on TSMC’s 2-nanometer process, while launching its Instinct MI450 accelerators and Helios rack-scale systems.^[124][125] AMD’s May 2026 announcement had described the Venice production ramp in Taiwan with plans for a future ramp at TSMC Arizona; future Arizona ramp remains the accurate description until the company’s stated event occurs.^[126]

“We delivered a record quarter with revenue exceeding the high end of our guidance as Data Center segment sales nearly doubled year-over-year. Demand for our Instinct and EPYC products continued to strengthen, and we expect growth to accelerate further in the fourth quarter as we scale Helios and ‘Venice’ production.”

— Dr. Lisa Su, Chair and CEO, AMD, November 3, 2026 ^[124]

These distinctions have economic implications. Chip firms must negotiate wafer capacity, process qualifications, advanced packaging, yields, memory supply, pricing, export restrictions, and delivery schedules. Datacenter owners negotiate power, cooling, land, networking, site approvals, and utility service. They are not exposed to identical policy risks. Layer Selectivity may improve the probability of domestic silicon production while raising the cost of certain domestic deployment sites. That does not necessarily reduce sales of accelerators, because buyers may move workloads or build in other jurisdictions. It can, however, change the domestic geographic distribution of computing and the cost of delivering an end-to-end system.

This is why Jensen Huang’s strategic calculus and Lisa Su’s calculus should not be reduced to a count of U.S. datacenter campuses. For NVIDIA and AMD, accelerated domestic fabs may provide supply-chain resilience and optionality. For customers, higher deployment costs may encourage more efficient chips and better software utilization. Those incentives could accelerate adoption of lower-cost inference configurations, CPU-heavy agentic architectures — NVIDIA noted that agentic AI is driving an acceleration in demand for datacenter CPUs — rack-scale networking, or custom accelerators.^[123] But a sufficiently restrictive domestic environment could also push incremental deployment offshore, reducing the national benefits of a robust chipmaking supply base. Strengthening Layer 2 without preserving workable Layer 3 markets creates a potential policy mismatch.

4.2. Intel and Micron: the manufacturing layer is not a single business model

Intel’s Ohio One investment exposes a different problem from TSMC’s Phoenix expansion. Intel seeks to expand its foundry presence while managing its own product portfolio, technology roadmap, and capital intensity. Its third-quarter 2026 results, reported October 29, showed revenue of $16.5 billion, up 21 percent; datacenter and AI revenue of $7.3 billion, up 78 percent — its strongest datacenter growth in fifteen years, according to the chief financial officer; foundry revenue of $6.7 billion, of which only $411 million was external, with a foundry operating loss of $1.9 billion; the release of the 14A 0.9 process design kit to select customers in October, with risk production planned for 2027 and volume production in 2028; and weak fourth-quarter guidance attributed to a PC slowdown that sent the shares down 21 percent.^{[127][128][99]} Chief Executive Lip-Bu Tan said the results “demonstrate improving execution, meaningful progress expanding supply and continued advancement of our process technology,” while declining, as foundry convention requires, to name 14A customers.^[127]

A postponed completion target means that the Ohio community may wait years for full operating employment and industrial output. A state must decide how long to maintain its support, what milestones trigger public payments, and whether incentives can be revised when the project’s schedule changes. The answer should depend on contractual terms, public risks, and strategic value, not either unconditional loyalty or reactive cancellation. The Government Accountability Office reported in August 2026 that a majority of CHIPS Act manufacturing projects are expected to be complete by 2028 and all by 2033, which frames Ohio One as an outlier in duration even among federally supported fabs.^[129]

Micron provides a complementary example, and its fiscal 2026 results are among the most extraordinary in the history of the memory industry. For the fiscal fourth quarter ended September 3, 2026, Micron reported revenue of $54.23 billion, nearly five times the $11.32 billion of a year earlier; GAAP net income of $37.70 billion; a gross margin of 87.0 percent; and full-year revenue of $133.19 billion with net capital expenditures of $27.37 billion.^[130][131] DRAM revenue of $39.8 billion was 73 percent of the total, and Chief Executive Sanjay Mehrotra told analysts that the company — the only U.S.-based maker of high-bandwidth memory — was working with NVIDIA on the industry’s “first custom HBM implementation.”^[132] Micron has committed to invest roughly $250 billion in U.S. manufacturing through 2035: construction of its second Idaho fab began in 2026 with initial wafer output expected by late calendar 2028; it broke ground on its first New York fab in January 2026, with supply in 2030 and beyond; and it has up to $6.1 billion in CHIPS direct funding and a non-binding term sheet from New York worth up to $5.5 billion over more than twenty years.^[105][133] Those announced figures span a long investment horizon and multiple states; they should not be treated as annual capital expenditure or as immediately available high-bandwidth-memory supply.

The policy lesson is that even Layer 2 requires more than one type of public bargain. Logic fabrication, DRAM production, high-bandwidth-memory packaging, specialty-chip manufacturing, and semiconductor research carry different strategic functions and labor needs. A tax credit that is valuable to one may be poorly targeted to another. States may be tempted to recruit glamorous leading-edge logic fabs while neglecting the less visible materials, test, assembly, maintenance, and skilled-worker systems that determine whether chips can actually be delivered. The federal government’s own record is mixed: GAO found that Commerce had canceled awards representing $7.8 billion of the $11 billion appropriated for advanced microelectronics research and development, including the National Semiconductor Technology Center award, and in October 2026 the department began signing CHIPS R&D letters of intent that condition funding on the government taking minority equity stakes in recipients.^{[134][135][136]}

For 2027–2030, the most meaningful semiconductor capacity measure is qualified, usable output, not the number of groundbreakings. A construction site may take years to move from structural work to equipment installation, process qualification, acceptable yields, and sustained shipments. The fact that a company identifies a site for several fabs does not establish simultaneous capacity additions, and an economic-impact forecast does not transform a projected job into an employed worker. Transparent public reporting should separate project authorization, actual cumulative investment, construction progress, installed tools, qualified process nodes, manufacturing output, and employment.

4.3. Google and Amazon: the hyperscaler is both a chip designer and a large-load customer

Hyperscalers make Layer Selectivity especially interesting because their businesses occupy multiple layers simultaneously. Amazon’s AWS operates datacenters, designs custom silicon such as Trainium and Graviton, supplies model-training and inference capacity, and sells software services consumed by developers and agents. Google operates Google Cloud, develops its Tensor Processing Units, trains and serves Gemini models, and builds applications. Microsoft and Meta likewise combine infrastructure ownership with AI research, model deployment, and large application ecosystems. A state may favor part of a company’s investment and impose additional conditions on another part without changing its view of the corporation as a whole.

The third-quarter 2026 disclosures illustrate the scale of this integration. Amazon reported revenue of $201.9 billion, up 12 percent; AWS revenue of $46.0 billion, up 39.3 percent — its sixth consecutive quarter of acceleration and fastest growth in nearly five years — with AWS operating income of $17.8 billion and a backlog of $578 billion; cash capital expenditures of roughly $75.9 billion in a single quarter, “primarily” for AWS and AI; reaffirmed 2026 cash capex of about $220 billion; and trailing free cash flow that turned negative.^{[137][138][139]} The company said revenue commitments for both Graviton and Trainium had each risen more than twentyfold year over year and that it intends to make Trainium available outside AWS beginning in late 2027.^[139][140]

“Our AWS investment in functionality and infrastructure is working, with over 39% year-over-year revenue growth in Q3 ($184 billion revenue run rate) and projected revenue growth rates of over 40% both in 2027 and in 2028.”

— Andy Jassy, President and CEO, Amazon, October 29, 2026 ^[137]

Alphabet, reporting the same week, missed on both revenue and earnings for the first time in nearly four years, posted capital expenditures of $51.4 billion in the quarter and negative free cash flow of $9.07 billion, held its 2026 capex guidance at $195 billion to $205 billion, and told analysts to expect capex to “increase significantly in 2027”; the shares fell about 10 percent.^[141] Google had introduced its TPU 8t and 8i chips at Cloud Next in April 2026, separating architectures aimed at training and inference, and Anthropic’s reported five-year, $200 billion commitment for Google Cloud capacity and Broadcom-built TPUs shows how a Layer 4 company’s location decisions flow through a Layer 3 partner’s balance sheet.^[142][143] These are company statements about products and plans, not independent proof of the savings that every customer will realize.

For a hyperscaler assessing Arizona, Virginia, or Texas, tax incentives are only one variable in a larger capital-allocation problem. The company must choose among owning land and buildings, leasing capacity, using an independent developer, deploying custom chips or merchant GPUs, entering multiyear utility contracts, and positioning compute near applications and customers. If a state withdraws a tax exemption but offers reliable transmission capacity, the net return may remain attractive. If a rival state promises relief without credible electrical service, the nominally favorable tax outcome may be worthless. Site selection increasingly resembles a portfolio optimization problem across five layers, not a subsidy auction for a single asset.

Some workloads are more geographically flexible than others. Large frontier-model training can tolerate location away from the end user if interconnect quality, storage, and communications costs are managed. Latency-sensitive inference may need to be closer to customers or critical enterprise data. Persistent agents, robotic systems, and real-time industrial applications can require predictable regional service, even when the underlying model is trained elsewhere. The location of compute therefore influences Layer 5 competitiveness as well as Layer 3 construction. States seeking to attract digital-service jobs should not assume they can restrict all local computation without consequences for the software and enterprise sectors they wish to develop.

4.4. Meta and Microsoft: the value of converting announced capex into usable capacity

Meta and Microsoft face the dual obligation of keeping up with rapid model improvement while building an expensive physical delivery network. Meta’s third quarter, reported October 21, 2026, showed revenue of $64.6 billion, up 26 percent, capital expenditures including finance-lease principal of $34.91 billion, slightly negative free cash flow, and a narrowed 2026 capex range of $135 billion to $145 billion.^[144][145] Microsoft’s fiscal 2027 first quarter, reported October 28, showed revenue of $92.3 billion, up 19 percent; Microsoft Cloud revenue of $63.9 billion, up 30 percent; Azure revenue disclosed in dollars for the first time at $33.2 billion, up 48 percent; a commercial remaining performance obligation of $677 billion, up 73 percent; and quarterly capital expenditures including finance leases of $45.2 billion, with Chief Financial Officer Amy Hood guiding to more than $50 billion in the following quarter and calendar-2026 spending of about $170 billion — a figure she said would have been higher had third-party capacity arrived on time.^{[146][147][148]}

“We still have demand that exceeds supply.”

— Amy Hood, Executive Vice President and CFO, Microsoft, October 28, 2026 ^[147]

Taken together, the four largest hyperscalers’ 2026 guidance — roughly $220 billion for Amazon, $195–205 billion for Alphabet, about $170 billion for Microsoft, and $135–145 billion for Meta — implies combined capital spending in the range of $720 billion to $740 billion for calendar 2026, an order of magnitude above the industry’s pre-2024 baseline and consistent with the IEA’s observation that the five largest datacenter companies spent more than $400 billion in 2025 and planned a further 75 percent increase in 2026.^[149][150] As policy becomes more selective, the key capital-market question shifts from how much capex was announced? to how much contracted, connected, productive capacity does that capex create? Microsoft’s acknowledgment that its spending was constrained by late third-party deliveries, and Alphabet’s and Amazon’s negative free cash flow, are early signs that the market is beginning to ask the second question. A project with a large funding announcement but no firm service date, reliable power pathway, or enforceable construction schedule may have lower economic value than a smaller project already integrated into the grid. Investors should separate capital spending on chips, servers, buildings, land, and electricity infrastructure, and recognize that each is depreciated over a different useful life.

There is a second-order consequence for established operators. A sudden change to new incentive eligibility can create differences between older certified campuses and new entrants. Existing large companies may possess early site options, utility agreements, and experienced regulatory teams; the 107 Arizona approvals went overwhelmingly to parties already present in the state. Startup compute providers can have more difficulty financing the same requirements. A tax exemption suspension that is neutral on its face can therefore influence competition through timing and balance-sheet strength. Policymakers should monitor whether new conditions preserve contestability or inadvertently restrict the market to a handful of incumbents.

This does not mean that existing companies should be punished for planning early. It means public policy must make explicit whether the objective is to reward early investment, protect community costs, create domestic competition, or some combination. If the state changes rules abruptly and retroactively, risk premiums rise. If it guarantees privileges indefinitely without performance conditions, public value may deteriorate. Predictable phase-ins, published eligibility tests, and enforceable milestone reporting can reduce both problems.

4.5. OpenAI, Anthropic, and the economics of model-location choice

For frontier-model companies, the marginal value of a datacenter depends on the productivity of the model-training or inference workload that it can support. OpenAI and Anthropic contract with hyperscalers, specialized cloud providers, or hardware partners rather than own every building — although both have moved toward direct ownership, OpenAI through the $500 billion Stargate program, which it says has secured roughly 10 gigawatts of U.S. capacity and whose Abilene flagship was operating at about 0.3 gigawatts in 2026 on the way to 1.2, and Anthropic through a $50 billion program of custom sites beginning in Texas and New York, with Fluidstack as the build partner.^[80][81] Reporting in 2026 put OpenAI’s planned compute at about 30 gigawatts by 2030 and Anthropic’s at 7 to 8 gigawatts.^[151] Their physical location decisions are therefore made both directly and through commercial counterparties, and a state datacenter policy can affect them through the price and availability of third-party capacity even if they never personally apply for an Arizona certification.

The indirect channel matters because model-development economics are changing. Larger training experiments, reinforcement learning, tool use, long-context inference, and multi-agent workflows have different ratios of GPUs, CPUs, memory, networking, and storage. A compute campus optimized for one workload may not be the least-cost platform for another by 2029. If state incentives favor equipment purchases without performance or flexibility conditions, they can reward hardware accumulation rather than productive utilization. Conversely, unusually burdensome recurring charges can discourage efficient domestic inference even when applications provide considerable economic value. The IEA’s technical observation that AI datacenters see swings of server load exceeding 50 percent of rated capacity within a second is a reminder that the physical profile of a training cluster is unlike that of a conventional industrial load, and that tariffs designed for steel mills may price it badly in both directions.^[29]

One emerging alternative is to reward verified performance that benefits the system, such as flexible demand, credible load forecasts, recoverable grid investments, or local workforce training, rather than count GPUs as a proxy for innovation. Such policies must be carefully designed to avoid disclosing proprietary model details or penalizing novel architectures. The public interest lies in infrastructure effects and contractual accountability, not in state officials choosing specific neural-network designs.

4.6. The overlooked Layer 5 dividend

Applications and agents may create the largest eventual public gains while receiving the least physical-infrastructure publicity. Consider an Arizona healthcare provider using a diagnostic assistance system, a Michigan automotive supplier using autonomous quality inspection, or an Ohio factory deploying predictive maintenance. The benefits are delivered through better output, reduced delays, safer operations, and worker augmentation. None necessarily requires a new hyperscale campus in the same locality. It may require reliable broadband, affordable inference, data governance, employee training, and a market for trustworthy applications.

The macroeconomic debate over how large that dividend will be is unresolved and sharply contested among the most prominent economists in the field. Daron Acemoglu of MIT, a Nobel laureate, estimated in his 2024 paper on the “simple macroeconomics of AI” that total-factor-productivity gains over the following decade would be below 0.55 percent even accounting for AI exposure, and he has since described much of the productivity discourse as speculative to the point of fiction.^[152][153] Erik Brynjolfsson of Stanford’s Digital Economy Lab argues the opposite case: that U.S. productivity grew roughly 2.7 percent in 2025, nearly double the prior decade’s average, and that the economy is moving from the investment phase to the “harvest phase” of the productivity J-curve.^[154]

“The updated 2025 US data suggests we are now transitioning out of this investment phase into a harvest phase where those earlier efforts begin to manifest as measurable output.”

— Erik Brynjolfsson, Director, Stanford Digital Economy Lab ^[154]

“We’re creating the potential to have massive productivity gains and a lot more wealth.”

— Erik Brynjolfsson, Stanford University, in the Federal Reserve Bank of Richmond’s Econ Focus ^[152]

Brynjolfsson’s further claim — that for every dollar of tangible investment in datacenters and software, firms historically make nine or ten dollars of intangible investment in process redesign, training, and organizational change — is the economic foundation of the Layer 5 argument.^[152] If it is even partly right, the public return on AI depends far more on what happens in hospitals, factories, and small businesses than on the location of any given hall of accelerators. The International Monetary Fund, whose managing director Kristalina Georgieva has compared the current buildout to the construction of the railways, estimates that AI could lift global growth by up to 0.8 percentage points, while warning that the benefits will accrue to countries that “make AI happen” rather than those who “watch it happen.”^[155][156]

“AI has arrived; it is here. Investment is surging into data centers, power plants, and neural networks. It is like when the world was building the railways: the old ways of doing things will be changed forever, and this time it will be much faster.”

— Kristalina Georgieva, Managing Director, International Monetary Fund, February 19, 2026 ^[155]

This observation cuts both ways. Governments should not overvalue a proposed local AI factory on the assumption that all national AI productivity will somehow accrue locally. But they should not assume that model and agentic capabilities are free-standing abstractions unrelated to physical capacity either. Inference prices and network quality affect adoption. Regions that develop application expertise, universities, startups, procurement standards, and workforce programs may capture more durable income than regions that focus solely on the largest physical project.

Layer Selectivity provides an incentive to ask whether subsidy dollars would produce higher social returns if spent on Layer 5 diffusion — small-business adoption grants, community-college training, public-sector procurement pilots, or shared evaluation infrastructure — rather than on another marginal equipment exemption. This is not an argument against large facilities. It is a demand for opportunity-cost accounting across the full chain. Industrial policy should not measure success only at the construction fence.

4.7. Financing consequences: not all project risk is sovereign risk

Investors evaluating 2027–2030 AI capacity need to distinguish several risks that frequently appear together. Regulatory risk concerns tax, siting, environmental, and utility changes. Execution risk concerns construction, equipment, cooling, labor, and commissioning. Demand risk concerns actual paid use of compute or semiconductor output. Counterparty risk concerns customers, suppliers, lenders, and power-contract partners. Technology risk concerns efficiency gains or obsolescence that change asset economics. A governor’s election can alter the first category without determining the others.

The macro-financial dimension of this risk is now a subject of official concern. Georgieva, presenting the IMF’s assessment of the eurozone in 2026, described the possibility of an AI bubble bursting as “low probability, very high impact,” at a moment when Big Tech’s AI-related capital commitments for 2026 exceeded $700 billion.^[157] A multi-method academic evaluation published in June 2026 reached the measured conclusion that AI could be economically significant without generating the productivity acceleration implied by the most optimistic market narratives, and that if value accrues primarily to firms controlling compute, distribution, and data, then application-layer firms without durable moats may be overvalued even as the technology succeeds.^[158] A state that stakes its budget on a datacenter exemption is, whether it knows it or not, taking a position on that question.

An appropriate financing model assigns probabilities and costs to actual project milestones: signed land rights, water supply, interconnection agreement, incremental infrastructure charges, zoning, equipment delivery, commissioning, and committed customers. The impact of a three-year tax-incentive pause on a fully funded, power-secured campus may be limited; on a thinly capitalized speculative project, it can be pivotal. Likewise, a fab award conditioned on future spending has different cash-flow implications from an unconditional check already disbursed.

The deeper consequence is a change in who owns the option to wait. A developer with multiple sites and a strong balance sheet can delay a tax-disadvantaged location while retaining strategic flexibility. A local government cannot as easily defer housing, schools, substations, or emergency services once demand arrives. A well-designed public agreement protects the community from unilateral changes in private schedules while preserving the ability of investors to adapt to technological progress. As selection becomes more explicit, the quality of risk allocation — not merely subsidy generosity — will determine which projects survive into operation.

Section 5: The New Congress, the White House, and Three Futures for 2027–2030

5.1. The institutional result of the November midterms

The 2026 elections changed the national political balance more decisively than most forecasters expected. Democrats captured both chambers of Congress — flipping at least seven Senate seats, including Texas, Ohio, Iowa, Kansas, Maine, North Carolina, and South Carolina, to reach at least 54 seats, and winning more than 250 House seats, their largest majority since 2008 — while President Donald Trump remains in office until January 2029.^{[12][13][159]} CNN’s election analysis noted that backlash against the environmental and community impact of new datacenters had turned AI spending into a key issue of the campaign, and a review of the gubernatorial field found pause or moratorium proposals in twenty of the thirty-six governor races on the ballot.^[159][15] The result creates divided government at a moment when technology companies are committing extraordinary sums to factories, datacenters, energy supply, models, and applications. It is important to say what the outcome does and does not establish. Congressional majorities can shape committee agendas, issue subpoenas where lawful, conduct oversight, develop appropriations bills, and negotiate legislation. They cannot by themselves erase statutory programs, force a president to sign every proposed bill, overcome a Senate filibuster that still requires sixty votes, or directly operate a state’s utility commission.

The immediate consequence may be an increase in scrutiny before a change in law. New committee leaders could request detailed reporting about federal CHIPS awards, the equity stakes Commerce has begun to demand in exchange for R&D funding, foreign chip suppliers, advanced accelerator exports to China, datacenter electricity obligations, labor conditions, and public return from subsidies. Senator Bernie Sanders and Representative Alexandria Ocasio-Cortez have introduced parallel bills to impose a national moratorium on AI datacenters; a bipartisan Ratepayer Protection Act has already cleared the House Energy and Commerce Committee; and Senator Chris Van Hollen’s Data Center Tax Accountability and Disclosure Act would condition federal tax benefits on public reporting of energy and water use.^{[160][114][161]} Oversight may affect project financing because contractors and investors must prepare disclosures, preserve records, and reassess compliance. But a congressional request should not be described as a confirmed violation, and a draft bill should not be treated as enacted policy.

Divided government can yield more than paralysis. Some semiconductor supply-chain measures may attract bipartisan support because both parties have endorsed forms of domestic industrial resilience; the CHIPS Act itself was a bipartisan statute, and the 35 percent investment credit was enacted by a Republican Congress in 2025. Electricity affordability is a cross-party concern in places with very different voting histories — it was a Republican governor of Texas who ordered the ERCOT audit and 23 Republican governors who signed the Ratepayer Protection Pledge, and it was Democratic governors in Pennsylvania and California who imposed conditions by executive order and statute. Similarly, an industrially focused House district may have strong reasons to support a specific fabrication project while questioning a neighboring speculative datacenter. The possible coalition is not neatly defined by national party labels. The question for 2027 is whether Congress can negotiate more transparent, performance-based public support without producing uncertainty that delays useful projects.

The incoming governors change this calculation further. Arizona voters reelected Hobbs, while the election brought new administrations to Texas, Ohio, Michigan, Georgia, California, Iowa, Nevada, and Alaska, among other states.^[14] A governor-elect can announce priorities and form a transition team, but official authority begins according to each state’s inauguration schedule — January 11, 2027, in Ohio, for example — and in Texas no statute can change before the legislature convenes in 2027. A federal majority cannot substitute for an individual state statute. These separate clocks — congressional session, gubernatorial inauguration, utility proceeding, environmental review, FERC compliance filing, and project construction — are central to 2027–2030 forecasting.

5.2. CHIPS funding after fiscal 2026: an important hinge, not an automatic cliff

The CHIPS and Science Act established major public support for domestic semiconductor manufacturing and research. The Government Accountability Office summarizes $39 billion for semiconductor manufacturing and related incentives and $11 billion for research and development in the relevant federal program; as of June 2026, Commerce had made awards representing about $31.5 billion in direct funding under the manufacturing incentives, and GAO’s August 2026 report found that awardees had continued to make progress against project milestones.^[134][162] But GAO also found that Commerce had canceled R&D awards worth $7.8 billion of the $11 billion appropriated, that its plan for reestablishing the National Semiconductor Technology Center was insufficiently detailed to demonstrate compliance with statute, and that it had not reestablished the Industrial Advisory Committee; Commerce agreed with all three recommendations and has one year to act.^[134][135] These are reminders that public accountability does not end when an award is announced. Agencies must track conditions, performance, risks, and statutory purposes, and Congress — now under new management — will decide whether to enforce them.

For 2027–2030, lawmakers may consider fresh money, changes in existing program administration, additional research priorities, or more rigorous oversight of the equity-stake model Commerce adopted in October 2026. But an end to scheduled appropriations is not equivalent to instant termination of a previously obligated multiyear project or automatic cancellation of a signed federal assistance agreement. The actual status of an award depends on contracts, appropriations law, milestones, agency decisions, and applicable conditions. A serious account separates available appropriations, obligations, disbursements, and promised private investment; the Congressional Research Service’s July 2026 report on CHIPS-funded fabrication facilities is the appropriate baseline for that accounting.^[163]

The distinction is especially important because semiconductor capital projects operate on longer horizons than most election cycles. Construction can extend past two gubernatorial administrations; tool installation and qualification can continue after a congressional majority changes. Intel’s Ohio schedule — announced under one Congress, rescheduled under a second, and now set to produce under a third — is the case in point. If political officials revise programs without preserving lawful contracts, they may raise the risk premium on every subsequent project. If they refuse to reexamine weak performance simply to avoid controversy, they may waste public resources. A credible bipartisan bargain should protect legitimate reliance while preserving audit rights and performance enforcement.

5.3. The 35 percent semiconductor credit and the December 2026 construction clock

One specific federal financial mechanism is especially consequential, and its deadline falls within weeks of this paper’s completion. The advanced manufacturing investment credit under Internal Revenue Code section 48D, created by the CHIPS Act at 25 percent, was raised by the One Big Beautiful Bill Act of July 4, 2025, to 35 percent of qualified investment for eligible property placed in service after December 31, 2025. The statute’s termination provision was not extended: no credit is available for property whose construction begins after December 31, 2026, with the beginning of construction determined under Treasury’s final regulations by a physical-work test or a 5 percent safe harbor.^{[164][165][166]} The applicable definitions of eligible property, foreign-entity restrictions, and ten-year recapture rules are legally technical; companies should not rely on the calendar date alone to determine eligibility.

The rate has a major theoretical effect. For every hypothetical $10 billion of qualifying investment, a 35 percent credit implies $3.5 billion before accounting for the tax law’s eligibility limitations, timing, recapture risk, financing effects, and actual basis. This is a mathematical illustration, not an assertion that any one TSMC, Intel, or Micron facility qualifies for a $3.5 billion credit. It illustrates why semiconductor developers made, and will continue to make through December, accelerated construction-start decisions in late 2026 even when physical output is years away — and why TSMC’s “early construction” on a fourth Arizona fab and first packaging facility, Micron’s 2026 start on its second Idaho fab, and Intel’s October truss all cluster in the same calendar year.

Here the Arizona paradox acquires a new dimension. A federally supported semiconductor investment faces a powerful incentive to establish a legally qualifying construction start by year-end 2026, while a new Arizona datacenter applicant faces a state tax-exemption pause lasting until mid-2029. The dates create two different economic timetables for adjacent parts of the AI system. One encourages certain eligible manufacturing commitments now; the other suspends a particular incentive for new datacenters for three years. They are governed by different laws and public purposes. Layer Selectivity is thus simultaneously vertical across the stack and horizontal across levels of government. Whether the new Congress extends the 48D construction deadline — a question that will arise almost immediately in 2027, because the credit will otherwise be closed to any fab not yet under construction — is among the first tests of whether bipartisan support for Layer 2 survives the change of majority.

5.4. Federal acceleration versus state fiscal discretion

The Trump administration’s July 23, 2025, AI Action Plan, Winning the Race, emphasized innovation, infrastructure expansion, and international technology leadership across more than ninety policy actions, and declared that “America must continue to be the dominant force in artificial intelligence to promote prosperity and protect our economic and national security.”^[167][168] Its accompanying executive order on accelerating federal permitting of datacenter infrastructure designated qualifying projects for FAST-41 expedited review, directed agencies to develop NEPA categorical exclusions, and instructed the Department of Energy to identify federal sites — Idaho National Laboratory, Oak Ridge, Paducah, and Savannah River among them — for datacenters and associated power.^[16][169] These policies signal federal support for rapid buildout, but they do not automatically create transmission capacity, appropriate new state tax credits, or displace every independent land-use and water-law requirement. A developer may qualify for one federal facilitation pathway and still require utility service, county approvals, or state environmental permits — a point Newsom made explicitly when he contrasted California’s laws with the administration’s “deregulation.”^[106]

The White House has also pursued export-oriented AI policy through its executive order on promoting the export of the American AI technology stack.^[170] That strategy has an internal logic: the United States wants to maintain leadership not merely in manufacturing chips but also in supplying systems, models, and related services internationally. Yet international sales do not resolve domestic infrastructure bottlenecks. If the country can manufacture advanced processors but cannot deploy reliable computing capacity at competitive cost, foreign customers may capture more of the high-margin downstream activity.

There is a legitimate counterargument to selective state restrictions: federal policy should prioritize overall AI capacity because U.S.–China competition is system-wide. Stanford’s 2026 AI Index found that the performance gap between the best American and Chinese models had narrowed to about 2.7 percentage points as of March 2026 even though U.S. private AI investment of $285.9 billion in 2025 dwarfed China’s $12.4 billion — evidence, as one analysis put it, that compute restriction is necessary but not sufficient as a competitive strategy.^[171] A restriction that looks fiscally prudent in one county might reduce domestic model-training options and ultimately weaken national strategic capabilities. But the converse is also plausible. Subsidizing every power-hungry project without sound grid planning can increase household costs, erode public support, and cause a backlash that impedes both chip production and future infrastructure; the 2026 election results are the clearest evidence yet that the backlash is electorally potent. The answer is not that federal or state priorities must always prevail. It is that the costs and national benefits of each project should be made explicit, with federal support reserved for demonstrable spillovers that a host community cannot capture alone.

5.5. Export controls and the problem of the missing domestic customer

Semiconductor policy cannot be separated from China. On January 13, 2026, the Bureau of Industry and Security issued a rule, effective January 15, revising its license-review policy for exports of the NVIDIA H200, the AMD MI325X, and equivalent commercially available chips to China and Macau from a presumption of denial to case-by-case review, subject to certifications that the chips fall below specified performance thresholds, that the exporter’s China volumes do not exceed its U.S. volumes, and that the chips pass through the United States for independent testing — the last requirement paired with a 25 percent Section 232 tariff announced the following day.^{[172][173][174]} Export rules can influence NVIDIA and AMD revenue, customer distribution, fab utilization, and incentives for regional deployment. NVIDIA’s own disclosure that China represented less than 1 percent of datacenter revenue in the quarter ended July 2026, and its guidance assuming no China datacenter compute revenue at all, show how completely the market has reorganized around the controls.^[121] They do not determine whether a particular state will approve a datacenter or whether a customer can actually use a delivered chip.

The strategic logic of domestic fabrication rests partly on reducing exposure to geopolitical disruption. TSMC’s Arizona program can diversify certain fabrication capacity relative to concentrated overseas production. But export controls can reshape the buyer market, while industrial policy alters the manufacturer market and datacenter policy alters the deployment market. National leaders therefore have three overlapping questions: where the chip is manufactured, to whom it may legally be sold, and where the resulting computing system can economically operate. A rule addressing one of those questions can create new costs in the others.

Consider two hypothetical outcomes. In the first, domestic chip manufacturing grows and abundant domestic datacenter projects become operational under fair infrastructure tariffs. U.S. model developers gain resilient access to compute and application companies gain competitive inference services. In the second, manufacturing grows but domestic service capacity is severely delayed and international restrictions prevent certain foreign sales. A chip company can encounter weaker near-term addressable demand even though the industrial rationale for domestic manufacturing remains strong. Neither is a forecast; both identify interactions that should be stress-tested before governments celebrate or restrict individual layers.

5.6. Scenario A, 2027–2030: selective abundance

In the first scenario, Layer Selectivity becomes more precise and economically productive. Congress preserves credible semiconductor research and manufacturing support subject to enforceable milestones, extends or replaces the 48D construction window with performance conditions, and legislates something resembling the Ratepayer Protection Act so that cost-causation becomes a national floor rather than a voluntary pledge. Arizona maintains its semiconductor ecosystem while designing an evidence-based successor to the temporary datacenter exemption pause. Virginia refines usage-related payments to protect ordinary ratepayers without double-charging for costs already recovered. Texas and PJM improve load verification and interconnection coordination, and FERC’s six show-cause proceedings produce consistent large-load definitions. Municipalities establish predictable development standards rather than ad hoc project-by-project surprises.

Under this scenario, private investors respond to a stable public bargain. They accept meaningful charges where they cause measurable costs because the rules are transparent and useful projects can proceed. New energy generation, storage, grid equipment, and workforce capacity expand alongside computing. Some datacenters move to lower-cost regions; others remain near demand centers. Chip manufacturing grows where supplier networks and skills are strongest. Model developers gain multiple credible deployment options, and application businesses gain affordable inference.

The indicators of success would be concrete: more projects reaching commercial operation, narrower divergence between requested and actual load, fewer unexpected utility charges to existing customers, improving semiconductor production qualifications, growing workforce retention, and higher measured productivity from applications. An abundance scenario is not defined by unchecked building. It is defined by productive capacity expanding without losing public consent.

5.7. Scenario B, 2027–2030: selective fragmentation

In the second scenario, states pursue defensible local goals but fail to coordinate. Tax preferences and large-load rates change abruptly, and moratoria of the New York type proliferate without sunset conditions. Utilities cannot distinguish credible projects from speculative reservations. Federal reviews move faster while state and local approvals become inconsistent. Some jurisdictions overinvest in generation after inflated demand forecasts; others cannot deliver power to projects with real customers. Companies with experienced legal departments and preexisting sites adapt, while independent entrants face high transaction costs.

The resulting physical AI economy is uneven. Domestic fab announcements continue because they carry national political prestige, but supporting packaging, electric infrastructure, and workforce programs lag. Datacenter capacity concentrates in a smaller number of jurisdictions and around firms that secured early grid access. Models and agents become more expensive to deploy for smaller companies, and public reaction to electricity prices hardens. Layer Selectivity ceases to be an evidence-based allocation tool and becomes an unpredictable series of political exceptions.

The warning signs would include a rising ratio of announced to completed projects, large unutilized interconnection reservations, repeated changes to eligibility criteria, growing delays after major award announcements, widening electricity-cost disparities, and reduced entry by smaller compute providers. The key danger is not that every state imposes identical restrictions. It is that uncertainty and misallocated investment reinforce each other.

5.8. Scenario C, 2027–2030: supply-chain success, deployment shortage

The third scenario is more counterintuitive. The United States achieves meaningful progress in domestic fabrication, memory, and packaging while its Layer 3 capacity expands more slowly than demand for models and agents. Successful manufacturing policy therefore exposes the limitation of a narrow industrial strategy. The chips can be made, but domestic generation, transmission, datacenter commissioning, fiber, and local approvals remain the binding constraints for deploying them at scale.

Under this scenario, chip firms may still sell into international markets or other domestic industries, and manufacturing success would remain valuable. But frontier-model firms could increasingly purchase compute where infrastructure is more readily available, and American application companies could face higher inference costs or more concentrated suppliers. For national security, the irony would be severe: a country might reduce dependence on imported advanced wafers while becoming more reliant on offshore deployment facilities or fragile third-party capacity arrangements. World Bank President Ajay Banga’s observation that the largest AI datacenters will remain in rich-country grids because “you need electricity, lots of it, more than what the developed world also has in some cases” is usually read as a warning for emerging markets; in Scenario C it becomes a warning for the United States as well.^[175]

This scenario argues against treating layers as substitutable. More fabs do not generate electricity; more datacenters do not automatically produce leading-edge chips; a powerful model does not guarantee that applications can use it safely and economically. A resilient AI economy requires sufficient capability at all five stages, even if different states specialize in different activities. The question is how to accomplish that specialization without abandoning the public interests that motivated Layer Selectivity in the first place.

5.9. A proposed Five-Layer Public Investment Scorecard

To turn the theory into a usable instrument for governors, legislators, and corporate boards, this paper proposes a Five-Layer Public Investment Scorecard. It would accompany major incentives and infrastructure approvals, not replace formal statutory analysis or engineering review. For each project, the responsible authority would publish the following information where confidentiality and security permit:

Layer and output: the project’s primary activity, measurable productive output, customers, and relation to the other four layers.
Additionality: the reason public support is expected to cause investment that would not occur in the same form without it.
Public exposure: grants, tax preferences, discounted services, public land, infrastructure spending, and contingent obligations expressed in present value.
Resource accounting: contracted power, forecast coincident peak, electricity consumption, water withdrawal and consumption, emissions, land, and interconnection costs.
Durable returns: permanent jobs, worker development, local suppliers, fiscal receipts, and national strategic capability, with attribution and counterfactual assumptions.
Execution milestones: permits, construction, equipment installation, testing, commercial operation, and operational performance.
Remedies: clawbacks where lawful, bonds, security deposits, minimum payments, transparent reporting, and treatment of project abandonment.

Scorecard element	Illustrative Layer 2 metric (fab)	Illustrative Layer 3 metric (hyperscale datacenter)
Output	Qualified wafer starts per month by node; yield parity with home fabs	Energized IT megawatts; contracted versus actual utilization
Additionality	Evidence that siting depended on incentive (e.g., competing-site record)	Evidence of relocation versus “would build anyway”; land/power as primary drivers
Public exposure	48D credit, CHIPS award, state credits, infrastructure, water contract	Sales/use-tax exemption value, discounted tariff, utility upgrades socialized
Resource accounting	Ultrapure water demand, reclamation rate, firm power requirement	Coincident peak, annual MWh, onsite and offsite water, backup-generator emissions
Durable returns	Permanent manufacturing and engineering jobs; supplier formation	Permanent operations jobs; property-tax base; local infrastructure funded
Milestones	Shell complete → tool-in → qualification → volume shipments	Permit → interconnection agreement → energization → commercial operation
Remedies	Milestone-based disbursement; clawback on abandonment	Minimum-bill or take-or-pay contracts; security deposits; curtailment terms

The scorecard should be designed to compare projects within their relevant categories as well as across the stack. A semiconductor fab cannot be fairly ranked against a model-safety laboratory simply by dividing announced dollars by employee count. Strategic resilience, knowledge spillovers, and public safety deserve different weights, and different public authorities may value them differently. The purpose of a common scorecard is to reveal assumptions and tradeoffs, not pretend that every activity can be condensed to a universal number. Several of its elements already exist in fragments: Pennsylvania’s GRID consent orders supply the remedies, California’s AB 2619 and SB 886 supply the resource accounting, Texas’s Batch Zero supplies the milestone discipline, and Arizona’s HB 2756 supplies the cost-of-service transparency. What no state has yet done is assemble them into one instrument applied evenly across layers.

5.10. A practical roadmap for governors and Congress

For 2027, the immediate tasks are transparency and baseline accounting. Governors should request public inventories of major energy-intensive projects, distinguishing prospective from contracted loads and privately financed from publicly supported infrastructure. Legislatures should publish the actual terms, fiscal costs, and expiration conditions of incentives — the Good Jobs First finding that only 11 of 36 subsidizing states disclose recipients is an indictment that a single statute per state could cure. Utilities should explain how they recover the incremental costs of service and why projections differ. Congress should examine the status of CHIPS obligations, the equity-stake model, research programs, and the 48D construction deadline without equating a headline investment commitment with delivered industrial capacity.

For 2028, the task is implementation quality. States can compare actual electricity use with forecasts, assess water and workforce programs, publish project completion rates, and review whether incentive requirements have unintentionally increased barriers to new entrants. Virginia’s consumption tax expires on June 30, 2028, and its renewal debate will be the first full test of a use-based instrument against actual revenue and load data. Where governments condition benefits on performance, measurements should be independently auditable and remedies proportionate. Where public data would reveal export-controlled information or security-sensitive facility details, agencies can release aggregated results or confidential oversight summaries rather than abandon accountability entirely.

For 2029, Arizona’s scheduled datacenter exemption pause presents a natural policy checkpoint. Lawmakers should evaluate the 107 applications approved before the pause, facilities built during it, actual tax collections, grid upgrades, water effects, economic benefits, and evidence of projects moving elsewhere. The choice need not be a binary return to the earlier blanket program or a permanent prohibition. It could be a redesigned exemption tied to credible power and infrastructure contributions, a narrower equipment category, a transparent cap, a Virginia-style consumption charge, a Pennsylvania-style conditional track, or a sunset with periodic review. Any proposal should be modeled for legal feasibility and administrative costs before adoption.

For 2030, the test is system-wide resilience. Have U.S. semiconductor projects reached qualified production — TSMC’s third and fourth Arizona fabs, Micron’s Idaho and New York fabs, Intel’s first Ohio module? Are grid investments productive? Can frontier-model companies obtain competitive domestic compute? Do startups and established firms both have routes into the market? Are agents delivering measurable benefits in hospitals, factories, government, transportation, and small businesses? Has the public received a defensible return for the concessions made? A state might be proud of its largest single investment while failing these broader tests. Layer Selectivity demands an assessment of the whole stack.

The overarching recommendation is to use conditional, measurable, and time-consistent industrial policy. Federal officials should support genuine national spillovers, states should protect local fiscal and resource interests, utilities should assign network costs credibly, and firms should be held to their actual delivery commitments. The tradeoffs will remain political. The contribution of the Five-Layer framework is to make them sufficiently visible that citizens and investors can judge the choices intelligently.

5.11. What would disprove the Layer Selectivity thesis?

A serious framework must state not only what could confirm it but also what evidence could weaken it. The first possible challenge is additionality. Suppose Arizona’s datacenter tax-exemption pause has little effect on actual investment, and nearly all credible projects proceed with unchanged timelines and capacity — an outcome the 107 late approvals make quite plausible for the next several years. That would suggest the exemption was less decisive than expected, at least for projects with strong customers and power access. In such a case, the policy could still protect public revenue, but its effect on infrastructure geography would be modest. Conversely, if the pause sharply reduces investment but yields no measurable improvement in local water or electricity conditions, lawmakers should ask whether the chosen instrument addressed the actual problem.

The second challenge is industrial spillover evidence. Supporters of semiconductor subsidies often assume that fabrication creates enduring clusters of engineers, suppliers, and technical knowledge. If the promised facilities remain delayed, local suppliers fail to establish meaningful business, or specialized workers leave after construction, the anticipated public return may be weaker than argued. A government might still support manufacturing for national-security reasons, but it should then say so plainly and seek appropriate national financing rather than disguise the objective as an exceptional local fiscal return. Similarly, the evidence may show that selected datacenters generate larger local property-tax receipts, infrastructure contributions, and construction employment than politically favored manufacturing projects — JLARC’s finding that Virginia datacenters paid $733 million in local property taxes in 2024 is the kind of figure that could tip the comparison.^[176] Layer Selectivity would then require reconsideration, not loyalty to a predetermined hierarchy.

A third test is system-wide productivity. If regions with stricter datacenter standards nevertheless achieve lower inference costs, higher reliability, and more application adoption because they require efficient infrastructure investment, that would support the proposition that transparent conditionality can improve economic performance. If rules instead raise prices, delay useful installations, and concentrate compute among dominant firms without commensurate public gains, the same framework would diagnose a failure of design. Neither result should be attributed to policy alone without controlling for electricity resources, demand growth, networking, tax differences, and preexisting industrial clusters.

The research should therefore construct a state-year and project-level evidence base for 2027–2030. It would track proposed facilities from first announcement through certification, permits, grid connection, operational status, actual payroll, and measured energy and water use. The analysis should compare projects approved just before and just after policy thresholds — the June 30, 2026, Arizona cutoff is an unusually clean one — but avoid treating them as automatically identical. Their sponsors may differ in size, creditworthiness, industry, and site readiness. Econometric studies could use matched projects, event studies, or carefully designed difference-in-differences comparisons, subject to the assumptions of those methods; the CEPR and Rutgers studies cited in Section 2 show both the promise and the fragility of such designs. Qualitative interviews with utilities, labor organizations, landowners, suppliers, and local communities would complement the numbers.

Finally, the framework must be alert to displaced effects. A tax change that moves a datacenter from Arizona into a neighboring state may reduce Arizona’s projected resource burden without reducing total regional electricity demand. A state that welcomes a semiconductor fab may import materials produced under less transparent environmental standards elsewhere. A hyperscaler may report lower onsite water use while buying electricity from water-intensive generation. The United Nations University’s finding that only 32 countries host AI-specialized datacenters and that more than 90 percent of global AI compute sits in just two of them is a reminder that displacement operates internationally as well as between states.^[55] An honest evaluation therefore distinguishes local, regional, national, and global consequences and guards against counting the same economic benefit multiple times. The framework succeeds only if it improves the accuracy of public choices; it should be abandoned or revised wherever evidence demonstrates that it systematically misprices value or risk.

Section 6: What Have We Learned? Eight Pillars of Layer Selectivity

Pillar 1 — The AI economy is one system but not one fiscal category

The first lesson is that a unified technology ecosystem can contain dramatically different local economic bargains. Energy facilities, fabrication plants, datacenters, model laboratories, and applications are connected through investment and operating decisions, yet they differ in labor intensity, physical footprint, technology risk, strategic importance, and the location of their benefits. An identical tax preference for every activity is not automatically fair or efficient. Neither is a universal presumption that one layer deserves preference. Arizona’s semiconductor expansion and datacenter exemption pause reveal why governments must look through the umbrella label artificial intelligence to the underlying business activity — and New York’s simultaneous datacenter ban and multibillion-dollar memory-fab partnership show the same instinct carried to its limit.

The Five-Layer AI Economy is therefore an analytical discipline. It prevents a lawmaker from confusing the economic effects of fabricating an advanced wafer with the effects of using an accelerator in a server. It also prevents the opposite error: failing to recognize that the chip factory’s long-term customers require viable datacenters, models, and applications somewhere. Public support can be different across layers while remaining coordinated across them. The correct question is not whether a state loves or opposes AI. It is which precise public commitments are warranted, at what cost, and with what evidence.

Pillar 2 — The scarce resource is credible capacity, not announced dollars

The second lesson concerns measurement. Investment announcements are important signals, but they are not expenditures: TSMC’s $265 billion is a multi-decade intention against a company-wide 2026 capital budget of $60 billion to $64 billion. Applications for tax relief are not certified projects, and certified projects are not energized halls — Arizona’s 113 applications are the case in point. Interconnection requests are not delivered megawatts: ERCOT’s 474 to 498 gigawatts of requests against roughly 65 gigawatts of screened base load is the case in point. A projected workforce is not a current payroll. A qualified process node is not the same as a proposed cleanroom, and a structural truss in New Albany is not a wafer. These distinctions may look pedestrian, but the AI boom makes them central to financial stability and public trust because project numbers have grown so large that small assumptions translate into billions of dollars and years of infrastructure commitments.

For state officials, credible capacity means facilities with financing, lawful permits, viable supply contracts, executable construction schedules, and ultimately productive output. For a utility, it means a verifiable load that can be served reliably and charged according to its actual system impact. For semiconductor firms, it means installed and qualified manufacturing, memory and packaging capability, and satisfied customers. For model companies, it means sustained useful computation at an economically viable price. Incentive programs should track the transition from promise to proof instead of celebrating each stage as though it were the final achievement.

Pillar 3 — Public costs must follow their causes, while public benefits should follow their evidence

The third lesson is about allocation. A community that hosts a large facility should not have to finance every new substation or absorb water risks simply because the investor describes itself as strategically important. Nor should a corporation be charged for network expansion that primarily benefits other customers. Utility rates should reflect defensible cost causation; environmental conditions should reflect measured impacts; fiscal subsidies should reflect documented incremental public returns. Virginia’s consumption tax, Texas’s large-load verification, Pennsylvania’s conditional consent orders, California’s cost-allocation statutes, APS’s extra-high-load-factor tariff, and Arizona’s reassessment of datacenter tax relief demonstrate different attempts to confront these issues — and the Harvard Electricity Law Initiative’s catalogue of nearly fifty rate proceedings shows how often the default outcome, absent such attempts, is socialization.

“Without systematic changes to prevailing utility ratemaking practices, the public faces significant risks that utilities will take advantage of opportunities to profit from new data centers by making major investments and then shifting costs to captive ratepayers. The industry’s current approaches of luring data centers with discounted contracts or lopsided tariffs is unsustainable.”

— Eliza Martin and Ari Peskoe, Harvard Law School Electricity Law Initiative, Extracting Profits from the Public ^[177]

Evidence can point in different directions. A datacenter may finance generation, create a substantial taxable asset, supply flexible load, or anchor a rural economic-development strategy. A fab may create enduring industrial expertise but also impose significant power and water obligations. A model laboratory may employ highly skilled workers while depending on compute outside the host state. The principle is not to minimize every burden at any price. It is to make the tradeoff explicit and assign it to parties that create, benefit from, or can manage the risk.

Pillar 4 — Elections change negotiating authority before they change physical infrastructure

The fourth lesson is institutional. The November elections changed Congress and several governorships, but power plants, fabrication facilities, and datacenter campuses cannot be moved by an election-night projection. New leaders will inherit laws, contracts, regulatory proceedings, unfinished construction, and investment commitments made before their victories. Governor-elect Hinojosa inherits SB 6 and the Batch Zero audit; Governor-elect Acton inherits Intel’s 2030 schedule; Governor-elect Becerra inherits seven statutes signed in September; the new Congress inherits a 48D deadline that expires before it convenes. Some outcomes can change through executive administration; others require legislatures, commissions, courts, utility boards, or federal agencies. The pace of institutional change is therefore different from the pace of market repricing — Alphabet lost a tenth of its value in a day on a capex disclosure, while the Texas legislature will not meet until 2027.

This difference creates both risk and opportunity. Abrupt, legally uncertain reversals increase costs and weaken the public’s credibility as a partner. Transparent review of existing programs can improve them. Congressional oversight may reveal whether CHIPS grants and research initiatives are working. Governors can ask harder questions about local electricity bills and workforce needs. Local governments can create predictable standards. A functioning democracy has a legitimate role in revising its industrial bargain; the challenge is to do so lawfully, prospectively where appropriate, and with an honest inventory of commitments already made.

Pillar 5 — National AI leadership requires coordination across all five layers

The fifth lesson is strategic. The United States cannot achieve durable leadership simply by maximizing one component of the stack. More manufacturing capacity will not substitute for affordable power, secure grids, model research, or application adoption. More datacenter floor space will not substitute for chips, high-bandwidth memory, networking, or innovative software. More powerful models will not automatically produce value without capable employees, trustworthy institutions, productive enterprises, and ways for ordinary businesses to afford deployment. The Five-Layer AI Economy is a system of complements, even though some investments compete for scarce resources.

The 2027–2030 opportunity is to create coordinated selectivity: specialize geographically, measure local costs honestly, support nationally important capabilities, and keep pathways open from one layer to another. A state may attract a fab while limiting a datacenter tax subsidy, but the nation must still ensure enough data-processing capacity is built under responsible terms. A state may favor applications and agentic services while relying on another region for energy and compute, but it should recognize its dependence and support resilient supply. The goal is not uniformity. It is a balanced, contestable, and economically productive system capable of adapting as AI technology changes.

Pillar 6 — Transparency is the precondition of every other pillar

The sixth lesson is procedural, and it emerged from the research for this paper more forcefully than any other. Nearly every quantitative dispute surveyed here — whether datacenters raise residential bills, whether the Arizona exemption was additional, whether Georgia’s incentive costs $400 million or $2.5 billion, whether a state’s queue represents 65 or 498 gigawatts — turns on data that is either undisclosed or disclosed too late to inform the decision. Good Jobs First found that most subsidizing states do not reveal which companies benefit; Texas underestimated its exemption cost by a factor of eight; Pennsylvania discovered that most of its hundred proposals had no permit application behind them; and the Harvard researchers found that utilities had in some cases concealed how much residential customers paid to offset special rates. Pennsylvania’s prohibition of nondisclosure agreements, California’s disclosure statutes, Arizona’s HB 2756 reporting, and the proposed federal Data Center Tax Accountability and Disclosure Act are all responses to the same deficiency. Layer Selectivity without disclosure is simply favoritism with a theory attached. Disclosure — of applications, certifications, contracts, loads, water, and fiscal cost — is not one policy option among many; it is the condition under which the other seven pillars can be tested at all.

Pillar 7 — Time horizons must be matched, not averaged

The seventh lesson concerns duration, and it is the lesson Ohio and Arizona teach jointly. Layer 1 assets last half a century; a leading-edge fab is built for several process generations; a datacenter shell may outlive three or four generations of accelerators; a frontier model is obsolete in eighteen months; an application may be rewritten in a quarter. A 48D credit that expires on a calendar date, a state exemption that runs for ten or twenty years, a consumption tax that sunsets in two, a datacenter ban that lasts one, and a pause that lasts three are all bets about which horizon matters. Policy fails when it averages these horizons into a single term or when it borrows the prestige of the long-lived asset to justify concessions to the short-lived one. The disciplined approach is to match each public commitment to the useful life of the local investment it induces and to the government’s realistic ability to recover its cost: milestone-based disbursement for fabs whose output is years away, use-based or performance-based charges for datacenters whose equipment turns over quickly, and sunset-and-review clauses for every instrument, so that 2028 and 2029 become scheduled decisions rather than accidental cliffs.

Pillar 8 — Public consent is a layer-specific resource and it is now binding

The eighth and final lesson is political, and 2026 established it beyond reasonable dispute. Seven in ten Americans oppose an AI datacenter in their community, a higher share than oppose a nuclear plant; candidates won governorships in Texas, Ohio, and Georgia on datacenter moratoria; twenty of thirty-six gubernatorial contests featured pause proposals; a Republican governor audited his state’s entire queue; a Democratic governor who had vetoed disclosure in 2025 signed seven bills in 2026; and the national election analysis identified datacenter backlash as a motive force in a congressional landslide.^{[34][15][159]} At the same time, no comparable movement has arisen against semiconductor fabs; TSMC’s expansion was greeted in Phoenix with the language of a “gold rush,” and Intel’s Ohio truss was celebrated rather than protested. Public consent, in other words, is not evenly distributed across the five layers, and it has become a binding constraint on Layer 3 in a way it is not on Layer 2. A framework that ignored this asymmetry would be describing a different country. Layer Selectivity is partly a response to that asymmetry — a way for governments to continue building the system their economies require while channeling the layer that lacks consent toward terms the public will accept. Whether consent can be rebuilt depends on whether the other seven pillars are honored; where costs are visibly borne by their causes and benefits are visibly measured, opposition has proved negotiable, and where they are not, it has proved decisive at the ballot box.

Conclusion: Why “Layer Selectivity” Names the Next Phase of America’s AI Industrial Policy

The Arizona story begins with a clock and an announcement. In June 2026, datacenter developers raced to submit 113 tax-exemption applications in sixteen days before a three-year pause took effect, and the state approved 107 of them. In July, TSMC announced four more advanced fabs and another $100 billion in planned Arizona investment, raising its cumulative stated commitment to $265 billion. In October, TSMC reported a $46.7 billion quarter, NVIDIA had already guided to $108 billion for the quarter it is now completing, and Microsoft, Alphabet, Amazon, and Meta together confirmed roughly $730 billion of 2026 capital spending. In November, Governor Katie Hobbs won reelection with a legislative majority behind her, as the electorate handed Congress to the opposition and elected governors in Texas, Ohio, and Georgia who had promised to pause the very facilities Arizona had declined to subsidize. Each event is clear in the record; none establishes by itself that every proposed factory or datacenter will be built. Taken together, however, they reveal an important change in the way American governments are negotiating with the AI industry.^{[1][4][116][12][15]}

The early AI infrastructure boom was frequently narrated as a race for everything: more chips, more megawatts, more land, more servers, more models, and more investment. That race remains real — the earnings reports of the third quarter of 2026 leave no doubt that demand exceeds supply at every layer from wafers to inference. But public officials increasingly confront its distributional consequences. Who receives the durable jobs? Who pays for new transmission? What is the opportunity cost of water used by an industrial project? How much of a large tax exemption represents investment that would have occurred anyway? Does a public subsidy create new national strategic capability, or merely increase an established company’s financial return? Do restrictions protect communities, or do they entrench incumbents that acquired favorable sites earlier?

These questions are not attacks on artificial intelligence. They are the normal questions of political economy entering a sector whose expansion has been so rapid that fiscal and regulatory institutions are struggling to catch up. The difficult choice is between indiscriminate encouragement and intelligent differentiation, not between innovation and stagnation. A state can welcome leading-edge semiconductor fabrication and still insist that new datacenter applicants demonstrate an adequate public return for special tax treatment. It can support electricity generation while charging large users their attributable network costs. It can promote frontier models while requiring credible safety and accountability practices. None of these decisions is self-justifying; each must be tested against evidence and alternatives.

This is why I chose the title Layer Selectivity. It is sufficiently specific to explain Arizona’s 2026 contrast and sufficiently general to describe an emerging national regime that now includes a Virginia tax, a Texas audit, a Pennsylvania consent order, a California disclosure code, and a New York ban. The term does not assume that Layer 2 deserves permanent preference over Layer 3. Instead, it names the government’s act of distinguishing investments across the five layers and allocating different incentives, conditions, and obligations according to their consequences. Its explanatory value comes from refusing to flatten the AI economy into one political slogan.

The Five-Layer AI Economy supplies the structure behind the term. Layer 1 delivers the physical energy system; Layer 2 manufactures and supplies computing machinery; Layer 3 makes scalable computation available; Layer 4 develops models; and Layer 5 converts machine intelligence into useful work. A policy decision in one layer travels into the costs, opportunities, and bargaining positions of the others. A state that understands only the project within its boundaries may miss the national opportunity; a national government that ignores local costs may lose the legitimacy needed to build the system at all.

Looking toward 2030, the most successful American jurisdictions will probably not be those that offer every conceivable subsidy or those that erect the highest barriers. They will be those capable of choosing openly, measuring carefully, coordinating across institutions, and delivering what they promise. Investors will reward execution, communities will demand evidence, and technological competition will continue to expose weaknesses in disconnected strategies. Arizona’s apparent contradiction is therefore not a historical curiosity. It may be an early illustration of the political and economic framework through which the next generation of America’s AI infrastructure is negotiated.

The enduring proposition is simple but consequential: the AI economy must grow as a system, even when governments support its layers differently. Layer Selectivity gives that proposition a name — and creates a research agenda for policymakers, corporations, investors, and scholars who must make it work.

Endnotes and Source Documentation:

[1] AZFamily (Arizona’s Family). “Arizona Saw a Surge in Data Center Tax Break Requests. Here’s Why.” November 6, 2026. Reports 113 applications June 15–30, 2026, 107 approved, with comments from Joseph Palomino (Arizona Center for Economic Progress) and Kasia Tarczynska (Good Jobs First). https://www.azfamily.com/2026/11/06/arizona-saw-surge-data-center-tax-break-requests-heres-why/

[2] Axios Phoenix. “Two-Week Surge in Data Center Tax Break Applications Rivals 13-Year Total.” July 8, 2026. Reports 113 applications June 15–30 versus 123 from August 2013 to June 14, 2026, and Data Center Coalition response. https://www.axios.com/local/phoenix/2026/07/08/data-center-tax-break-applications-double-moratorium

[3] KJZZ. “Agency Received Flood of Applications Ahead of Pause on Arizona Data Center Tax Breaks.” July 14, 2026. Describes the 2013 Computer Data Center Program terms (10- and 20-year exemptions through 2033) and 83 beneficiary developments. https://www.kjzz.org/business/2026-07-14/agency-received-flood-of-applications-ahead-of-pause-on-arizona-data-center-tax-breaks

[4] City of Phoenix, Community and Economic Development. “TSMC Announces Additional $100 Billion Investment in Arizona.” July 16, 2026. Includes statements by Governor Katie Hobbs and Mayor Kate Gallego. https://www.phoenix.gov/newsroom/ced-news/tsmc-announces-additional–100-billion-investment-in-arizona.html

[5] Engineering News-Record. “TSMC Arizona Megaproject Grows to $265 Billion.” July 22, 2026. https://www.enr.com/articles/63349-tsmc-arizona-megaproject-grows-to-265-billion

[6] Tom’s Hardware. “TSMC Commits Another $100 Billion to Arizona for at Least Four More 2nm Fabs.” July 16, 2026. Discusses fab module costs, the U.S.–Taiwan trade agreement, and labor, water, and visa constraints. https://www.tomshardware.com/tech-industry/tsmc-commits-another-100-billion-to-arizona-for-at-least-four-more-2nm-fabs

[7] Taiwan Semiconductor Manufacturing Company. “TSMC Intends to Expand Its Investment in the United States to US$165 Billion to Power the Future of AI.” Form 6-K, March 4, 2025. https://www.sec.gov/Archives/edgar/data/1046179/000104617925000024/tsmcexpandinvestmentintheu.htm

[8] Wikipedia. “TSMC Arizona.” Accessed November 2026; summarizes investment chronology and Apple’s February 2026 commitment to purchase more than 100 million Arizona-made chips in 2026. https://en.wikipedia.org/wiki/TSMC_Arizona

[9] Converge Digest. “TSMC Posts $46.72B Q3 Revenue, Reaffirms $60–64B CapEx.” October 15, 2026. https://convergedigest.com/tsmc-q3-2026-capex-taiwan-arizona-japan/

[10] MultiState. “Arizona 2026 Election Results.” Reports Hobbs’s reelection and the first Democratic legislative trifecta since 1966. https://www.multistate.us/elections/results/arizona-2026

[11] KTAR News. “Katie Hobbs Defeats Andy Biggs, Wins Second Term as Arizona Governor.” November 4, 2026. https://ktar.com/arizona-politics-news/election/katie-hobbs-win-andy-biggs-race-arizonas-governor

[12] NBC News. “U.S. Senate Election 2026 Live Results: Democrats Win Majority.” https://www.nbcnews.com/politics/2026-elections/senate-results

[13] CBS News. “2026 Midterm Election Updates as Fallout Begins Over Sweeping Democratic Victories.” November 2026. https://www.cbsnews.com/live-updates/election-day-2026-midterms/

[14] MultiState. “Governors — 2026 State Election Results.” Lists winners in all 36 gubernatorial races and party flips. https://www.multistate.us/resources/governors-2026-state-election-results

[15] FourWeekMBA. “AP Calls 4 of 5 Data Center Governor Races; Cost Terms Vary.” November 4, 2026. Compiles data-center positions of Hinojosa, Bottoms, Acton, and Abbott, and counts pause proposals in 20 of 36 governor races. https://fourweekmba.com/ai-ap-calls-4-of-5-data-center-governor-races-cost-terms-vary/

[16] The White House. “Accelerating Federal Permitting of Data Center Infrastructure.” Executive Order, July 23, 2025. https://www.whitehouse.gov/presidential-actions/2025/07/accelerating-federal-permitting-of-data-center-infrastructure/

[17] IEEE Spectrum. “Stanford’s AI Index for 2026 Shows the State of AI.” April 2026. Summarizes Epoch AI compute data in the Stanford HAI 2026 AI Index. https://spectrum.ieee.org/state-of-ai-index-2026

[18] Site Selection Magazine. “AI & the Workforce: How AI Is Changing the Future of Work.” Quoting Mark Muro, Brookings Metro, on the geography of AI readiness. https://siteselection.com/ai-the-workforce-how-ai-is-changing-the-future-of-work/

[19] Server Country. “Arizona Data Center Policy.” Summarizes Laws 2026, Chapter 140 (HB 4168), HB 2756, and application status. https://servercountry.org/policy/arizona/

[20] Eliza Martin and Ari Peskoe. Extracting Profits from the Public: How Utility Ratepayers Are Paying for Big Tech’s Power. Harvard Law School Electricity Law Initiative, March 2025. https://eelp.law.harvard.edu/wp-content/uploads/2025/03/Harvard-ELI-Extracting-Profits-from-the-Public.pdf

[21] Virginia Joint Legislative Audit and Review Commission. Data Centers in Virginia. Report 598, December 9, 2024. https://jlarc.virginia.gov/pdfs/reports/Rpt598.pdf

[22] George Mason University, Center for Assurance Research and Engineering. “STEP Brief 05: Data Centers.” September 3, 2026. Tabulates JLARC 2024 and 2025 findings (74,000 jobs; 59,000 construction; 15,000 operations; ~$0.48 return per $1 exempted; 11.3 GW capacity) and the 2026 electricity tax. https://care.gmu.edu/wp-content/uploads/STEP_Brief_05_Data_Centers_2026-09-03.pdf

[23] Investigative Post. “Subsidies for STAMP Data Center Among Largest Ever.” March 18, 2026. Quoting Kasia Tarczynska, Good Jobs First. https://investigativepost.org/2026/03/18/subsidies-for-stamp-data-center-among-largest-ever/

[24] Data Center Dynamics. “Research: US State Subsidies Pay $2 Million per Data Center Job.” Summarizing Good Jobs First, Money Lost to the Cloud. https://www.datacenterdynamics.com/news/research-us-state-subsidies-pay-2-million-per-data-center-job/

[25] Tufts University, Fletcher School. “Culture, Costs Unlikely to Be Decisive for TSMC in Arizona, Academic Says.” Interview with Chris Miller. https://sites.tufts.edu/fletcherrussia/culture-costs-unlikely-to-be-decisive-for-tsmc-in-arizona-academic-says/

[26] Global Finance Magazine. “Chip Beef.” Interview with Chris Miller, Tufts University, on the CHIPS Act and Taiwan risk. https://gfmag.com/economics-policy-regulation/semiconductor-chip-industrial-policy-sustainability/

[27] Sarah J. Smith, Alex Hubbard, Alex Newkirk, Mohan Ganeshalingam, Billie Holecek, Dale Sartor, Michael Mills, and Arman Shehabi. United States Data Center Energy Usage Report: 2025 Update. Lawrence Berkeley National Laboratory for the U.S. Department of Energy, LBNL-2001758, June 2026. https://www.energy.gov/documents/united-states-data-center-energy-usage-report-2025-update

[28] Lawrence Berkeley National Laboratory, Center of Expertise for Data Center Energy. “Modeling and Forecasting.” Summarizes the 2025 Update (11.8% central, 9.5–15.3% range for 2030) and the 2024 Report (6.7–12% by 2028). https://datacenters.lbl.gov/modeling-forecasting

[29] International Energy Agency. Key Questions on Energy and AI — Executive Summary. World Energy Outlook Special Report, 2026. https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary

[30] International Energy Agency. Energy and AI — Executive Summary. World Energy Outlook Special Report, 2025. https://www.iea.org/reports/energy-and-ai/executive-summary

[31] International Energy Agency. “AI Is Set to Drive Surging Electricity Demand from Data Centres While Offering the Potential to Transform How the Energy Sector Works.” Press release, April 2025. https://www.iea.org/news/ai-is-set-to-drive-surging-electricity-demand-from-data-centres-while-offering-the-potential-to-transform-how-the-energy-sector-works

[32] Global Data Center Hub. “ERCOT Data Center Interconnection Pause — Docket 59220.” August 2026. Reports 1,800+ large-load requests totaling ~474 GW, ~90% data centers, and the 250–300 project audit due December 10, 2026. https://www.globaldatacenterhub.com/p/ercot-data-center-interconnection-pause-docket-59220

[33] Baxtel. “Texas Grid Audit Pauses New Data Center Approvals Across ERCOT.” August 6, 2026. Reports 498 GW across 688 requests and 65 GW across 150 requests preliminarily eligible as base load. https://baxtel.com/news/texas-grid-audit-pauses-new-data-center-approvals-across-ercot

[34] Gallup (Jeffrey M. Jones). “Americans Oppose Data Centers in Their Area.” May 13, 2026. March 2–18, 2026 survey: 71% oppose, 48% strongly; 53% oppose a nearby nuclear plant. https://news.gallup.com/poll/709772/americans-oppose-data-centers-area.aspx

[35] Yahoo Finance / AFP. “Taiwan Chipmaker TSMC to Invest Another US$100 bn in Arizona Fabs.” July 16, 2026. Quotes C.C. Wei and CFO Wendell Huang on the $60–64 billion 2026 capex budget. https://finance.yahoo.com/technology/ai/articles/taiwan-chipmaker-tsmc-invest-another-065311994.html

[36] Metro Phoenix. “TSMC Accelerates AZ Production Amid Soaring Earnings.” October 2026. Reports completion of the second Phoenix fab and early construction of a fourth fab and first packaging facility. https://metrophoenix.com/2026/10/tsmc-accelerates-az-production-amid-soaring-earnings/

[37] Arizona Commerce Authority. “ACA and U of A Semiconductor Nano Fabrication Center.” February 17, 2026. https://www.azcommerce.com/news-events/news/2026/2/aca-and-u-of-a-semiconductor-nano-fabrication-center/

[38] Arizona Commerce Authority. “Future48 Workforce Accelerators.” https://www.azcommerce.com/future48-workforce-accelerators/

[39] Arizona Commerce Authority. “SEMICON West.” July 30, 2026. https://www.azcommerce.com/news-events/news/2026/7/semicon-west/

[40] AZFamily. “TSMC Adds $100B to Arizona Plans as Leaders Call It a Chipmaking ‘Gold Rush.’” July 27, 2026. Quotes on Arizona’s $620 billion GDP and 12,000 construction-trade workers. https://www.azfamily.com/2026/07/27/tsmc-adds-100b-arizona-plans-leaders-call-it-chipmaking-gold-rush/

[41] The Motley Fool. “TSMC (TSM) Q3 2026 Earnings Call Transcript.” October 15, 2026. https://www.fool.com/earnings/call-transcripts/2026/10/15/tsmc-tsm-q3-2026-earnings-call-transcript/

[42] Arizona State Legislature. HB 4168 / SB 1861, Senate Fact Sheet as Enacted. 57th Legislature, Second Regular Session, 2026. https://www.azleg.gov/legtext/57leg/2R/summary/S.4168-1861ATT_ASENACTED.DOCX.htm

[43] Good Jobs First. “Arizona’s Data Center Tax Break Surged 98% in One Year.” 2026. Reports FY2024 cost of $19.4 million and FY2025 cost of $38.5 million per the Arizona Department of Revenue Tax Expenditure Report. https://goodjobsfirst.org/arizona-data-center-tax-break-costs-surge-98-percent/

[44] Arizona Commerce Authority. “Computer Data Center Program.” https://www.azcommerce.com/incentives/computer-data-center-program/

[45] BillTrack50. “AZ HB2756 — Utilities; High Load Factor Customers.” Signed June 4, 2026. https://www.billtrack50.com/billdetail/1938738

[46] Arizona Public Service. “Data Centers: How We’re Protecting Customers While Planning for Big Energy Needs.” Describes the 2017 high-load-factor rate and the June 2025 proposal for a 45% increase for extra-large users. https://www.aps.com/en/About/Our-Company/Newsroom/Articles/Data_Centers_How_Were_Protecting_Customers_While_Planning-for_Big_Energy_Needs

[47] Arizona Daily Star / Capitol Media Services. “Gov. Hobbs Wants to Build More Energy Sources, Cut State Use.” September 2025. Quotes Hobbs on the Arizona Energy Promise Task Force. https://tucson.com/news/state-regional/government-politics/article_5e1133a7-1935-4403-bd45-3604e6a5f46c.html

[48] pv magazine USA. “Arizona Task Force Roadmap Prioritizes Virtual Power Plants and Distributed Solar to Cut Energy Costs.” April 8, 2026. https://pv-magazine-usa.com/2026/04/08/arizona-task-force-roadmap-prioritizes-virtual-power-plants-and-distributed-solar-to-cut-energy-costs/

[49] Construction Reporter. “Arizona Senate Passes Bill Requiring More Detailed Information on Data Center Energy Usage.” 2026. Quotes the Arizona Energy Promise Taskforce report. https://constructionreporter.com/arizona-senate-passes-bill-requiring-more-detailed-information-on-data-center-energy-usage/

[50] Arizona Daily Star. “Arizona’s Pause on Data Center Tax Breaks Prompts Rush Before Deadline.” July 2026. https://tucson.com/news/state-regional/government-politics/article_99e67946-a818-48d6-9559-be754a7b2d9f.html

[51] Copper Courier. “Data Center Developers Rush for Arizona Tax Breaks Before Three-Year Moratorium.” July 13, 2026. https://coppercourier.com/news/politics/data-center-developers-rush-for-tax-relief/

[52] Arizona Department of Water Resources. “Phoenix Active Management Area Groundwater Supply Updates.” https://www.azwater.gov/phoenix-ama-groundwater-supply-updates

[53] Data Center Dynamics. “TSMC Breaks Ground on Water Reclamation Project in Phoenix, Arizona.” 2025. Reports the 11.4 million gallons per day development agreement and 85%/90% recycling targets. https://www.datacenterdynamics.com/en/news/tsmc-breaks-ground-on-water-reclamation-project-in-phoenix-arizona/

[54] Arizona Technology Council / In Business PHX. “TSMC Breaks Ground on ‘Near-Zero’ Discharge Water Plant to Back Fabs.” 2025. Quotes Rose Castanares, President, TSMC Arizona. https://www.aztechcouncil.org/tsmc-breaks-ground-on-near-zero-discharge-water-plant-to-back-fabs/

[55] United Nations University Institute for Water, Environment and Health. Environmental Cost of AI’s Energy Use: Carbon, Water and Land Footprints. June 2026, as reported by Business Standard, June 4, 2026. https://www.business-standard.com/world-news/ai-data-centres-electricity-water-use-un-report-2030-126060400318_1.html

[56] Salt River Project. “Data Centers and SRP.” https://www.srpnet.com/about/data-center-facts

[57] Salt River Project. “Integrated System Plan.” https://www.srpnet.com/grid-water-management/future-planning/integrated-system-plan

[58] NPR, Short Wave. “Why AI Data Center Costs May Be Driving Up Your Electric Bill.” Summarizing the Carnegie Mellon University–North Carolina State University analysis (8% nationally, 25% in Virginia by 2030). https://www.npr.org/transcripts/nx-s1-5626900

[59] North Carolina State University / EurekAlert. “Study Suggests Data Centers Will Boost Power Bills Up to 57% by 2030.” May 18, 2026. Lead author Jeremiah Johnson. https://www.eurekalert.org/news-releases/1128659

[60] CEPR VoxEU. “How Data Centres Affect Electricity Prices and for Whom.” Based on CEPR Discussion Paper 21854, “Large-Load Electricity Customers and Spillovers on Household Prices: Evidence from U.S. Data Centers.” https://cepr.org/voxeu/columns/how-data-centres-affect-electricity-prices-and-whom

[61] Rutgers Policy Lab. “Do Data Centers Raise Residential Electricity Bills?” June 22, 2026. https://policylab.rutgers.edu/?p=11153

[62] Insper. “Inteligência Artificial e o Paradoxo Energético da Sustentabilidade.” Report on the keynote “Transition to Expansion: AI and the New Era of Energy” by Shon Hiatt, University of Southern California, at the São Paulo School of Advanced Science in Systems Change and Sustainability. https://www.insper.edu.br/pt/conteudos/acontece-no-insper/inteligencia-artificial-e-o-paradoxo-energetico-da-sustentabilidade

[63] Compute Law Blog. “Arizona Grid, Corporation Commission Rules, and AI Data Centers.” 2026. Describes APS XHLF demand charges, HB 2756, and ACC proceedings. https://compute-law-blog.pages.dev/power/arizona-grid-corporation-commission-rules-ai-data-centers/

[64] ABC15 Arizona. “Gov. Katie Hobbs Says Arizona Should Rethink State Incentives for Data Centers.” https://www.abc15.com/news/local-news/gov-katie-hobbs-says-arizona-should-rethink-state-incentives-for-data-centers

[65] Commonwealth of Virginia. HB 30 — Enrolled Budget, Item 3-5.24: Data Center Electricity Consumption Tax. 2026 Special Session I. https://budget.lis.virginia.gov/item/2026/2/HB30/Enrolled/3/3-5.24/

[66] Williams Mullen. “Virginia Budget Creates New Electricity Consumption Tax for Data Centers.” July 1, 2026. https://www.williamsmullen.com/insights/news/legal-news/virginia-budget-creates-new-electricity-consumption-tax-data-centers

[67] Virginia Business. “Virginia Finalizes $200B+ 2026–28 Budget.” June 30, 2026. Describes the $600 million annual cap and refund mechanism. https://virginiabusiness.com/virginia-finalizes-2026-budget-data-center-tax-cannabis-sales/

[68] Virginia Business. “Virginia Budget with Data Center Tax Moves to Spanberger’s Desk.” June 22, 2026. Quotes Governor Spanberger. https://virginiabusiness.com/virginia-budget-with-data-center-tax-moves-to-spanbergers-desk/

[69] Broadband Breakfast. “Virginia Lawmakers Strike Data Center Tax Deal After Budget Standoff.” June 24, 2026. https://broadbandbreakfast.com/virginia-lawmakers-strike-data-center-tax-deal-after-budget-standoff/

[70] ARLnow. “Virginia Legislators Propose New Tax on Data Centers’ Power Usage to Break Budget Deadlock.” June 22, 2026. Quotes Brennan Gilmore, Clean Virginia. https://www.arlnow.com/2026/06/22/virginia-legislators-propose-new-tax-on-data-centers-power-usage-to-break-budget-deadlock/

[71] Whiteford. “Client Alert: Virginia Enacts Energy Consumption Tax on Data Centers and Signs Broader Energy Affordability Package.” July 30, 2026. https://www.whitefordlaw.com/news-events/client-alert-update-virginia-enacts-energy-consumption-tax-on-data-centers-and-signs-broader-energy-affordability-package

[72] Akin Gump Strauss Hauer & Feld. “Texas Pauses Data Center Interconnections Pending Statewide Audit.” August 11, 2026. Summarizes SB 6 (signed June 20, 2025), Batch Zero, and the audit directive. https://www.akingump.com/en/insights/alerts/texas-pauses-data-center-interconnections-pending-statewide-audit

[73] Office of the Texas Governor. “Governor Abbott Directs Comprehensive Data Center Audit.” August 3, 2026. https://gov.texas.gov/news/post/governor-abbott-directs-comprehensive-data-center-audit

[74] E&E News. “Texas Governor Talks Tough on Data Centers, Calls for Clampdown.” June 11, 2026. Quotes Abbott’s June 10 directive. https://www.eenews.net/articles/texas-governor-talks-tough-on-data-centers-calls-for-clampdown/

[75] S&P Global Commodity Insights. “Democrat Hinojosa Wins Texas Governor’s Race, Ends 3-Decade Republican Run.” November 4, 2026. https://www.spglobal.com/energy/en/news-research/latest-news/electric-power/110426-democrat-hinojosa-wins-texas-governors-race-ends-3-decade-republican-run

[76] The Texas Tribune via States Newsroom. “Gina Hinojosa Won the Governor’s Race. What Happens to Texas Data Centers Now?” November 4, 2026. https://www.newsfromthestates.com/node/436869

[77] Blackridge Research. “Texas Governor-Elect Gina Hinojosa Pledges Extended Data Center Moratorium and New Legislation.” November 5, 2026. https://www.blackridgeresearch.com/news-releases/texas-governor-elect-gina-hinojosa-pledges-extended-data-center-moratorium-and-new-legislation

[78] The Atlanta Journal-Constitution. “Georgia’s Sales Tax Breaks for Data Centers Tally More Than $2.5 Billion.” February 2026. Reports Georgia and Texas cost revisions compiled by Good Jobs First. https://www.ajc.com/business/2026/02/georgias-sales-tax-breaks-for-data-centers-tally-more-than-25-billion

[79] Reuters via The Star. “Texas Data Centers Face Uncertain Future After Hinojosa Wins Governor’s Race.” November 4, 2026. Quotes Brandon Rottinghaus, University of Houston. https://www.thestar.com.my/tech/tech-news/2026/11/05/texas-data-centers-face-uncertain-future-after-hinojosa-wins-governors-race

[80] Epoch AI. “OpenAI Stargate: Where the US Sites Stand.” 2026. https://epoch.ai/publications/openai-stargate-where-the-us-sites-stand

[81] Anthropic. “Anthropic Invests $50 Billion in American AI Infrastructure.” November 12, 2025. https://www.anthropic.com/news/anthropic-invests-50-billion-in-american-ai-infrastructure

[82] Tech-Insider. “TSMC’s $265B Arizona GigaFab: Reshaping US Chips.” 2026. Reports TSMC’s October 15, 2026 confirmation of Texas site evaluations. https://tech-insider.org/tsmc-arizona-165-billion-expansion-gigafab-2026/

[83] DLA Piper. “Pennsylvania Executive Order Establishes State Compliance Framework for Data Center Development.” August 27, 2026. https://www.dlapiper.com/insights/publications/2026/08/pennsylvania-executive-order-establishes-state-compliance-framework

[84] Altoona Mirror. “Shapiro Order Puts Curbs on Data Center Growth.” August 2026. https://www.altoonamirror.com/news/local-news/2026/08/shapiro-order-puts-curbs-on-data-center-growth/

[85] K&L Gates. “Pennsylvania Data Center Executive Order Adds New Development and Permitting Requirements.” August 19, 2026. https://www.klgates.com/thought-leadership/Pennsylvania-Data-Center-Executive-Order-Adds-New-Development-and-Permitting-Requirements-8-19-2026

[86] Ballard Spahr. “Pennsylvania’s Data Center Executive Order Ties Compliance With State’s GRID Requirements to Streamlined Development.” August 21, 2026. https://www.ballardspahr.com/insights/alerts-and-articles/2026/08/pa-data-center-executive-order-ties-compliance-with-the-grid-requirements-to-streamlined-development

[87] American Public Power Association. “Pennsylvania Governor Signs Executive Order on Data Center Development.” August 2026. https://www.publicpower.org/node/31007

[88] FOX 29 Philadelphia. “Gov. Shapiro Signs Executive Order Putting Guardrails on Pennsylvania Data Centers.” August 18, 2026. https://fox29.com/news/gov-shapiro-signs-executive-order-places-new-guardrails-data-center-development

[89] Environment+Energy Leader. “Pennsylvania Ties Data Center Permits to New GRID Rules.” August 2026. https://environmentenergyleader.com/stories/pennsylvania-ties-data-center-permits-to-new-grid-rules,137601

[90] Office of Governor Josh Shapiro. “Gov. Shapiro Secures Extension of PJM Price Cap, Saving Consumers $45B in Total.” February 13, 2026. https://pubt.io/view/156C38D13BE97F01C012150158E3A964193981BC

[91] Utility Dive. “PJM Approves First Expedited Interconnection Track Projects from Engie, LS Power.” October 2026. https://www.utilitydive.com/news/engie-ls-power-pjm-expedited-interconnection-eit/832613/

[92] American Public Power Association. “PJM Accepts First Three Projects in Expedited Interconnection Track.” October 2026. https://www.publicpower.org/node/31434

[93] Vinson & Elkins. “FERC Institutes Six Simultaneous Section 206 Proceedings Targeting Large Load Interconnection Across All RTO/ISO Markets.” June 2026. https://www.velaw.com/insights/ferc-institutes-six-simultaneous-section-206-proceedings-targeting-large-load-interconnection-across-all-rto-iso-markets/

[94] Akin Gump. “FERC Issues Landmark Show Cause Orders on Large Load Interconnection.” June 18, 2026. https://www.akingump.com/en/insights/blogs/speaking-energy/ferc-issues-landmark-show-cause-orders-on-large-load-interconnection

[95] Intel Corporation. “Ohio One Construction Timeline Update.” February 28, 2025. https://www.intel.com/content/www/us/en/newsroom/news/corporate/ohio-one-construction-timeline-update.html

[96] Magica. “Intel Ohio Fab Truss Milestone, Production Delay.” October 21, 2026. https://magica.com/news/intel-ohio-fab-truss-milestone-production-delay

[97] ServeTheHome. “Intel Moves Forward at Ohio One With First Major Structural Truss Installation.” October 2026. https://www.servethehome.com/intel-moves-forward-at-ohio-one-with-first-major-structural-truss-installation/

[98] WHIO. “Intel Spent $1.53B in Ohio in 2025; Updated Construction Timeline Outlined in New State Report.” 2026. Reports cumulative spending of $5.26 billion. https://www.whio.com/news/local/intel-spent-153b-ohio-2025-updated-construction-timeline-outlined-new-state-report/SRGWM5UWE5DF5IPDKODZP3F2YU

[99] The Motley Fool. “Intel (INTC) Q3 2026 Earnings Call Transcript.” October 30, 2026. https://www.fool.com/earnings/call-transcripts/2026/10/30/intel-intc-q3-2026-earnings-call-transcript/

[100] MultiState. “Ohio Governor Election Results 2026: Amy Acton Wins.” https://www.multistate.us/elections/governor-race/ohio-2026

[101] Indiana Economic Development Corporation. “Data Center Sales Tax Exemption — Overview.” https://www.iedc.in.gov/indiana-advantages/investments/data-center-sales-tax-exemption/overview

[102] Fortune. “Meta’s Hyperion Data Center Tax Break in Louisiana.” May 14, 2026. Reports Good Jobs First’s $8.2 billion estimate for Amazon’s New Carlisle, Indiana subsidies. https://fortune.com/2026/05/14/meta-data-center-tax-break-hyperion-louisiana/

[103] Site Selection Magazine. “AI & the Workforce.” Notes Google’s $2 billion Fort Wayne, Indiana data center cluster. https://siteselection.com/ai-the-workforce-how-ai-is-changing-the-future-of-work/

[104] Washington Examiner. “Trump Expands Ratepayer Protection Pledge to Prevent Data Center Costs Hitting Households.” July 23, 2026. Also reports New York’s one-year ban on AI data centers. https://www.washingtonexaminer.com/?p=4661057

[105] Micron Technology, Inc. Form 10-K for fiscal year ended September 3, 2026. Describes Idaho and New York fab schedules, up to $6.1 billion in CHIPS direct funding, and the New York term sheet of up to $5.5 billion. https://www.sec.gov/Archives/edgar/data/0000723125/000072312526000023/mu-20260903.htm

[106] Office of Governor Gavin Newsom. “Governor Newsom Signs Most Comprehensive Data Center Laws in the Nation, Providing Communities More Control on Water, Electricity, and Land Use.” September 21, 2026. https://www.gov.ca.gov/2026/09/21/governor-newsom-signs-most-comprehensive-data-center-laws-in-the-nation/

[107] Mayer Brown. “California’s Governor Signs Nation’s ‘Most Comprehensive’ Data Center Laws, Reshaping Regulation of Energy, Water, and Land Use.” September 2026. https://www.mayerbrown.com/en/insights/publications/2026/09/californias-governor-signs-nations-most-comprehensive-data-center-laws-reshaping-regulation-of-energy-water-and-land-use

[108] CalMatters. “New California Laws on Data Centers.” September 2026. https://calmatters.org/economy/technology/2026/09/new-california-laws-data-centers/

[109] Fox News. “2026 Governor Election Results by State.” https://www.foxnews.com/elections/2026/midterm-results/governor

[110] Georgia Department of Revenue. “High-Technology Data Center Equipment Exemption.” https://dor.georgia.gov/high-technology-data-center-equipment-exemption

[111] AutoInformed. “Data Center Tax Abatements Threaten State Budgets — and You.” 2026. Summarizes Good Jobs First’s analysis of Georgia (+664% to $2.5 billion) and Texas (>$1 billion annually) and quotes Kasia Tarczynska. https://autoinformed.com/data-center-tax-abatements-threaten-state-budgets-and-you/

[112] WRDW. “Data Centers, Power Bills Become Ga. Election Flashpoint.” September 14, 2026. https://www.wrdw.com/2026/09/14/data-centers-power-bills-become-ga-election-flashpoint/

[113] Stateline. “Most States Don’t Disclose Which Companies Get Data Center Incentives, Report Finds.” November 12, 2025. https://stateline.org/2025/11/12/most-states-dont-disclose-which-companies-get-data-center-incentives-report-finds/

[114] WABE / States Newsroom. “Kemp, Other GOP Govs Join Trump Pledge Designed to Shield Ratepayers from Data Center Costs.” July 23, 2026. https://www.wabe.org/?p=972696

[115] Gray DC. “President Trump Expands Data Center Utility Bill Pledge.” July 23–24, 2026. Quotes President Trump and David Klaus. https://www.wafb.com/2026/07/24/president-trump-expands-data-center-utility-bill-pledge/

[116] Taiwan Semiconductor Manufacturing Company. “TSMC Reports Third Quarter EPS of NT$29.43.” Form 6-K, October 15, 2026. https://www.sec.gov/Archives/edgar/data/0001046179/000104617926000726/a3q26e_withguidancexfinal.htm

[117] Taiwan Semiconductor Manufacturing Company. “2026 Q3 Quarterly Results.” Investor Relations. https://investor.tsmc.com/english/quarterly-results/2026/q3

[118] Global Brands Magazine. “TSMC’s Third Quarter: The Numbers Behind a $46.7bn Quarter and a Supply Gap That Is ‘Still Quite Big.’” October 2026. https://www.globalbrandsmagazine.com/tsmc-q3-2026-results-analysis

[119] NVIDIA Corporation. “NVIDIA Announces Financial Results for Third Quarter Fiscal 2026.” November 2025. https://investor.nvidia.com/news/press-release-details/2025/NVIDIA-Announces-Financial-Results-for-Third-Quarter-Fiscal-2026/default.aspx

[120] NVIDIA Corporation. “NVIDIA Announces Financial Results for Second Quarter Fiscal 2027.” Form 8-K, August 26, 2026. https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000073/q2fy27pr.htm

[121] NVIDIA Corporation. “NVIDIA Announces Financial Results for Second Quarter Fiscal 2027.” Press release with third-quarter outlook, August 26, 2026. https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2027

[122] CNBC. “Nvidia Earnings Takeaways: Huang Forecasts 70% Fiscal 2028 Revenue Growth, Far Above Estimates.” August 26, 2026. https://www.cnbc.com/2026/08/26/nvidia-nvda-earnings-report-q2-2027-live-updates.html

[123] MarketBeat. “NVIDIA (NVDA) Q2 2027 Earnings Results & Call Transcript.” August 26, 2026. Includes Huang’s remark on the cost per gigawatt rising from ~$30 billion to ~$60 billion. https://www.marketbeat.com/earnings/reports/2026-8-26-nvidia-co-stock/

[124] Advanced Micro Devices, Inc. “AMD Reports Third Quarter 2026 Financial Results.” Form 8-K, November 3, 2026. https://www.sec.gov/Archives/edgar/data/0000002488/000000248826000188/q32026991.htm

[125] Benzinga. “Advanced Micro Devices Q3 2026 Earnings Call Transcript.” November 2026. https://www.benzinga.com/news/26/11/62849372/advanced-micro-devices-q3-2026-earnings-call-transcript

[126] AMD. “AMD Announces Production Ramp of Next-Generation AMD EPYC Processor ‘Venice’ on TSMC 2nm Process Technology.” May 21, 2026. https://ir.amd.com/news-events/press-releases/detail/1287/amd-announces-production-ramp-of-next-generation-amd-epyc-processor-venice-on-tsmc-2nm-process-technology

[127] Intel Corporation. “Intel Reports Third-Quarter 2026 Financial Results.” Form 8-K, October 29, 2026. https://www.sec.gov/Archives/edgar/data/0000050863/000005086326000188/q326earningsrelease.htm

[128] CNBC. “Intel Stock Plunges 21% as Chipmaker Blames PC Slowdown for Weak Guidance.” October 29, 2026. https://www.cnbc.com/2026/10/29/intel-intc-q3-earnings-report-2026-.html

[129] IEEE Spectrum. “The CHIPS Act’s Impact on Semiconductor Manufacturing.” August 2026. Summarizes GAO’s August 6, 2026 report. https://spectrum.ieee.org/chips-act-impact-semiconductor-manufacturing

[130] Micron Technology, Inc. “Micron Technology, Inc. Reports Record Fiscal Fourth-Quarter and Full-Year 2026 Results.” Form 8-K Exhibit 99.1, September 30, 2026. https://www.sec.gov/Archives/edgar/data/0000723125/000072312526000018/a2026q4ex991-pressrelease.htm

[131] Micron Technology, Inc. “Fiscal Q4 2026 Earnings Call Prepared Remarks.” September 30, 2026. https://s25.q4cdn.com/621799436/files/doc_financials/2026/q4/Q4-FY26-Prepared-Remarks.pdf

[132] CNBC. “Micron (MU) Q4 Earnings Report 2026.” September 30, 2026. https://www.cnbc.com/2026/09/30/micron-mu-q4-earnings-report-2026.html

[133] Micron Technology, Inc. “Micron Accelerates U.S. Investments, Pours First Concrete at New York Fab.” July 9, 2026. https://investors.micron.com/news/press-release/2026/Micron-Accelerates-U-S–Investments-Pours-First-Concrete-at-New-York-Fab/default.aspx

[134] U.S. Government Accountability Office. Semiconductors: Commerce Needs Plan to Meet CHIPS for America R&D Requirements. GAO-26-109121, August 6, 2026. https://www.gao.gov/products/gao-26-109121

[135] Legis1. “CHIPS for America R&D: Commerce Cancels $7.8 Billion.” August 2026. https://legis1.com/news/chips-for-america-rd-commerce-cancels-78-billion

[136] Global Trade Alert. “United States of America: Department of Commerce Signs Letters of Intent for CHIPS R&D Awards to Four Companies.” October 22, 2026. https://globaltradealert.org/state-act/101524-united-states-of-america-department-of-commerce-announces-rd-awards-for-four-semiconductor-companies

[137] Amazon.com, Inc. “Amazon.com Announces Third Quarter Results.” October 29, 2026. https://ir.aboutamazon.com/news-release/news-release-details/2026/Amazon-com-Announces-Third-Quarter-Results/default.aspx

[138] Benzinga. “Amazon.com Q3 2026 Earnings Call Transcript.” October 2026. https://www.benzinga.com/news/26/10/62724825/amazoncom-q3-2026-earnings-call-transcript

[139] Data Studios. “Amazon Q3 2026 Results: AWS Grows 39% to $46B, $578B Backlog, About $220B 2026 Capex.” October 2026. https://www.datastudios.org/post/amazon-q3-2026-results-aws-39-percent-growth-46-billion-220-billion-capex

[140] TradingKey. “Amazon (AMZN) Q3 2026 Earnings Call: AWS Growth Hits 39.3%.” October 2026. https://www.tradingkey.com/news/transcripts/262256105-tradingkey

[141] CNBC. “Alphabet Sinks 10% as It Holds 2026 Capex Unchanged, Expects Significant Increase in 2027.” October 28, 2026. https://www.cnbc.com/2026/10/28/alphabet-googl-earnings-q3-2026.html

[142] Google. “Google Introduces New TPUs at Cloud Next ’26.” April 22, 2026. https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/tpus-8t-8i-cloud-next/

[143] TensorFeed. “AI Infrastructure Tracker.” 2026. Reports Anthropic’s five-year, $200 billion Google Cloud/TPU commitment (May 5, 2026) and Fluidstack-built sites. https://c72f82d6.tensorfeed.pages.dev/ai-infrastructure

[144] Stock Metric Lab. “AI CapEx Comparison: Microsoft, Alphabet, Amazon & Meta.” November 2026. https://stockmetriclab.com/ai-capex-comparison-2026/

[145] Tech-Insider. “Microsoft, Meta, Google AI Capex Hits $725B [2026].” October 2026. https://tech-insider.org/microsoft-meta-google-ai-capex-725-billion-2026/

[146] Microsoft Corporation. “Earnings Release FY27 Q1.” October 28, 2026. https://www.microsoft.com/en-us/investor/earnings/fy-2027-q1/press-release-webcast

[147] CNBC. “Microsoft Stock Jumps 4% as Cloud Growth Drives Earnings and Revenue Beat.” October 28, 2026. https://www.cnbc.com/2026/10/28/microsoft-msft-q1-earnings-report-2027.html

[148] Data Studios. “Microsoft Q1 FY27 Results: Azure Grows 48% to $33.2 Billion as Quarterly Capex Reaches $45.2 Billion.” October 2026. https://www.datastudios.org/post/microsoft-q1-fy27-results-azure-48-percent-33-2-billion-capex-45-2-billion

[149] Value Add VC. “Big Tech AI Spending 2026: ~$730B Capex, Company by Company.” Updated November 4, 2026. https://valueaddvc.com/ai-spending

[150] Enlit World. “AI and Data Centre Electricity Use Continue to Surge, IEA Finds.” 2026. https://www.enlit.world/library/ai-and-data-centre-electricity-use-continue-to-surge-iea-finds

[151] Bossblog Research Desk. “OpenAI Tells Investors 30 Gigawatts of Compute Planned by 2030 — Dwarfing Anthropic’s 7–8 Gigawatts.” April 12, 2026. https://bossblog-alpha.vercel.app/blog/2026-04-12-openai-compute

[152] Federal Reserve Bank of Richmond. “Will AI Investments Pay Off?” Econ Focus, First/Second Quarter 2026. Discusses Acemoglu’s 2024 estimate and quotes Erik Brynjolfsson. https://www.richmondfed.org/publications/research/econ_focus/2026/q1-q2_feature2

[153] Fortune. “‘It’s Not Going Away’: The Stanford Economist Who Called the AI Entry-Level Jobs Crisis Early Has the Receipts.” June 27, 2026. https://fortune.com/2026/06/27/what-is-ai-impact-entry-level-jobs-stanford-adp-canaries-brynjolfsson-richardson/

[154] Fortune. “One of Stanford’s Original AI Gurus Says Productivity Liftoff Has Begun After Doubling in 2025.” February 15, 2026. https://fortune.com/2026/02/15/ai-productivity-liftoff-doubling-2025-jobs-report-transition-harvest-phase-j-curve/

[155] International Monetary Fund. “A Global Vision for Indian AI.” Remarks by Managing Director Kristalina Georgieva at the India AI Impact Summit, New Delhi, February 19, 2026. https://pubt.io/view/447B47C3A158B3BDCA53E39AED626216AD5391A1

[156] Digital Watch Observatory. “AI Power Play, No Referees.” World Economic Forum session summary, January 20, 2026. https://dig.watch/?p=240229

[157] The Next Web. “IMF Chief Warns Advanced AI Models Like Mythos ‘Can Be Used to Destroy the Financial System.’” 2026. https://thenextweb.com/news/imf-georgieva-mythos-destroy-financial-system-ai-bubble

[158] “Boom, Bubble, or Buildout? A Multi-Method Evaluation of Whether Artificial Intelligence Is in an Ongoing Financial Bubble.” arXiv:2606.01575, June 2026. https://arxiv.org/pdf/2606.01575

[159] CNN. “2026 Midterm Election Results: Democrats Projected to Win Over 250 House Seats.” November 5, 2026. https://www.cnn.com/2026/11/05/politics/live-news/midterm-election-congress

[160] Louisiana Illuminator / States Newsroom. “Landry, GOP Govs Join Trump Pledge to Shield Utility Ratepayers from Data Center Costs.” July 23, 2026. Reports the Sanders/Ocasio-Cortez moratorium bills and New York’s ban. https://lailluminator.com/?p=34126

[161] Brownstein Hyatt Farber Schreck. “Data Center Friendliness Tracker — Virginia.” October 27, 2026 edition. Notes the Data Center Tax Accountability and Disclosure Act (July 2026). https://www.bhfs.com/data-center-tracker/virginia/

[162] U.S. Government Accountability Office. Semiconductors: Information on Projects Funded to Strengthen U.S. Supply Chain. GAO-26-107882, December 11, 2025. https://www.gao.gov/products/gao-26-107882

[163] Congressional Research Service. Semiconductor Fabrication Facilities Funded by the CHIPS Act: Project Status and Considerations for Congress. R49031, July 14, 2026. https://www.everycrsreport.com/files/2026-07-14_R49031_157331cbaf561ee49780ef5b60b2f38407b33cde.html

[164] BDO. “Section 48D Credit Opportunities for Semiconductor Manufacturing and Production Equipment Companies.” 2026. https://www.bdo.com/insights/industries/technology/section-48d-credit-opportunities-for-semiconductor-manufacturing-and-production-equipment-companies

[165] Instead. “Section 48D Credit Rises to 35% and Why You Must Act Before 2027.” March 31, 2026. https://instead.com/resources/blog/section-48d-credit-rises-to-35-and-why-you-must-act-before-2027

[166] PwC. “Final Regulations Are Issued on Section 48D Credit.” November 21, 2024. Describes the physical work test and 5% safe harbor for the December 31, 2026 termination date. https://www.pwc.com/us/en/services/tax/library/pwc-final-regulations-are-issued-on-section-48d-credit.html

[167] Inflect. “What the US AI Action Plan Means for Data Centers and Telcos.” July 31, 2025. Quotes the White House on AI dominance and the executive order fact sheet. https://inflect.com/blog/what-the-us-ai-action-plan-means-for-data-centers-and-telcos

[168] King & Spalding. “The White House’s Action Plan to Win the AI Race.” July 28, 2025. https://www.kslaw.com/insights/articles/the-white-houses-action-plan-to-win-the-ai-race

[169] Hunton Andrews Kurth. “President Trump Announces AI Action Plan and Accelerated Permitting and Financing for Datacenters.” July 2025. https://www.hunton.com/the-nickel-report/president-trump-announces-ai-action-plan-and-accelerated-permitting-and-financing-for-datacenters

[170] Special Competitive Studies Project. “Winning the AI Race: The White House Strategy for Innovation, Infrastructure, and International Diplomacy.” July 2025. https://scsp222.substack.com/p/winning-the-ai-race

[171] AllMind. “Stanford’s Big AI Report: The Most Important Takeaways.” April 2026. Summarizes the 2026 AI Index findings on the 2.7-point U.S.–China performance gap and $285.9 billion in U.S. private AI investment. https://allmind.ai/news/b3f63ebc2ccd704d769841a62cfe28395ef22cba76eeb9f24579d29c6e14f646

[172] Bureau of Industry and Security. “Revision to License Review Policy for Advanced Computing Commodities.” Federal Register, Vol. 91, No. 10, January 15, 2026. https://www.govinfo.gov/content/pkg/FR-2026-01-15/pdf/2026-00789.pdf

[173] Morgan Lewis. “BIS Revises Export Review Policy for Advanced AI Chips Destined for China and Macau.” January 16, 2026. https://www.morganlewis.com/pubs/2026/01/bis-revises-export-review-policy-for-advanced-ai-chips-destined-for-china-and-macau

[174] U.S. Department of Commerce, Bureau of Industry and Security. “Department of Commerce Revises License Review Policy for Semiconductors Exported to China.” January 13, 2026. https://bis.gov/press-release/department-commerce-revises-license-review-policy-semiconductors-exported-china

[175] International Business Times UK. “‘It’s Not Coming, Trust Me’: World Bank Chief Says Big AI Data Centres Will Stay in Rich Grids.” September 2026. Quotes Ajay Banga at Climate Week NYC, September 23, 2026, and cites the World Bank’s World Development Report 2026: The Promise of Artificial Intelligence. https://www.ibtimes.co.uk/emerging-markets-ai-data-centre-infrastructure-1821787

[176] Minnesota House of Representatives committee document summarizing Virginia JLARC 2024 findings, including $733 million in data center property taxes paid in 2024. https://www.house.mn.gov/comm/docs/quu3lDMDSUWjcccsskZ2xw.pdf

[177] Utility Dive. “Utilities May Subsidize Data Center Growth by Shifting Costs to Ratepayers: Harvard Paper.” March 2025. Quotes Martin and Peskoe. https://www.utilitydive.com/news/utilities-subsidize-data-center-growth-ratepayer-cost-shif/742001/

Featured Posts

Counterparty Cascade: When Artificial Intelligence Infrastructure Becomes a Network of Interdependent Financial Promises — How Hyperscaler Contracts, Semiconductor Suppliers, New Clouds, Private Credit, and Frontier Models Could Transmit Economic Risk Across the Five-Layer AI Economy Through 2030