Introduction: The Reuters Report That Was About More Than Texas
On September 30, 2026, Reuters correspondents Wen-Yee Lee and Fanny Potkin reported from Taipei that Taiwan Semiconductor Manufacturing Company, the world’s largest contract chipmaker, was evaluating a potential investment in Texas to expand its chip production in the United States. Two people familiar with the matter said the plans had not been finalized, and TSMC did not immediately respond to a request for comment. On its surface, the report read like another entry in the now-familiar genre of semiconductor expansion headlines, the sort of story that has appeared with such regularity since 2020 that readers have grown accustomed to numbers in the tens and hundreds of billions of dollars. Yet the location, and the single sentence that placed the possible Texas project in relation to Arizona, made this report far more consequential than the usual capital-spending anecdote. According to one of the sources, any Texas investment would come in addition to the $265 billion that TSMC had already committed to Arizona, where the company’s American footprint is evolving from a handful of wafer fabs into a dense manufacturing ecosystem that will eventually comprise twelve fabrication and advanced packaging facilities as well as a research and development center. [1,2]
Only a few years earlier, the central question surrounding TSMC in the United States had been existential rather than geographic. Could the world’s most important pure-play foundry, whose manufacturing culture had been forged in the Hsinchu Science Park and whose yields depended on an extraordinarily disciplined ecosystem of Taiwanese suppliers and engineers, manufacture advanced chips in America at meaningful scale and acceptable cost? The first Arizona fab was delayed, criticized, and treated by many observers as an experiment whose outcome remained uncertain. That question has not disappeared entirely, but it has been overtaken by a different one. TSMC’s first Phoenix fab entered high-volume production in the fourth quarter of 2024, its second fab is expected to enter high-volume manufacturing in the second half of 2027, and in July 2026 the company announced another $100 billion of planned Arizona investment to build four or more additional fabs for two-nanometer and more advanced technologies as well as advanced packaging plants. [3,4,5] Apple announced in February 2026 that it would purchase well over 100 million chips from the Arizona site during 2026, and TSMC paid $197 million at a state land auction for 900 additional acres adjacent to its north Phoenix campus to support further expansion. [10] If Arizona succeeds, and the evidence increasingly suggests that it is succeeding, the question becomes whether the United States simply makes Arizona larger, or whether semiconductor manufacturing begins spreading into a collection of specialized regional clusters.
A possible TSMC investment in Texas therefore matters even if it never proceeds in the form currently being considered. It introduces a structural question that extends far beyond one company’s capital-spending decision: what happens when America’s semiconductor strategy becomes geographically competitive inside America itself? For most of the period between the 2020 announcement of the first Arizona fab and the passage of the CHIPS and Science Act in 2022, the relevant geography was international. The United States was competing with Taiwan, South Korea, Japan, China and the European Union for a share of leading-edge manufacturing capacity, and the policy conversation in Washington was organized around national totals: how many fabs, how many wafers, what share of global advanced logic capacity could be brought onshore by 2030. The Reuters report hints at a second geography layered beneath the first, one in which states, utilities, universities, community colleges, water districts and municipal governments have become active participants in deciding where the physical infrastructure of artificial intelligence will be built.
Arizona has already demonstrated how rapidly this process can compound once the first anchor facility begins operating. When Governor Katie Hobbs and Phoenix Mayor Kate Gallego announced the July 2026 expansion, the City of Phoenix noted that since 2020 the state had attracted more than seventy semiconductor expansions representing over $314 billion in investment, the largest total in the nation. [3] Arizona economic-development officials pointed out that the state’s nominal gross domestic product is roughly $620 billion, which means that one company’s planned investment alone exceeds forty percent of the state’s annual output, and that TSMC’s commitment combined with Intel’s $32 billion Chandler expansion amounts to more than $300 billion relative to that same base. [8] A building-trades council estimated that roughly 12,000 construction workers could be required to build out the additional fabs. [8] These are not the numbers of a factory siting decision. They are the numbers of a regional economic transformation.
“TSMC’s historic investment firmly establishes Arizona as the nation’s epicenter for advanced semiconductor manufacturing and innovation.” — Governor Katie Hobbs [3]
But a modern semiconductor cluster cannot be measured simply by counting fabs, and this is the first analytical insight that the rest of this paper develops at length. Advanced chips require silicon wafers, ultrapure chemicals, specialty gases, precision lithography and deposition tools, metrology instruments, photomasks, advanced packaging, test capacity, water-treatment systems, exceptionally reliable electricity, specialized construction trades, process engineers, research universities, logistics infrastructure and thousands of suppliers capable of operating within tolerances that have no equivalent in any other manufacturing industry. Arizona’s own development illustrates this widening industrial perimeter. TSMC and Amkor Technology announced a ten-year advanced packaging and test procurement agreement in June 2026, tying TSMC’s north Phoenix fabs to Amkor’s $7 billion packaging and test campus in Peoria, and a month later Nvidia committed a $1.5 billion prepayment to expand Amkor’s Arizona capacity. [11,13,14] Semiconductor materials and equipment suppliers have been opening or enlarging facilities around the Phoenix region at a pace that local economic development officials have begun describing, with some justification, as a gold rush. [8]
Texas presents a different starting position, and much of the analytical interest of the Reuters report lies in that difference. Texas already possesses substantial semiconductor manufacturing, design, research and supplier activity, built over decades around Texas Instruments in Dallas and Sherman, Samsung in Austin and now Taylor, a dense chip-design community in the Austin corridor, and a growing photonics and optical-networking base in the Houston and Dallas regions. The state’s Texas Semiconductor Innovation Fund was created under the Texas CHIPS Act of 2023 and was replenished by the legislature in 2025, and during 2026 the Texas CHIPS Office has been distributing grants across a remarkably wide range of activities: $11.6 million to Soulbrain for a new semiconductor materials plant in Taylor, $14.1 million to Coherent for what the company describes as the world’s first six-inch indium phosphide wafer fab in Sherman, $20.9 million to Applied Optoelectronics for AI datacenter transceiver manufacturing in Sugar Land, and $3.9 million to Schunk Xycarb Technology for silicon-carbide-coated graphite and quartz components in Georgetown. [31,32,33,34] Samsung began trial production of two-nanometer wafers for Tesla at Taylor in September 2026, and Texas Instruments started production at SM1, the first of four planned 300-millimeter fabs at its Sherman megasite, in December 2025. [36,39]
Nor is this becoming merely an Arizona versus Texas contest. Micron is building an enormous memory-manufacturing geography that stretches across Idaho and New York, with more than $250 billion of planned U.S. fab and technology investment through 2035. Its first new Idaho fab is projected to produce initial DRAM wafers in mid-calendar 2027, its second Idaho fab is expected to be operational by the end of 2028, and its Clay, New York complex, where the first concrete was poured in July 2026 ahead of schedule, could eventually contain as many as four fabs supported by up to $6.1 billion of federal direct funding and as much as $5.5 billion in New York State GREEN CHIPS incentives over twenty years. [40,41,42] Micron has also committed up to $3 billion to the domestic supply chain, including $500 million of strategic financing to GlobalWafers’ 300-millimeter raw silicon wafer facility in Sherman, Texas, alongside a ten-year wafer supply agreement. [43] A memory company headquartered in Boise is therefore financing a Taiwanese wafer supplier’s expansion in north Texas in order to supply fabs in Idaho and upstate New York. The supply chain is already interstate.
The result is an emerging American semiconductor map in which states do not necessarily need to replicate one another. Arizona could concentrate leading-edge logic fabrication and advanced packaging. Texas could combine logic, design, equipment, materials, photonics, analog and foundational semiconductors, and eventually additional leading-edge foundry production. Idaho could deepen the connection between memory research and memory manufacturing. New York could become one of the world’s largest long-duration memory-manufacturing centers. Other states could specialize in equipment, substrates, power semiconductors, defense microelectronics, chemicals, packaging, research, or the AI datacenters that ultimately consume the resulting silicon.
That possibility changes the meaning of semiconductor industrial policy. The first phase of the American semiconductor revival was largely national. Washington asked how much leading-edge manufacturing the United States could bring back through the CHIPS and Science Act, the Section 48D investment tax credit, trade measures, research funding and federal partnerships. The geographic unit was the country. The next phase may be increasingly federal in the constitutional sense of that word. Arizona, Texas, New York, Idaho and other states are not simply waiting for Washington to allocate semiconductor capacity among them. They are constructing their own incentive systems, workforce programs, university partnerships, infrastructure packages, permitting structures and supplier-recruitment strategies, and companies can therefore compare not only countries but states, and increasingly complete regional industrial ecosystems.
This paper calls that emerging system Foundry Federalism. Foundry Federalism describes an American semiconductor economy in which federal policy establishes national strategic objectives while states compete, cooperate and specialize in assembling the physical conditions required to manufacture advanced computing systems. The competition is not limited to tax subsidies. It increasingly encompasses electricity, water, land, housing, transportation, universities, skilled labor, industrial suppliers, permitting speed, advanced packaging, research capabilities and proximity to the enormous AI datacenters that will consume the resulting silicon.
Seen through the Five-Layer AI Economy framework that organizes this series of papers, this geographic transformation begins at Layer 1 with electricity and industrial infrastructure; moves through Layer 2 with GPUs, CPUs, memory, networking silicon and foundry capacity; connects to Layer 3 through datacenters and hyperscaler demand; determines what becomes economically possible for Layer 4 models; and ultimately shapes the price, availability and geography of Layer 5 applications and autonomous agents. The scale of the demand pulling on this stack is no longer speculative. Nvidia reported revenue of $96.2 billion for its second quarter of fiscal 2027, up 106 percent from a year earlier, with datacenter revenue of $89.0 billion; TSMC’s August 2026 revenue rose 53.3 percent year over year; and Micron reported record fiscal third-quarter revenue of $41.46 billion against $9.30 billion in the same quarter a year earlier. [59,9,41] The semiconductor map, in other words, is becoming part of the AI map.
“Now, compute is revenue.” — Jensen Huang, Nvidia [59]
By 2030, America’s semiconductor strength may therefore be measured not by whether the country possesses one successful advanced-fab cluster, but by whether it has created a distributed federation of complementary semiconductor regions capable of supporting one another while reducing critical geographic dependencies. The Reuters report about Texas is an anecdote about something much larger than Texas. It asks whether Arizona is the destination of America’s semiconductor reshoring, or merely its beginning.
Why I Chose the Title “Foundry Federalism”
I chose Foundry Federalism because foundry represents the physical manufacturing foundation of the semiconductor economy, while federalism captures the distinctive American system in which national industrial policy interacts with fifty states possessing their own economic-development strategies, universities, utilities, infrastructure, incentives and regulatory institutions. The word foundry is deliberately chosen over the broader word semiconductor because it insists on the physical. Artificial intelligence may appear increasingly digital and autonomous at the upper layers of the economy, but every model ultimately rests on processors fabricated with extraordinarily demanding industrial processes in buildings that consume millions of gallons of water per day and hundreds of megawatts of electricity, staffed by people who must be trained somewhere, supplied by chemicals that must be manufactured somewhere, and packaged by facilities that, until very recently, existed almost exclusively in Asia. The word federalism is chosen because it carries two meanings simultaneously, and the paper depends on both. In its competitive sense, federalism describes states bidding against one another for mobile capital, a phenomenon that economists have studied, and often criticized, for decades. In its cooperative sense, federalism describes a system in which different levels and different units of government perform different functions within a shared national framework, and in which the whole can achieve what no single unit could achieve alone.
The title therefore shifts the analysis from one fab or one federal program toward the geographic organization of an entire industry. It also captures the paper’s central argument: America’s next semiconductor advantage may come not from forcing every region to become another Arizona, but from allowing states to develop complementary industrial specializations. Fabrication may concentrate in one region, advanced packaging in another, memory in another, semiconductor materials somewhere else, and hyperscale AI consumption in still another. Foundry Federalism is therefore both competition among states and coordination across them, and the central research problem the paper returns to repeatedly is how to tell the difference between the wasteful duplication that competitive federalism can produce and the strategic redundancy that a well-designed federation can provide.
Section 1: Arizona and the Creation of America’s First Advanced-Fab Anchor
Every durable industrial geography begins with an anchor, and the anchor of America’s semiconductor revival is unambiguously the TSMC campus in north Phoenix. To understand why the Texas question has become askable at all, it is necessary first to understand how Arizona became the place where the question could be posed. This section traces the compounding sequence through which a $12 billion foreign direct investment announced in May 2020 became a $265 billion commitment by July 2026, and it argues that the significance of this sequence lies less in the headline totals than in what each stage revealed about execution risk, yield, customer commitment and the gravitational pull that a working fab exerts on everything around it. Arizona matters to this paper not simply because it is large, but because it is the empirical demonstration that the ecosystem logic described in the Introduction actually operates on American soil.
1.1 From a $12 Billion Experiment to a $265 Billion Industrial Cluster
The original TSMC Arizona commitment, announced in the spring of 2020 amid pandemic disruption and intensifying U.S.-China technology tensions, called for roughly $12 billion of spending over eight years beginning in 2021 to build a single fab producing five-nanometer chips. [6] At the time, this was treated as a diplomatic gesture as much as an industrial strategy, a way for Taiwan’s most important company to demonstrate commitment to its most important security partner. The skeptics were numerous and their skepticism was not unreasonable. TSMC had never operated a leading-edge fab outside Taiwan at scale, American construction and operating costs were widely understood to be substantially higher, and the surrounding ecosystem of Taiwanese chemical, gas, equipment-service and packaging suppliers that made Hsinchu function did not exist in the Sonoran Desert.
In December 2022, TSMC announced that in addition to the first fab, which was scheduled to begin production of N4 process technology in 2024, it had started construction of a second fab scheduled to begin three-nanometer production in 2026, bringing overall investment for the two fabs to approximately $40 billion, which the company described as the largest foreign direct investment in Arizona history and one of the largest in the history of the United States. The same announcement projected that the two fabs would eventually manufacture more than 600,000 wafers per year and that the site would include an on-site industrial water reclamation plant designed to achieve near-zero liquid discharge. [6] The Commerce Department subsequently finalized up to $6.6 billion in direct CHIPS Act funding and up to $5 billion in loans in November 2024 to support what had by then grown into a three-fab plan with roughly $65 billion of committed investment. [45]
The decisive turn came in March 2025, when TSMC announced a further $100 billion of U.S. investment, raising the planned total to $165 billion and the planned number of Arizona fabs from three to six, with two advanced packaging facilities and an R&D center added to the campus plan. Chief Executive C.C. Wei described the completed configuration as a gigafab cluster, and indicated that it would eventually account for roughly thirty percent of TSMC’s two-nanometer and more advanced capacity. Then, on July 16, 2026, alongside second-quarter results that showed net profit up 77 percent year over year to NT$706.56 billion and quarterly sales of $39.45 billion, Wei announced another $100 billion, bringing the Arizona total to $265 billion and the planned number of facilities to twelve, including four or more additional fabs for two-nanometer and below technologies and additional advanced packaging plants. [4,5,7] Chief Financial Officer Wendell Huang simultaneously raised the company’s 2026 capital expenditure budget to between $60 billion and $64 billion from an earlier range of $52 billion to $56 billion. [4]
“The demand and the supply, the gap is so big.” — C.C. Wei, Chairman and CEO, TSMC [4]
Table 1. The compounding of TSMC’s Arizona commitment, 2020–2026
| Date | Cumulative plan | Facilities | What changed |
| May 2020 | ~$12 billion | 1 fab (N5) | Initial commitment; treated as a strategic gesture under U.S.-Taiwan pressure [6] |
| Dec 2022 | ~$40 billion | 2 fabs (N4, N3) | Second fab under construction; 600,000 wafers/year target; water reclamation planned [6] |
| Nov 2024 | ~$65 billion | 3 fabs | CHIPS award finalized: up to $6.6B direct funding, $5B loans [45] |
| Mar 2025 | $165 billion | 6 fabs + 2 packaging + R&D | Gigafab cluster concept; ~30% of 2nm-and-below capacity to be in Arizona |
| Jul 2026 | $265 billion | 12 fab/packaging facilities + R&D | Four or more additional 2nm-and-below fabs; 2026 capex raised to $60–64B [4,5] |
| Sep 2026 | (Texas under evaluation) | — | Reuters: possible second U.S. state, in addition to Arizona; not finalized [1] |
Why did the first fab matter so disproportionately, when its dollar value was a small fraction of what followed? The answer is that the first fab converted a question of belief into a question of measurement. Before high-volume production began in the fourth quarter of 2024, every argument about whether TSMC could manufacture in America was an argument about culture, cost structure, labor practices and national industrial character. After the first fab reached volume production with yields that TSMC publicly described as comparable to its Taiwanese fabs, the arguments shifted to the ordinary language of manufacturing: cycle time, tool utilization, defect density, customer qualification. Chris Miller of the Fletcher School at Tufts University, whose book Chip War has become the standard account of the industry’s geopolitics, argued before the fab opened that the conventional objections were unlikely to prove decisive, precisely because cost had never been the source of TSMC’s international dominance. [19]
“TSMC’s success internationally has not been due to cost efficiency.” — Chris Miller, Fletcher School, Tufts University [19]
This is why manufacturing yield matters more than ribbon-cutting announcements, and why the chronology in Table 1 should be read as a chain of de-risking events rather than as a sequence of press releases. A fab that announces itself but cannot reach competitive yields is a liability that consumes capital and reputation. A fab that reaches competitive yields is an asset that can be financed, expanded and copied, and its existence changes the risk calculation for every subsequent decision by the same company, by its customers and by its suppliers. Apple’s February 2026 statement that it would buy well over 100 million Arizona-made chips in 2026, a significant increase from 2025, was not a political gesture; it was a procurement decision by a company whose products cannot tolerate supply failures, and it signaled to every supplier contemplating a Phoenix location that the anchor customer’s demand was real. [10]
1.2 The Difference Between a Fab and a Cluster
The paper’s first major conceptual distinction can be stated compactly: a fab manufactures wafers, while a semiconductor ecosystem makes fabs reproducible. The distinction matters because the policy debate has often conflated the two, treating the announcement of a fab as if it were the achievement of a cluster, when in fact the fab is the beginning of a long process of industrial accretion whose outcome is not guaranteed. A leading-edge fab is an extraordinarily sensitive organism. It requires semiconductor-grade chemicals delivered in purities measured in parts per trillion, ultrapure water produced on site from municipal or reclaimed supplies, industrial gases such as nitrogen, argon and specialty etch gases supplied through dedicated pipelines, 300-millimeter silicon wafers whose surface flatness is specified at the atomic scale, continuous tool maintenance by technicians trained on each vendor’s equipment, cleanroom components, metrology and inspection systems, photomasks, advanced packaging and test capacity, logistics capable of moving wafers and tools without contamination, specialty construction trades, process engineers with years of experience, university research partnerships and electricity whose reliability and power quality exceed anything an ordinary industrial customer would demand.
John VerWey, East Asia national security advisor at the Pacific Northwest National Laboratory, has described the ecosystem effect in terms that the Arizona experience has borne out: a leading-edge fab brings with it the network of chemical, material, mask and equipment suppliers needed to keep the factory running. [21] The economic moat, therefore, is not merely TSMC’s factory. It is the accumulation of capabilities surrounding the factory, and the question for any second location is how much of that accumulation must be rebuilt from scratch and how much can be shared across a national network.
“You’re getting the ecosystem of chemical suppliers, material suppliers, mask suppliers, equipment suppliers.” — John VerWey, Pacific Northwest National Laboratory [21]
The density of the Asian clusters is the benchmark against which any American cluster will be measured, and Harvard Business School’s Willy Shih, whose research on manufacturing and supply chains spans four decades of semiconductor cycles, has been blunt about the gap. In an April 2026 Harvard Business Review analysis based on interviews with managers at eight major companies across the supply chain, Shih and PJ Lin found that even chips made at TSMC’s Fab 21 in Phoenix were still being shipped back to Asia for packaging, that multiple firms in the Phoenix area were competing for the same small pool of workers, and that one company found hiring packaging engineers in Arizona so difficult that it flew in expatriates from South Korea. [16]
“The Asian semiconductor cluster benefits from a high level of density.” — Willy C. Shih, Harvard Business School [16]
1.3 Advanced Packaging Becomes Part of the Geography
Advanced packaging deserves substantial treatment because it has become the hinge on which the entire AI hardware economy turns, and because its geographic distribution exposes most clearly the gap between fabricating a chip and producing a usable AI processor. For most of the industry’s history, packaging was the unglamorous back end, a lower-margin, more labor-intensive activity that could be located wherever labor was cheapest. The AI accelerator changed that calculus permanently. A modern training or inference accelerator is not a single die; it is a system of compute dies, stacks of high-bandwidth memory and dense interconnects integrated on silicon interposers or through three-dimensional stacking techniques such as TSMC’s CoWoS family. The packaging step now determines a substantial share of the device’s performance, power efficiency and yield, and packaging capacity has repeatedly been the binding constraint on how many accelerators Nvidia and its peers can ship in a given quarter. [12,14]
This is why the June 2026 agreement between TSMC and Amkor is a more significant event for Foundry Federalism than its modest press coverage suggested. Under the ten-year agreement, TSMC will procure advanced packaging and testing services from Amkor at facilities in Arizona, where Amkor is building what it has described as a $7 billion advanced packaging and test campus in Peoria that will be, in the words of state officials, the first place in the country where millions of leading-edge microchips are packaged and tested. [11,13] Amkor’s chief executive, Kevin Engel, framed the partnership as an effort to give customers a fully U.S.-based supply chain from advanced wafer manufacturing through packaging and test. [12] A month later, on July 23, 2026, Nvidia announced a $1.5 billion multi-year agreement with Amkor under which it would prepay to support the expansion of Amkor’s U.S. advanced packaging capacity, including capacity in Arizona, for next-generation AI and accelerated computing platforms combining different types of chips in a single package. [14,15] Amkor’s own second-quarter 2026 results, with record revenue of $1.90 billion and gross margin that improved to 16.8 percent from 12.0 percent a year earlier, suggest that the economics of domestic packaging are beginning to move in the direction its advocates hoped. [63]
“Underscores the central role advanced packaging plays in enabling the future of AI.” — Kevin Engel, CEO, Amkor Technology [14]
The policy question that this subsection poses is therefore not rhetorical: what does domestic fabrication accomplish if advanced chips must still leave the country before they become usable AI processors? Until TSMC’s own first Arizona advanced packaging facility, planned for around 2029 with CoWoS and 3D-IC capacity, and Amkor’s Peoria lines are operating at scale, the honest answer is that domestic fabrication accomplishes a great deal less than the investment totals imply. [12] Federal policy has recognized this explicitly; CHIPS for America established an advanced packaging research facility in Arizona, and the Commerce Department’s manufacturing incentive guidance has repeatedly treated packaging as a strategic component of domestic capability. But the geography of packaging is also where Arizona’s concentration advantage is most visible, because packaging facilities benefit enormously from proximity to the fabs whose wafers they process. Shih and Lin noted that the capacity of TSMC’s two planned Arizona packaging facilities had been mostly booked by Nvidia before they were built. [16]
1.4 The Supplier Multiplier
The second- and third-order investment that follows an anchor fab is what economists would call an agglomeration externality and what Arizona’s development officials call, more vividly, a gold rush. The mechanism is straightforward to describe and difficult to engineer. A fab’s demand for chemicals is large, continuous and intolerant of supply interruptions, so chemical-delivery system vendors and specialty-chemical producers locate blending, purification and distribution facilities within trucking distance. Equipment vendors such as ASML, Applied Materials, Lam Research, Tokyo Electron and KLA station field-service engineers and spare-parts inventories near the fabs they support, and as the number of tools grows, those field operations become local service centers employing hundreds of people. Materials suppliers producing photoresists, slurries, targets and cleanroom consumables follow the same logic. Engineering and construction specialists who have learned to build to semiconductor tolerances stay in the region because the next fab will need them. Over time, these suppliers begin serving one another, and a secondary economy of calibration, logistics, waste treatment and specialized staffing forms around the primary one.
Enrico Moretti of the University of California, Berkeley, whose work on the economics of place has shaped how policymakers think about clusters, has shown in the American Economic Review that geographical agglomeration produces significant productivity gains for inventors and that the effect strengthens with cluster size. [24] The implication for semiconductor policy is double-edged. Concentration is productive, which argues for making Arizona larger. But the same research implies that spreading activity across smaller clusters sacrifices some of that productivity, which is precisely the trade-off that any second TSMC state would have to overcome.
“Agglomeration economies make workers inside large clusters more productive.” — Enrico Moretti, University of California, Berkeley [24]
Arizona’s expanding supplier base demonstrates that semiconductor industrialization occurs through industrial multiplication rather than through one-time megaprojects, and the figure of more than seventy semiconductor-related expansions since 2020 captures that multiplication better than any single announcement. [3] It also establishes the yardstick for Texas: a second cluster will not be judged by whether it attracts a fab, but by whether the fab attracts the seventy suppliers that follow.
1.5 Arizona’s Resource Equation
Concentration has physical limits, and the honest evaluation of Arizona’s future requires confronting them directly rather than treating them as rhetorical objections to be dismissed. The most frequently discussed limit is water. TSMC’s first Arizona fab consumes approximately 4.75 million gallons of water per day; the second fab is expected to consume roughly 5.7 million gallons; and when the first three fabs are fully operational, combined demand is projected at 17.2 million gallons per day, against a development agreement with the City of Phoenix for access to 11.4 million gallons per day. [27,28] The company currently recycles about 65 percent of its intake and in August 2026 broke ground on a fifteen-acre Industrial Reclamation Water Plant intended to raise reuse to 90 percent or better, which TSMC says would reduce the first fab’s net demand to fewer than 1.2 million gallons per day. [27,28] Sarah Porter, director of the Kyl Center for Water Policy at Arizona State University, has observed that the planned ten-fab configuration could make TSMC the City of Phoenix’s largest tap-water customer, placing semiconductor manufacturing near the top of regional industrial water users, above golf courses, beverage plants and datacenters but below power generation. [26] Porter’s own framing of the trade-off is instructive because it refuses to treat water consumption in isolation from economic return.
“A million gallons of water can provide 200 high-paying semiconductor jobs.” — Sarah Porter, Kyl Center for Water Policy, Arizona State University [29]
Water is only the most visible constraint. Electricity for a twelve-facility campus will require generation and transmission investments that Arizona Public Service and the Salt River Project are already planning around, and the same utilities are simultaneously absorbing hyperscale datacenter load. Land is finite even in the desert, which is why TSMC’s purchase of 900 additional acres was newsworthy. [10] Housing costs in the north Phoenix corridor have risen as thousands of construction and engineering workers have arrived. The construction labor force itself is a bottleneck: 12,000 trades workers for the additional fabs must be found in a regional market that is also building datacenters, Intel’s Chandler expansion and Amkor’s Peoria campus. [8] And the engineering talent pool, as Shih and Lin documented, is already being fought over by multiple foundries and equipment vendors. [16] This creates the transition into the question that drives the rest of the paper: at what point does adding another fab to an existing cluster become less attractive than creating a second cluster somewhere else?
1.6 Arizona as Template, or Arizona as Exception?
Section 1 closes with the central ambiguity that the Texas evaluation has brought into focus. Arizona may be the template that other states reproduce, in which case the correct reading of the Reuters report is that TSMC, having learned how to build an American gigafab once, is preparing to apply the same playbook in a second state with a different resource profile. Or Arizona may be a unique semiconductor supercluster, the product of a specific convergence of federal money, Taiwanese strategic anxiety, an existing Intel ecosystem in Chandler, Arizona State University’s engineering scale, a cooperative city government with a hundred-year water plan, and a first-mover advantage that cannot be replicated, in which case its very success encourages companies to diversify elsewhere not to copy it but to hedge against its concentration. TSMC Arizona’s president, Rose Castanares, has pointed to the state’s long-horizon water planning as one of the reasons the company chose Phoenix, and that kind of institutional preparation is not something a second state can improvise in a single legislative session. [28]
“Including a water supply plan for 100 years.” — Rose Castanares, President, TSMC Arizona [28]
That unresolved question leads directly to Texas, but it cannot be answered by looking at Texas alone. It requires first understanding, in Section 2, what a semiconductor ecosystem actually consists of once the subsidy headline is set aside.
Section 2: A Semiconductor Ecosystem Is More Than Subsidies
The public conversation about semiconductor location decisions has been dominated by subsidy arithmetic, and the arithmetic is genuinely large: $52.7 billion appropriated under the CHIPS and Science Act, of which $39 billion was reserved for manufacturing incentives; a Section 48D investment tax credit raised from 25 percent to 35 percent by the One Big Beautiful Bill Act for property placed in service after 2025; state packages such as New York’s $5.5 billion GREEN CHIPS term sheet with Micron and Texas’s roughly $948 million in cumulative appropriations to its Semiconductor Innovation Fund. [45,49,40] Yet anyone who has spent time with the managers who actually make these decisions knows that the subsidy is rarely the decisive variable once a short list of credible locations has been assembled. This section argues that the relevant inputs to a semiconductor region are far broader than land plus subsidy plus factory, that each of the broader inputs is itself a form of infrastructure that must be deliberately built, and that the state which can assemble these inputs reliably may possess an advantage even when another state offers a larger nominal incentive package. The deep explanation that follows is intended to establish the vocabulary that Sections 3 through 5 use to compare states.
2.1 The Hidden Bill of Materials Behind a Semiconductor Region
Engineers describe a product by its bill of materials, the full list of components and sub-assemblies required to build it. A semiconductor region has a hidden bill of materials of its own, and the purpose of this subsection is to make it visible. The naive version of the location equation is land plus subsidy plus factory. The realistic version is power plus water plus talent plus chemicals plus equipment plus packaging plus logistics plus research plus permitting plus suppliers plus customers, with each term multiplied by a reliability coefficient that reflects how confident a company can be that the input will remain available over the thirty-year life of a fab. A state that scores well on subsidy but poorly on the reliability of electricity, or that has an excellent university but no construction workforce capable of building cleanrooms, will struggle to convert an announcement into a functioning cluster. Table 2 presents the hidden bill of materials in the form that corporate site-selection teams increasingly use, and the remainder of Section 2 takes each major term in turn.
Table 2. The hidden bill of materials of a semiconductor region
| Input | What a leading-edge fab actually requires | Why states, not Washington, largely decide it |
| Electricity | Hundreds of MW per fab; near-perfect reliability and power quality; redundant feeds; multi-decade price visibility | State PUCs, utilities and ISOs plan generation, transmission and large-load interconnection rules |
| Water | Millions of gallons/day of ultrapure feed; reclamation to 65–90%+; industrial pretreatment permits | Municipal supply agreements, state water law, drought management |
| Talent | Process, equipment and chemical engineers; technicians; operators; construction trades | State universities, community colleges, apprenticeship funding, K-12 pipeline |
| Chemicals and gases | Ultra-high-purity acids, solvents, photoresists, specialty gases; on-site or near-site supply | State grants to materials suppliers; industrial zoning; environmental permitting |
| Equipment and service | Tool vendors’ field engineers, spare parts, calibration labs within hours of the fab | Supplier recruitment, airport and logistics capacity |
| Advanced packaging and test | CoWoS-class 2.5D/3D integration, HBM stacking, final test near the fab | State incentives and sites for OSAT campuses (e.g., Peoria, Arizona) |
| Construction capacity | Trades able to build vibration-isolated cleanrooms, process piping, UPW networks at scale | Trade-union pipelines, state building codes, concurrent project load |
| Research institutions | Universities and national labs that co-develop processes and supply graduate talent | State higher-education funding and industry partnership structures |
| Permitting | Predictable environmental, air, water and zoning approvals on compressed timelines | State and municipal permitting reform |
| Customers and datacenters | Hyperscalers, automakers, defense primes, AI labs within the regional economy | State datacenter policy, sales-tax exemptions, grid access for large loads |
2.2 Electricity as Semiconductor Infrastructure
It has become commonplace to say that electricity is the binding constraint on artificial intelligence, and the International Energy Agency’s projections give that commonplace quantitative weight: global datacenter electricity consumption is expected to more than double from roughly 415 terawatt-hours in 2024 to around 945 terawatt-hours by 2030, with the United States accounting for by far the largest share of the increase and datacenters representing nearly half of American electricity demand growth over the period. [58] But the semiconductor fab’s relationship to electricity is different from the datacenter’s in ways that matter for geography, and the distinction deserves careful statement because it determines which regions can host which activities.
A datacenter needs a great deal of electricity, but a well-designed datacenter can tolerate brief interruptions through on-site batteries and generators, can shift some workloads in time, and can, under new regulatory frameworks such as Texas Senate Bill 6, agree to curtail during grid emergencies in exchange for faster interconnection. [56] A fab cannot do any of these things. Fabrication runs continuously, with individual wafers requiring hundreds of process steps over periods of weeks to months. A voltage sag lasting a fraction of a second can scrap wafers in process, misalign tools and force recalibration that takes days. A sustained outage can destroy work-in-progress worth tens or hundreds of millions of dollars and can damage equipment whose replacement lead times are measured in quarters. A fab therefore demands exceptional reliability, exceptionally high power quality, redundant transmission feeds, planning horizons of decades rather than years, and price predictability sufficient to underwrite a thirty-year asset. Shutdown risk is economically extreme in a way that has no close analogue elsewhere in heavy industry.
This distinction connects directly to Layer 1 of the Five-Layer AI Economy, and it corrects a common simplification. Electricity does not merely power AI datacenters; it manufactures the silicon that makes those datacenters possible, and it does so under stricter conditions than the datacenters themselves require. A region that can satisfy fab-grade power requirements can almost certainly satisfy datacenter requirements, but the reverse is not true. Texas illustrates the resulting tension vividly. ERCOT, the Texas grid operator, was tracking more than 438 gigawatts of large-load interconnection requests by mid-2026, nearly ninety percent of them from datacenters, against an all-time peak demand of roughly 85.5 gigawatts, and Governor Greg Abbott in August 2026 called for an audit of every datacenter in the queue, noting that requests had reached approximately 474 gigawatts. [56,55] ERCOT responded by delaying its first batch-study process for new large loads. [55] A foundry evaluating Texas is therefore evaluating a state with abundant generation, a famously independent grid and a datacenter-driven interconnection queue that has become one of the most congested in the world.
“More than five times Texas’ record peak electricity demand for ERCOT.” — Governor Greg Abbott, letter to ERCOT and the Public Utility Commission of Texas [55]
2.3 Water Becomes an Industrial-Strategy Variable
The water discussion in semiconductor policy has too often been reduced to the observation that fabs consume water, which is true but analytically unhelpful. The important policy question is whether regions can create closed-loop or highly efficient semiconductor water systems capable of supporting larger clusters without placing unsustainable pressure on other users, and the Arizona experience shows that the answer depends on engineering choices, municipal planning and regulatory design as much as on hydrology. Ultrapure water, the feedstock for wafer cleaning and rinsing, must be produced on site from whatever municipal or reclaimed supply is available; the raw-water requirement therefore depends on how much of the process water can be captured, treated and returned to ultrapure quality. TSMC’s progression from 65 percent reuse to a planned 90 percent or better through its Industrial Reclamation Water Plant demonstrates that the gross and net figures can differ by a factor of four. [27,28] The City of Phoenix, for its part, expects to recover a large majority of the water it delivers to the campus through its own wastewater system, which is consistent with the roughly 80 percent recovery that Chandler achieves with Intel. [29]
The institutional dimension is as important as the technical one. Phoenix operates under a hundred-year assured water supply framework, draws from four distinct sources including Colorado River water via the Central Arizona Project and Salt and Verde River water via the Salt River Project, and has structured its industrial pretreatment and discharge permits so that fab wastewater can be reclaimed rather than lost. [29,27] Grist’s reporting in early 2026 found that despite nearly two hundred datacenters and chip factories, non-residential water use in Phoenix had risen only modestly as a share of the city total since 2021. [30] The lesson for a second cluster is that water constraints are less about whether a region is arid and more about whether it has the legal, municipal and engineering capacity to run a semiconductor water loop at scale. Texas, which has its own severe statewide supply shortfall projections and a datacenter boom drawing on the same aquifers and reservoirs, will need the same capacity whether or not it is wetter than Arizona on average.
2.4 Chemicals, Gases and Materials
This is one of the least commonly discussed parts of the semiconductor geography and one of the most strategically important. The leading-edge fab consumes photoresists, specialty etch and deposition gases, wet chemicals such as hydrofluoric acid and isopropyl alcohol at semiconductor grade, polished and epitaxial silicon wafers, substrates for packaging, rare and specialty materials for targets and precursors, and the cleaning and delivery systems that move all of these around a cleanroom without contamination. Each of these supply chains has its own geography, often concentrated in Japan, South Korea and Taiwan, and the question of how much of it must be domesticated to make an American cluster resilient has no settled answer. What is clear from 2026 is that the chipmakers themselves have begun treating upstream domestic suppliers as part of strategic capacity planning rather than as commodity procurement.
Micron’s July 2026 commitment of up to $3 billion to the U.S. supply chain, anchored by $500 million of strategic financing to GlobalWafers’ 300-millimeter raw silicon wafer facility in Sherman, Texas, together with a ten-year supply agreement, is the clearest example. Sherman is the only operating 300-millimeter raw wafer plant in the United States, and GlobalWafers’ chief executive Doris Hsu had publicly set conditions for expanding it, including long-term customer contracts, prepayments and government support. [43] Micron’s financing met those conditions, converting a dependency into contracted domestic capacity. The Texas Semiconductor Innovation Fund’s 2026 grants show the same logic operating at the state level: Soulbrain’s $11.6 million award supports its first U.S. advanced-materials plant in Taylor, next to Samsung’s fab; Schunk Xycarb’s $3.9 million grant doubles a Georgetown facility that manufactures silicon-carbide-coated graphite, quartz and ceramic components used inside wafer-processing tools; and Coherent’s $14.1 million grant supports a $154 million indium phosphide wafer fab in Sherman serving the photonics that connect AI datacenters. [31,34,32] None of these is a fab, and all of them are indispensable to fabs.
“Further build a strong and resilient semiconductor supply chain.” — Governor Greg Abbott [31]
2.5 Semiconductor Construction as a Specialized Industry
Building a leading-edge fab differs from ordinary industrial construction in ways that make construction capacity a genuine regional asset rather than a commodity. The fab floor must be isolated from vibration to tolerances that lithography tools require, which means massive, independently founded waffle slabs and structural systems engineered to damp both ground motion and the movement of thousands of workers and tools. The cleanroom itself is a building within a building, with laminar airflow, filtration and pressure cascades that must be commissioned and certified before a single tool is installed. The mechanical systems that supply ultrapure water, process cooling, exhaust scrubbing and specialty gases are larger and more complex than the production equipment they serve, and the process piping that connects them is installed to orbital-welded, high-purity standards by trades who must be trained and certified for the work. Specialty electrical systems deliver power with the quality and redundancy described above. Contamination control governs every stage of construction, so that a fab under construction looks and operates differently from any other job site.
The implication is important and frequently overlooked: a state can have capital, land and even a signed anchor tenant and still lack the construction ecosystem required to scale semiconductor projects simultaneously. Intel’s Ohio One site in New Albany, where contractors led by Bechtel are building 2.5 million square feet including 600,000 square feet of cleanroom, illustrates the inverse problem of a construction ecosystem assembled for a schedule that then slipped: first announced for 2025 production, the first module is now expected to complete construction in 2030 and begin operations between 2030 and 2031, with the second module operational in 2032. [44] Intel’s foundry chief, Naga Chandrasekaran, explained the decision in terms of demand alignment, and construction has continued at a slower pace, but the episode shows how a region’s investment in specialized construction capacity is exposed to the anchor company’s corporate fortunes. [44]
“Align the start of production of our fabs with the needs of our business.” — Naga Chandrasekaran, Intel Foundry [44]
2.6 Talent Density
Semiconductor labor is not a single market but a stack of markets with different training pipelines, different time constants and different geographic mobility. Research scientists and process-development engineers typically hold graduate degrees and are internationally mobile. Process, electrical, chemical and equipment engineers typically hold bachelor’s or master’s degrees and are regionally mobile. Equipment and tool technicians can be trained in months through community-college or apprenticeship programs and are largely local. Packaging specialists, as Shih and Lin found, are scarce enough in the United States that companies import them. [16] Operators and construction trades are local and are drawn from labor pools that compete with every other large industrial project in the region. The Semiconductor Industry Association’s study with Oxford Economics, still the most cited baseline, projected that the U.S. industry would need to grow from roughly 345,000 workers to about 460,000 by 2030 and that approximately 67,000 of those positions, 39 percent of them technicians, 41 percent engineers and 20 percent computer scientists, risked going unfilled. [51] More recent 2026 analyses by McKinsey and the SEMI Foundation have put the manufacturing-specific shortfall as high as 127,000 to 157,000 workers as AI laboratories compete for the same engineering graduates, with only about 1,500 engineers entering semiconductor manufacturing annually. [53] Samsung’s Jon Taylor, executive vice president of the company’s Austin semiconductor operations, told CNBC in September 2026 that he was concerned about staffing the Taylor fab. [54]
“Tens of thousands of new post-secondary-trained workers will need to fill the roles.” — Dan Martin, Oxford Economics [51]
Universities and community colleges therefore become semiconductor infrastructure in the literal sense, and the regions that have treated them as such are visible in the data. The Maricopa County Community College District’s ten-day Semiconductor Technician Quick Start program, developed with Intel and TSMC and taught by fab workers, has become a national model for the technician layer. [52] Arizona State University has become one of Intel’s largest sources of U.S. graduates and in May 2026 joined Stanford and Rensselaer Polytechnic Institute as an inaugural academic partner in Applied Materials’ EPIC Center, the largest U.S. investment in semiconductor equipment research. [61] Texas, with the University of Texas system, Texas A&M and an extensive community-college network, begins from a comparably strong base, and the Texas Semiconductor Innovation Consortium was explicitly designed to channel higher-education capacity toward industry needs.
2.7 The University–Fab Feedback Loop
The relationship between universities and fabs is best understood as a feedback loop rather than a one-way supply of graduates, and the loop is what makes an established cluster so difficult for competitors to reproduce. Research partnerships produce talent with hands-on experience on industry-relevant tools. Fabs create employment that makes the region attractive to students choosing where to study and to faculty choosing where to build laboratories. Employment and faculty attract federal and corporate research funding, which follows industrial concentration because funders want their research to be used. Suppliers that locate near the fabs create additional engineering demand and additional sponsored research. Each turn of the loop increases the region’s density of tacit knowledge, the kind of know-how that is transmitted through people rather than documents and that Moretti’s research identifies as the source of agglomeration’s productivity premium. [24] Over a decade or more, the region becomes a place where semiconductor manufacturing is a normal career path rather than an exotic one, and that normality is itself a competitive asset that cannot be purchased with a grant.
2.8 From Incentive Competition to Ecosystem Competition
Section 2 ends with one of the paper’s central propositions, stated as plainly as possible: the first semiconductor subsidy may attract a plant, but the ecosystem determines whether the second, third and fourth plants follow. The CHIPS Act’s own early evidence supports this reading. The Brookings Papers on Economic Activity study by Bilge Erten of Northeastern University, Nobel laureate Joseph Stiglitz of Columbia University and Eric Verhoogen of Columbia found that the Act directly created roughly 15,000 to 16,000 jobs in the core semiconductor sector and 28,000 to 35,000 indirect jobs in construction and related industries through the end of 2024, with the effects concentrated in counties that already had semiconductor facilities and beginning in anticipation of the legislation rather than after its signing. [22,23] The authors were explicit that the Act did not reduce spatial inequality in where the industry is located; it reinforced existing clusters. [22] That finding is uncomfortable for states hoping that a subsidy alone can conjure an industry from nothing, and it is precisely why Texas’s existing base matters so much to the evaluation that Section 3 takes up.
“Industrial policies can deliver measurable employment benefits in targeted strategic sectors.” — Erten, Stiglitz and Verhoogen, Brookings Papers on Economic Activity [22]
“Companies aren’t simply getting money to do something they would’ve done anyway.” — Eric Verhoogen, Columbia University [23]
Section 3: What Would a Second TSMC State Mean?
The Reuters report of September 30 is the paper’s opening anecdote, and this section returns to it directly, but with the analytical vocabulary of Sections 1 and 2 now available. The purpose of the section is not to predict whether TSMC will build in Texas; the sources themselves emphasized that the plans had not been finalized, and that caveat should remain explicit throughout. [1] The purpose is instead to ask what the mere evaluation reveals about how the world’s largest foundry thinks about American geography, what Texas offers that Arizona does not, what Arizona offers that Texas cannot, and why the question of Arizona versus Texas is, on close inspection, the wrong question to ask. The deeper claim is that if a second major TSMC state ever materializes, its strategic significance would be greater than the additional wafer output it provides, because it would mark the point at which American advanced manufacturing progressed from a project to a cluster to a network.
3.1 The September 30 Texas Signal
The facts of the report are few and should be stated precisely. Two people familiar with the matter told Reuters that TSMC was evaluating a potential investment in Texas to expand U.S. chip production; one of them said the potential investment would be in addition to the $265 billion committed to Arizona; both said the plans had not been finalized; TSMC did not immediately comment; and Chief Financial Officer Wendell Huang had told Reuters in July that the company would continue to invest in the United States. [1,2] The size, timing, process node and purpose of any Texas project remain unknown. It could be a leading-edge logic fab, an advanced packaging complex, a specialty-technology fab, a research facility or some combination; it could be sited in the Austin-Taylor corridor near Samsung and Soulbrain, in the Dallas-Sherman corridor near Texas Instruments, GlobalWafers and Coherent, or somewhere else entirely.
The analytical value of the report does not depend on resolving any of these unknowns. The mere evaluation asks an important question that would not have been askable in 2022: why might the world’s largest foundry want more than one major U.S. geography? A company that was merely satisfying political demands for American investment would keep adding to the campus it already understands, where the permits, the water agreement, the utility relationships, the supplier base and the trained workforce already exist and where each incremental fab is cheaper than the last. A company that is evaluating a second state is, by implication, weighing something other than incremental cost, and the candidates for that something are the subject of the next subsection.
3.2 From Concentration Economics to Diversification Economics
Arizona offers enormous cluster benefits, and nothing in this section should be read as minimizing them. But a second region could offer a different set of benefits that concentration cannot provide, and it is useful to enumerate them because they define what a second cluster would have to deliver to justify forgoing Arizona’s agglomeration premium. The first is geographic redundancy against physical and infrastructural risk: a single campus, however large, is exposed to a single utility system, a single municipal water supply, a single transmission corridor and a single regional labor market, and the entire logic of reshoring was to reduce exposure to single points of failure. The second is workforce diversification, drawing on a second state’s universities and community colleges rather than competing with Intel, Amkor and the equipment vendors for the same Phoenix engineers. The third is access to additional utility systems and, in Texas’s case, to a state with very large generation capacity and a political culture oriented toward building more of it. The fourth is proximity to a different set of suppliers and, critically, customers: Tesla’s Austin operations, Texas Instruments, the Dallas and Austin design houses, the defense and aerospace primes, and the hyperscale datacenter clusters in Dallas, Austin, Houston and San Antonio. The fifth is political diversification, in the sense that a company with major operations in two of the nation’s largest states has a broader base of congressional and gubernatorial support than a company with operations in one. The sixth is construction capacity, since Texas’s industrial construction market is among the largest in the country and would not have to compete with the Phoenix pipeline. The seventh is future expansion optionality: a second campus with its own land bank gives the company two places to add the next fab rather than one.
This paper introduces the concept of domestic geographic redundancy to describe the strategic logic that these benefits share. Traditionally, semiconductor diversification meant Taiwan versus the United States, or more precisely Taiwan versus anywhere else. The next version could mean Arizona plus Texas plus other specialized U.S. nodes, a diversification inside the diversification. Chris Miller observed in a March 2026 lecture at Carnegie Mellon that the primary concern about semiconductors today is access but that the defining question of tomorrow will be trust, and a foundry with redundant domestic capacity in two states is better positioned to offer both to American customers than a foundry with one very large campus. [20]
3.3 Texas Begins With an Existing Semiconductor Base
Texas would not be starting from zero, and this is the single most important fact distinguishing a Texas evaluation from the dozens of state pitches that TSMC presumably receives and declines. Samsung has manufactured chips in Austin since 1996, holds a CHIPS award of up to $4.745 billion to support a Central Texas project expected to exceed $37 billion in capital expenditure including two leading-edge logic fabs and an R&D fab in Taylor, and in September 2026 began production at Taylor of Tesla’s AI5 chip on its two-nanometer process, moving the fab’s utilization toward thirty percent ahead of an originally planned November start under a $16.5 billion Tesla supply contract. [35,36,37] Texas Instruments, the largest analog and embedded-processing semiconductor manufacturer in the United States, started production at SM1 in Sherman in December 2025, the first of four planned 300-millimeter fabs at a site that will receive up to $40 billion of the company’s $60 billion U.S. investment plan across seven fabs in Texas and Utah, supported by up to $1.6 billion in CHIPS funding. [38,39,64] The Austin corridor hosts one of the country’s densest concentrations of semiconductor design talent, including Arm’s expanding research operations, AMD’s large Austin campus, Apple’s Austin engineering center and a long tail of design houses. GlobalWafers operates the nation’s only 300-millimeter raw wafer plant in Sherman. Coherent and Applied Optoelectronics are building photonics capacity directly relevant to AI connectivity. The defense and aerospace demand base is substantial. Electricity generation is the largest of any state. And the state possesses established advanced-manufacturing corridors from Dallas to Austin to Houston.
“Owning every part of the manufacturing process.” — Haviv Ilan, President and CEO, Texas Instruments [64]
The Texas Semiconductor Innovation Fund, created when Governor Abbott signed the Texas CHIPS Act in 2023 with roughly $698 million in initial appropriations and replenished with another $250 million in 2025, now has approximately $948 million in cumulative appropriations, which is small relative to federal awards but is being deployed in a manner that reveals a deliberate ecosystem strategy. [31,34] A list of the 2026 grants reads like a deliberate attempt to fill the hidden bill of materials in Table 2: a materials supplier in Taylor, a tool-component supplier in Georgetown, a compound-semiconductor wafer fab in Sherman, an optical-transceiver manufacturer in Sugar Land. [31,32,33,34]
3.4 The Texas Semiconductor Innovation Fund as State Industrial Policy
It is tempting to read Texas’s semiconductor strategy through the familiar lens of low taxes and light regulation, and those features are real and relevant. But the Texas CHIPS Act created something more specific: an explicit state industrial policy for semiconductors, administered by a Texas CHIPS Office inside the Governor’s economic development apparatus, advised by a Texas Semiconductor Innovation Consortium drawn from industry and higher education, and empowered to make grants across research, design, manufacturing, materials, workforce and services. [31,32] The state’s own framing of the fund’s purposes, to leverage Texas’s semiconductor investments, encourage semiconductor-related companies to expand, develop the expertise of the state’s institutions of higher education and maintain the state’s position in manufacturing, is a portfolio description rather than a fab-recruitment description. [34] That diversity strongly supports the Foundry Federalism thesis. States are beginning to manage semiconductor ecosystems as portfolios, allocating modest state dollars to the suppliers, materials and workforce programs that make large private and federal dollars productive, rather than concentrating everything on landing a single marquee fab.
The contrast with the early CHIPS Act design is instructive. Washington’s manufacturing incentives flowed overwhelmingly to fabs, and the Erten-Stiglitz-Verhoogen study found correspondingly that the employment effects were concentrated in counties that already hosted fabs. [22] Texas’s fund, operating at one-fiftieth the scale, has been allocating a far larger proportion to the connective tissue around fabs. Whether this reflects strategic insight or simply the fact that a $948 million fund cannot buy a fab is a fair question, but the effect is the same: Texas is building the parts of the ecosystem that a foundry would otherwise have to import.
3.5 Arizona Versus Texas Is the Wrong Question
Much commentary after the Reuters report framed the Texas evaluation as a contest, and the framing is natural because state economic development is organized competitively. But it is the wrong question for understanding what is actually happening, and a more interesting question is what comparative advantage each cluster could develop if both continue to grow. Arizona may deepen frontier logic fabrication, advanced packaging through the TSMC-Amkor-Nvidia triangle, the Taiwanese supplier networks that have followed TSMC to Phoenix, and high-volume manufacturing of the most advanced nodes, consolidating its position as the place where the newest process technology first reaches American production. Texas may combine foundry manufacturing at Samsung and potentially TSMC with chip design in Austin, analog and mature-node production at Texas Instruments and NXP, semiconductor materials at GlobalWafers and Soulbrain, optical networking at Coherent and Applied Optoelectronics, test and tool-component manufacturing, aerospace and defense demand, and one of the largest AI datacenter concentrations in the country. These are not the same profile, and a national system in which both exist is more capable than one in which either is simply a larger version of the other. Federalism permits specialization rather than requiring duplication, and the paper’s argument is that specialization is the more resilient outcome.
3.6 A Two-Cluster TSMC America
The hypothetical must be developed carefully because it is only a hypothetical. If TSMC ultimately built major capacity in a second U.S. state, the strategic significance could be greater than the additional wafer output for three reasons. First, it would demonstrate that the Arizona playbook is transferable, which would change the risk calculation for every other state and every other foundry considering American investment. Second, it would create the first genuinely interstate advanced-logic supply chain within the United States, with wafers, packaging, materials and engineers moving between two clusters under one corporate roof, and that movement would itself become a form of national infrastructure. Third, it would indicate that advanced semiconductor manufacturing in America had progressed through three distinct stages: from project, a single fab whose success was uncertain; to cluster, a campus with an ecosystem forming around it; to network, multiple clusters with complementary roles connected by domestic logistics. That progression would represent a fundamental maturation of U.S. semiconductor geography, and it is the empirical signature that the remainder of this paper watches for.
The honest counterargument should also be stated. Willy Shih has long warned that market forces continue to operate regardless of policy intent, that U.S. fab construction and operating costs remain higher than in Asia, and that he favors demand-side incentives that would cause customers to preferentially buy American-made chips while the new fabs improve their cost positions. [17] A second TSMC cluster would face every one of the ecosystem gaps that Section 2 described, at least initially, and it would face them without the first-mover advantages that Arizona accumulated. Whether the benefits of domestic geographic redundancy outweigh the costs of forgoing Arizona’s agglomeration premium is an empirical question that TSMC’s board, not this paper, will answer. What the paper can say is that the fact the question is being seriously asked is itself the news.
“Market forces are still at work.” — Willy C. Shih, Harvard Business School [17]
Section 4: The Emerging Semiconductor Federation
If Sections 1 through 3 established that clusters are more than fabs and that a second TSMC state would signal the transition from cluster to network, this section attempts to draw the network as it currently exists in late 2026, not as a prediction but as a description of the roles that different states are already, in practice, beginning to play. The exercise is useful for two reasons. First, it corrects the tendency to discuss American semiconductor policy as though it were a single national project with a single national outcome, when in fact it is already a patchwork of regional projects with different technologies, different anchor companies, different timelines and different exposures to the AI demand cycle. Second, it allows the paper to pose the question of redundancy versus waste concretely, by asking of each region not whether it has a leading-edge fab but which part of the semiconductor system it could realistically host. The deep explanation here deliberately resists the temptation to assign every state a role; the purpose is to demonstrate that national resilience may emerge from complementary regional specialization rather than identical regional duplication.
4.1 Arizona: Leading-Edge Logic and Advanced Packaging
Arizona is the first regional archetype and the most fully developed. Its potential specialization is leading-edge logic plus advanced packaging plus research and development plus the localization of TSMC’s Taiwanese supplier base. The twelve-facility TSMC plan, Intel’s Fabs 52 and 62 in Chandler producing Intel 18A, Amkor’s Peoria campus, the CHIPS for America advanced packaging research facility, Applied Materials’ $270 million Materials-to-Fab Center at Arizona State University’s Research Park, and the more than seventy supplier expansions since 2020 together constitute something that no other American region currently possesses: a complete front-end and emerging back-end ecosystem for the most advanced process nodes. [3,13,61] The region’s constraints, water, power, construction labor and engineering talent, are the constraints of success, and the state’s institutions have so far managed them with a degree of long-horizon planning that Castanares credited as a reason for TSMC’s original site choice. [28]
4.2 Texas: A Diversified Semiconductor Industrial System
Texas is the second archetype and the most diversified. Its potential specialization is logic plus design plus analog plus materials plus photonics plus equipment components plus test plus potential additional leading-edge foundry capacity. Samsung’s Taylor fab is the state’s leading-edge anchor; the CHIPS award of up to $4.745 billion supports two leading-edge logic fabs and an R&D fab in Taylor plus an Austin expansion, and the company’s own disclosures describe a staged ramp with Fab 1 operating in 2026, customer mass production in 2027 and a second Taylor fab under construction. [35,36] Texas Instruments’ Sherman megasite is the state’s foundational-semiconductor anchor, producing the analog and embedded chips that go into vehicles, industrial systems, medical devices and the power-delivery subsystems of datacenters. [38,39] Around these anchors the state’s fund is assembling materials, components and photonics, and the state’s datacenter boom is creating a local demand base. The limiting factor is the grid: the same abundance that attracts datacenters is producing an interconnection queue of more than 438 gigawatts that ERCOT and the Public Utility Commission are only beginning to manage through the batch-study process created under Senate Bill 6. [56,57]
4.3 Idaho: Memory Research Meets Manufacturing
Micron’s Boise strategy is the third archetype and an unusually clean example of the university-fab feedback loop described in Section 2.7, because the new fabs are co-located with the company’s existing research and development operations. Construction of the first Idaho fab, ID1, began in October 2023, and the company’s June 2026 Form 10-Q projects first DRAM wafer output in mid-calendar 2027; construction of the second Idaho fab begins in 2026 with operations expected by the end of 2028. [40] Micron amended its $6.1 billion CHIPS direct-funding agreement in June 2025 to add the second Idaho fab and reallocate funding toward it, accelerating Idaho relative to New York. [40] The company’s fiscal third-quarter results, revenue of $41.46 billion against $9.30 billion a year earlier and operating cash flow of $25.39 billion, together with a fiscal 2026 capital expenditure plan above $25 billion net of government incentives, show how the AI-driven memory super-cycle is financing the Idaho buildout. [41,40] Chief Executive Sanjay Mehrotra has framed the expansion directly in terms of high-bandwidth memory demand. Idaho’s potential specialization is therefore memory research plus DRAM plus HBM-related manufacturing knowledge, and its distinctive asset is the proximity of process development to volume production.
“Reflect the strategic value of memory in the AI era.” — Sanjay Mehrotra, Chairman, President and CEO, Micron Technology [41]
4.4 New York: Memory at Gigafab Scale
Micron’s Clay project in Onondaga County creates the fourth archetype, defined by scale and duration rather than by research co-location. The site, where the first concrete was poured in July 2026 more than a quarter ahead of schedule, is planned to eventually hold four fabs forming one of the world’s largest memory-manufacturing complexes, with the first fab providing supply in 2030 and beyond and the later fabs extending into the 2030s and 2040s. [42,40] The project is supported by the federal direct-funding agreement and by a non-binding term sheet with the State of New York for up to $5.5 billion over more than twenty years through a combination of tax credits for qualified capital investment and incentives tied to new job wages under the state’s GREEN CHIPS program. [40] Brookings researchers Joseph Parilla, Mark Muro and Xavier de Souza Briggs identified Central New York early as a test of whether the CHIPS Act could deliver regional economic revitalization rather than merely fab capacity, and the twenty-year horizon of the state’s commitment is itself a form of policy durability that corporate planners value. [60] New York’s potential specialization is high-volume leading-edge memory manufacturing, and its exposure, as Shih has warned, is to the historically brutal memory cycle, in which the right time to build the fab is always earlier than the moment the market makes it feel urgent. [18]
“It has already almost doubled the price for memory that goes into servers.” — Willy C. Shih, Harvard Business School [18]
4.5 Other Semiconductor Geographies
The map must be broadened beyond the four archetypes, because the Foundry Federalism thesis does not depend on four states but on a national distribution of roles. Ohio’s Intel One site, delayed but under continued construction, is the clearest example of a region that assembled a construction and workforce ecosystem around an anchor whose timeline then moved by half a decade, and it poses the question of how a federation should treat a cluster whose anchor is slow. [44] Oregon remains Intel’s primary process-development geography and the home of its most advanced research fabs, which makes it the natural locus for process research and semiconductor engineering regardless of where volume production lands. California remains the center of architecture, design and AI-system development, and Applied Materials’ EPIC Center in Silicon Valley, with Stanford, ASU and RPI as academic partners, is an attempt to re-anchor equipment research in the state that invented the industry. [61] New Mexico hosts Intel’s Rio Rancho advanced packaging operations. Virginia hosts Micron’s Manassas fab, which received up to $275 million in CHIPS funding for modernization and began producing 1-alpha DRAM in May 2026, as well as the largest datacenter concentration on earth in Loudoun County. [40,43] Utah hosts Texas Instruments’ Lehi fabs. Indiana is becoming a high-bandwidth-memory packaging location through SK Hynix’s West Lafayette investment near Purdue University, which has built one of the country’s most ambitious semiconductor degree programs. [53] North Carolina and Michigan have pursued wide-bandgap and automotive semiconductor roles respectively.
The paper does not argue that every state needs a leading-edge fab; the opposite is true. It asks instead which part of the semiconductor system each state could realistically host, given its existing industrial base, its universities, its utilities and its proximity to customers. The Stanford Emerging Technology Review’s 2026 assessment of semiconductors emphasizes that securing American leadership requires manufacturing capacity, sustained innovation and talent in combination, and that framing applies as much to the division of labor among states as to the nation as a whole. [62]
4.6 From Megafabs to Specialized Regions
Drawing these threads together produces the possibility of a national division of semiconductor labor, presented in Table 3 as a description of emerging tendencies rather than a prediction of an exact map. The purpose of the table is to demonstrate that national semiconductor resilience may emerge from complementary regional specialization rather than identical regional duplication, and that the roles are already differentiating in practice through corporate decisions, state policies and the gravitational pull of existing institutions.
Table 3. An emerging national division of semiconductor labor, late 2026
| Region | Emerging specialization | Principal anchors | Principal constraint |
| Arizona | Frontier logic, advanced packaging, Taiwanese supplier localization, R&D | TSMC (12 facilities), Intel Chandler, Amkor Peoria, ASU [3,13] | Water, power, construction and engineering labor |
| Texas | Diversified: leading-edge foundry, design, analog, materials, photonics, test, datacenters | Samsung Taylor, TI Sherman, GlobalWafers, Coherent, AOI, Soulbrain [35,38,43] | Grid interconnection queue; workforce |
| Idaho | Memory R&D co-located with DRAM/HBM production | Micron ID1 (2027), ID2 (2028) [40] | Single-company dependence; memory cycle |
| New York | High-volume leading-edge DRAM at gigafab scale | Micron Clay (first fab 2030+); GlobalFoundries Malta [40,42] | Long timeline; memory cycle |
| Oregon | Process research and engineering | Intel Hillsboro development fabs | Anchor company’s corporate health |
| California | Architecture, design, AI systems, equipment R&D | Nvidia, AMD, Apple, Applied EPIC Center [61] | Cost; limited volume manufacturing |
| Ohio | Future leading-edge logic (delayed) | Intel Ohio One (2030–2032) [44] | Anchor timeline slippage |
| Virginia | Specialty DRAM; hyperscale datacenter consumption | Micron Manassas; Loudoun County datacenters [40] | Power |
| Indiana | HBM packaging and university pipeline | SK Hynix West Lafayette; Purdue [53] | New ecosystem |
Other regions could specialize in materials, power electronics, equipment, substrates, photonics, defense chips, advanced packaging or test, and the table is deliberately incomplete. Its value lies in making visible the pattern that is already forming: no two of the major regions have the same profile, and the profiles are complementary in the specific sense that a product built from American silicon would draw on several of them.
4.7 Interstate Semiconductor Corridors
Once the regions are specialized, the connections between them become economic infrastructure in their own right, and future semiconductor policy may need to consider logistics and interdependence among states as seriously as it now considers the siting of individual facilities. Consider the path of a hypothetical AI accelerator built entirely within the United States by the end of the decade. It would be designed in California, fabricated on a two-nanometer-class process in Arizona or Texas, packaged in Arizona at TSMC’s or Amkor’s facilities, combined with high-bandwidth memory whose DRAM was produced in Idaho or New York and whose stacks were assembled in Indiana, mounted on boards and installed in servers assembled elsewhere, and ultimately deployed in a Texas or Virginia AI datacenter. Each arrow in that path crosses a state line. Each depends on wafers, dies and modules moving under contamination-controlled conditions between facilities that are hundreds or thousands of miles apart. The semiconductor supply chain therefore becomes interstate economic infrastructure, and the regions become nodes in a network whose resilience depends on the links as much as on the nodes.
This has practical implications that have received little attention. Air cargo capacity at Phoenix, Austin, Dallas and Boise becomes semiconductor infrastructure. The ability of a packaging facility in Arizona to receive wafers from a Texas fab on the same timeline it receives them from the adjacent TSMC campus becomes a design parameter for the national network. The qualification of a second domestic source for a critical material becomes a multi-state coordination problem rather than a single-state recruitment problem. And the data that would allow policymakers to see the network as a whole, which facility supplies which, with what lead times and what single points of failure, largely does not exist in any public form.
4.8 Federalism as Redundancy, or Federalism as Waste
Federalism creates inefficiencies, and the paper would be dishonest if it pretended otherwise. States duplicate incentives, bidding against one another for the same mobile investment and transferring surplus from taxpayers to firms that would have located somewhere in the country regardless. States duplicate institutions, each building its own semiconductor workforce office, its own innovation consortium, its own research center, when a shared national institution might serve the same function at lower cost. Smaller clusters forgo the agglomeration productivity that Moretti’s research documents. [24] And geographic dispersion can strand investment when an anchor’s timeline slips, as Ohio has learned. [44]
But geographic dispersion can also create resilience, and the paper’s central research problem is determining when redundancy becomes waste and when redundancy becomes resilience. Both sides deserve neutral examination. Redundancy is waste when two states subsidize identical capabilities that serve the same customers from the same supply chains, so that the second facility adds no protection against the failure of the first and merely divides a fixed demand. Redundancy is resilience when a second facility draws on a different utility system, a different water supply, a different labor market and a different supplier base, so that a disruption affecting one region leaves the other operating, and when the two facilities together give customers a credible domestic second source. The distinction maps onto the distinction between duplication and specialization that runs through this paper. A second TSMC cluster in Texas that simply replicated the Arizona campus would be closer to the waste end of the spectrum; a Texas cluster that combined TSMC capacity with Samsung, Texas Instruments, GlobalWafers, Austin design and the ERCOT datacenter base would be closer to the resilience end. Mark Muro and Bruce Katz’s classic Brookings work on regional innovation clusters defined them as geographic concentrations of interconnected businesses, suppliers, service providers and associated institutions including universities and community colleges, and the definition is useful here precisely because it locates the value of a cluster in its interconnections rather than in any single facility. [25]
Section 5: From Federal Semiconductor Policy to the 2030 AI Geography
The final analytical section assembles the layers of government and the layers of the AI economy into a single account. It explains what the federal government does and cannot do, what states do and cannot do, how the toolkit available to states has evolved, how AI demand is reshaping the geography of chip manufacturing, whether chip geography and datacenter geography are converging, what the 2027 to 2030 map might look like under four scenarios, and what the analysis implies for corporate strategists, governors and federal policymakers. The section is deliberately the longest in the paper because it is where the descriptive material of Sections 1 through 4 is converted into the framework that gives the title its meaning.
5.1 Washington Sets the Strategic Floor
The federal role in semiconductor geography is foundational but bounded. Washington controls national security policy, including the export controls that have shaped which chips can be sold where; federal research funding through the National Science Foundation, the Department of Energy’s national laboratories and the CHIPS research programs; the manufacturing incentives administered by the CHIPS Program Office; trade policy including tariffs and the reciprocal-tariff framework under which TSMC’s March 2025 expansion was negotiated; tax policy including Section 48D; defense procurement, which creates demand for trusted domestic sources; and supply-chain security authorities. The CHIPS program demonstrated how federal funding can catalyze regional clusters: the original TSMC Arizona award of up to $6.6 billion in direct funding supported a $65 billion plan that has since quadrupled, and the Erten-Stiglitz-Verhoogen findings show measurable employment effects in fab counties beginning even before the law was signed. [45,22]
The federal floor has also shifted in character since 2025 in ways that matter for how states plan. The Commerce Department under Secretary Howard Lutnick converted Intel’s grant into an $8.9 billion equity investment for a 9.9 percent non-voting stake, explored equity in other CHIPS recipients before indicating that TSMC and Micron, having expanded their U.S. commitments, would not be asked for equity, and in July 2026 signed letters of intent to provide $874 million to seven smaller compute-supply-chain companies on condition of minority equity stakes. [46,48,47] The Congressional Research Service’s July 2026 review of CHIPS fab projects notes that TSMC and Micron have increased their capital commitments while other recipients have not, and that the department’s incentive program for semiconductor materials and equipment was accepting concept plans until November 2026. [45] Meanwhile the Section 48D credit, now 35 percent for property placed in service after December 31, 2025, terminates for property whose construction begins after December 31, 2026, and the Semiconductor Industry Association with seventeen allied trade groups has urged Congress to enact a multi-year extension and to broaden the credit to design and research. [49,50] The expiration deadline is itself a geographic force: it compresses groundbreaking decisions into 2026 and rewards states whose permitting and site readiness allow construction to begin before the cliff.
“We should get an equity stake for our money.” — Howard Lutnick, U.S. Secretary of Commerce [46]
5.2 States Build Above the Federal Floor
State governments determine many of the conditions that ultimately decide whether industrial projects can actually operate, and this is the heart of Foundry Federalism. States and their municipalities control sites and the zoning that governs them, the roads and interchanges that serve them, the utility franchises and public utility commissions that regulate electricity, the water agreements and industrial discharge permits that make reclamation possible, the workforce programs that train technicians, the environmental permitting that governs construction timelines, the university systems that produce engineers, the local property-tax abatements and infrastructure contributions that constitute most state incentive packages, and increasingly the datacenter and large-load policies that determine whether chip customers can locate nearby. Washington can establish national strategy and can write very large checks. It cannot manufacture a regional ecosystem by itself, because the ecosystem consists of institutions, permits, pipes, wires, people and relationships that are created and governed at the state and local level. The Arizona water story is the clearest illustration: no federal program created Phoenix’s hundred-year supply framework, its four-source water portfolio or its industrial pretreatment regime, yet those state and municipal institutions were, by TSMC’s own account, among the reasons the company chose the site. [28,29]
5.3 The New State Semiconductor Toolkit
The instruments available to states have expanded considerably since 2020, and it is useful to compare them without ranking the political actors who deploy them, evaluating outcomes rather than branding. The toolkit now includes direct grants such as those of the Texas Semiconductor Innovation Fund; refundable and transferable tax credits such as New York’s GREEN CHIPS; infrastructure investment in water, roads and transmission dedicated to industrial sites; university funding for semiconductor research centers and degree programs; workforce training through community colleges and apprenticeships; permitting reform to compress environmental and construction approvals; utility arrangements including dedicated substations, rate structures and reclaimed-water agreements; industrial parks and megasites assembled in advance of a tenant; supplier recruitment programs that target the hidden bill of materials rather than the anchor fab; and community-benefit commitments that address housing, schools and local hiring. Table 4 summarizes the instruments and the outcomes against which they can be evaluated.
Table 4. The state semiconductor toolkit and how to evaluate it
| Instrument | Illustrative use, 2023–2026 | Outcome measure |
| Direct grants to suppliers | Texas TSIF: Soulbrain $11.6M, Coherent $14.1M, AOI $20.9M, Xycarb $3.9M [31,32,33,34] | Share of hidden bill of materials sourced in-state |
| Long-duration tax credits | New York GREEN CHIPS: up to $5.5B over 20+ years for Micron Clay [40] | Policy durability; second- and third-fab commitments |
| Water infrastructure and permits | Phoenix 11.4M gpd agreement; TSMC reclamation to 90%+ [27,28] | Net municipal water per wafer; reuse rate |
| Large-load grid rules | Texas SB 6: 75 MW threshold, cost contribution, curtailment [56] | Interconnection time; reliability events |
| Workforce pipelines | Maricopa Quick Start; ASU; Purdue; UT system [52,61] | Vacancy duration; share of local hires |
| Research partnerships | Applied Materials Materials-to-Fab Center at ASU; EPIC Center [61] | Sponsored research; process transfers to fabs |
| Land assembly | TSMC 900-acre state land purchase, $197M [10] | Expansion optionality |
| Community-benefit commitments | Housing, schools, local hiring in Phoenix and Clay [60] | Local support; permitting friction |
5.4 AI Demand Changes Semiconductor Geography
The connection between this paper and the Five-Layer AI Economy becomes explicit here. Artificial intelligence increases demand not simply for graphics processors but for an entire system of silicon: CPUs that orchestrate accelerators, high-bandwidth memory that feeds them, networking chips and optical transceivers that connect them, storage, power semiconductors that convert and regulate the enormous currents that datacenters draw, switches and custom accelerators designed by hyperscalers themselves. Each of these categories has its own manufacturing geography. Nvidia’s second-quarter fiscal 2027 datacenter revenue of $89.0 billion, up 117 percent from a year earlier, is the demand signal at the top of the chain; Micron’s record memory results and TSMC’s 53 percent August revenue growth are the same signal propagating to Layer 2; Coherent’s indium phosphide fab and Applied Optoelectronics’ transceiver expansion in Texas are the signal reaching the optical layer; and Texas Instruments’ Sherman analog fabs are the signal reaching power delivery. [59,41,9,32,33,39] Layer 4 and Layer 5 growth therefore pull investment backward toward Layer 2 manufacturing, and the pull is distributed across states according to which part of the silicon system each state produces.
This creates a feedback loop that can be written as a chain: more models produce more inference, which requires more datacenters, which require more chips, which require more fabs, which require more power, water, talent and infrastructure. Each link in the chain has a geography and each geography has a governor. The loop also has a cyclical risk that the memory industry knows better than any other, and Shih’s warning that curves which go to the sky never continue forever applies to the fab-building enthusiasm of states as much as to the share prices of chipmakers. [18] A state that builds its ecosystem around the demand peak of a cycle will face the trough with stranded infrastructure, and the states with the most durable positions are those whose specializations span multiple demand sources rather than depending on a single AI product category.
5.5 When Semiconductor Geography and Datacenter Geography Converge
Do chip manufacturing and AI computing increasingly locate near some of the same regions? There are reasons to expect convergence. Both require reliable electricity at large scale, skilled technical labor, industrial infrastructure, very large capital investments, supplier networks and political support, and both are drawn to states that have made those inputs available. Arizona hosts both TSMC’s fabs and a rapidly growing hyperscale datacenter population; Texas hosts Samsung, Texas Instruments and more than six hundred datacenters; Virginia hosts Micron’s Manassas fab and Loudoun County’s datacenter alley. [30,56,40] The IEA’s observation that datacenters will account for nearly half of U.S. electricity demand growth through 2030 means that fabs and datacenters are competing for the same incremental megawatts in the same states. [58]
But there are equally strong reasons they may remain geographically separate, and these reasons support the specialized interstate roles that Section 4 described. Datacenters can tolerate far greater physical distance from chip factories than packaging facilities or materials suppliers can; a GPU manufactured in Phoenix and packaged in Peoria can be installed in a server in Dallas, Ashburn or Columbus with no loss of performance, whereas a wafer that must travel a thousand miles to reach a packaging line incurs cost, risk and time at every step. Datacenters are also drawn to fiber routes, land prices, tax regimes and latency requirements that fabs are indifferent to. The practical consequence is that the tight-coupled parts of the semiconductor system, fabrication, packaging, test, chemicals and equipment service, will tend to concentrate within a region, while the loose-coupled parts, design, memory, final assembly and consumption, can distribute across the federation. Convergence will be partial, and the partiality is what makes specialization possible.
5.6 The 2027–2030 Semiconductor Map: Four Scenarios
The paper develops scenarios rather than predictions because the variables are too numerous and too interdependent for confident forecasting. Table 5 presents four, and the paragraphs that follow evaluate each rather than presuming the outcome.
Table 5. Four scenarios for the U.S. semiconductor map, 2027–2030
| Scenario | Description | Signals that it is unfolding | Principal risk |
| A. Continued concentration | Arizona absorbs most incremental frontier-foundry investment; Texas evaluation does not proceed at scale | Further TSMC Arizona fab announcements; Texas limited to Samsung and TI | Single-region dependence; water and power ceilings |
| B. Second-cluster formation | Texas or another state develops a second major leading-edge cluster | TSMC Texas final investment decision; Samsung Taylor Fab 2; supplier migration | Diluted agglomeration; duplicated incentives |
| C. Functional specialization | Different states dominate different stages of the stack | Idaho/New York memory ramps; Indiana HBM packaging; Texas photonics and materials | Interstate logistics fragility |
| D. Distributed federation | Multiple specialized clusters connected through domestic supply chains under one national framework | Interstate wafer and die flows; shared qualification; federal network data | Coordination failure; cyclical over-build |
Scenario A, continued concentration, is the baseline implied by agglomeration economics and by the sheer momentum of the Arizona commitment. Under this scenario, TSMC adds its next fabs to the existing campus, Amkor and TSMC complete the packaging buildout, the supplier base densifies further, and Texas remains a Samsung and Texas Instruments state with a strong design and materials base but no second TSMC-scale anchor. The scenario is efficient and plausible, and its principal risk is the one that the reshoring project was designed to avoid, namely dependence on a single region whose water, power and labor ceilings are already visible.
Scenario B, second-cluster formation, is the scenario that the Reuters report made thinkable. Under it, TSMC makes a final investment decision in Texas, Samsung proceeds with its second Taylor fab, and the supplier migration that followed TSMC to Phoenix begins to follow it to Central or North Texas. The scenario sacrifices some agglomeration productivity and invites duplicated incentives, but it delivers domestic geographic redundancy at the leading edge. The scenario is distinguishable from Scenario D by whether the second cluster is a copy or a complement.
Scenario C, functional specialization, is already partially underway in memory, where Idaho and New York are differentiating by timeline and function, and in packaging, where Arizona and Indiana are differentiating by product. Under this scenario, the stack divides geographically without any single state achieving a second full leading-edge logic cluster, and the national system’s resilience depends on the interstate corridors described in Section 4.7.
Scenario D, the distributed semiconductor federation, most fully expresses Foundry Federalism. Under it, multiple specialized clusters, Arizona for frontier logic and packaging, Texas for diversified manufacturing and design, Idaho and New York for memory, Oregon and California for research and architecture, Indiana for HBM packaging and others for materials, power electronics and defense, are connected through domestic supply chains and coordinated through a federal framework that treats the network, rather than the individual facility, as the unit of strategic planning. The scenario’s risks are coordination failure among fifty states and the cyclical over-build that distributed enthusiasm can produce. The paper evaluates all four rather than presuming the fourth, and it notes that the scenarios are not mutually exclusive: the most likely 2030 map contains elements of A in Arizona, B or C in Texas, C in memory and the beginnings of D in the corridors that connect them.
5.7 Implications for Corporate Strategy
Companies should increasingly evaluate regions according to an ecosystem matrix rather than subsidy totals alone, and the analysis in this paper suggests the dimensions of that matrix: energy reliability and power quality, not merely generation capacity; water availability and the institutional capacity to run a reclamation loop; supplier density across the hidden bill of materials; talent at every layer of the labor stack; the university and community-college pipeline; packaging and test availability within the region; transportation and air-cargo capacity; permitting predictability; construction capacity net of concurrent projects; expansion land; customer proximity; and policy durability across electoral cycles. Nvidia’s $1.5 billion prepayment to Amkor and Micron’s $500 million financing of GlobalWafers show that the most sophisticated firms are no longer passive consumers of regional ecosystems but active investors in them, using balance sheets to create the suppliers they need where they need them. [14,43] That behavior is itself a form of private-sector federalism, and states that understand it will design their incentives to co-invest with anchor customers rather than to substitute for them.
5.8 Implications for Governors and State Leaders
The paper offers neutral guidelines rather than political endorsements, and it frames them as questions that state policymakers could ask of their own strategies. First, which semiconductor capability does the state realistically possess, given its existing firms, universities and infrastructure, rather than which capability it would like to possess? Second, which missing suppliers in the hidden bill of materials prevent that capability from scaling, and can state dollars fill those gaps more productively than they can chase an anchor? Third, does the region possess sufficient electricity and water, measured by fab-grade reliability and reclamation capacity rather than by gross availability? Fourth, which universities or community colleges can develop the necessary workforce at each layer, and are they funded to do so on the fab’s timeline? Fifth, are incentives building an ecosystem or merely subsidizing a single project, and what evidence would distinguish the two? Sixth, what happens when incentive periods expire, and is the region’s cost position improving fast enough to retain the facility without them? Seventh, can infrastructure support second- and third-generation expansions, or was it sized for the announcement? Eighth, how does semiconductor policy connect with AI datacenter, energy and manufacturing policy, given that all four are competing for the same megawatts, acres and engineers?
5.9 Implications for Federal Strategy
Federal policymakers face a different question: how much geographic concentration is efficient, and how much diversification is strategically desirable? Too much concentration can reproduce geographic dependence inside the United States, replacing exposure to a single island with exposure to a single desert valley. Too much dispersion can destroy cluster economics, producing a dozen subscale regions none of which achieves the density that makes Hsinchu or Phoenix productive. Foundry Federalism operates precisely inside this tension, and the federal role is to manage it rather than to resolve it by fiat. Concretely, that suggests three federal functions that do not currently exist in developed form: a network-level view of the domestic semiconductor supply chain that maps interstate dependencies; incentive design that rewards complementary specialization rather than duplicative recruitment, for example by weighting materials and packaging awards toward regions that lack them; and policy durability, since the single most frequently cited corporate concern in 2026 is not the size of any incentive but the uncertainty created by shifting award terms, equity conditions and the Section 48D expiration. [45,50]
5.10 From Reshoring to Regionalization
Section 5 ends with the paper’s major historical progression, which places the current moment in a sequence that began before the CHIPS Act and extends beyond 2030. In Phase I, globalization, semiconductor manufacturing migrated toward specialized Asian clusters over four decades, driven by cost, by the fabless business model and by the extraordinary ecosystem density that Taiwan in particular achieved. In Phase II, reshoring, the United States incentivized domestic advanced manufacturing through the CHIPS Act, Section 48D, trade measures and federal partnerships, and the geographic unit was the nation. In Phase III, clustering, Arizona, Texas, Idaho, New York and other states assembled regional ecosystems around anchor facilities, and the geographic unit became the region. In Phase IV, specialization, which is beginning now, states develop complementary semiconductor roles rather than identical ones. And in Phase V, Foundry Federalism, the United States functions as a network of interdependent semiconductor regions operating under one national strategic framework, in which the geographic unit is simultaneously the nation, the region and the corridor between regions. The Reuters report of September 30, 2026 is a data point at the boundary between Phases III and IV.
Section 6: What Have We Learned? Eight Pillars of Foundry Federalism
The preceding sections have moved from a single Reuters report to a framework for understanding the geographic organization of an entire industry. This section distills that movement into eight pillars, each stated as a proposition, developed with the evidence assembled above, and closed with a lesson. The pillars are not a summary in the ordinary sense; they are the load-bearing claims on which the Foundry Federalism thesis rests, and a reader who disagrees with any of them should be able to locate precisely where the argument would fail.
Pillar 1: A Fab Is Not an Ecosystem
The semiconductor resurgence cannot be measured by factory announcements alone. The sustainable unit of industrial capacity is the cluster: fabs plus packaging, materials, suppliers, equipment service, utilities, talent, research and logistics. Arizona’s importance comes not simply from the dollar value of TSMC’s investment but from the surrounding capabilities now forming around it, the seventy-plus supplier expansions, the Amkor campus, the Applied Materials research center, the Maricopa technician pipeline and the water reclamation plant. [3,13,61,28] Shih and Lin’s finding that Phoenix-made wafers were still traveling to Asia for packaging in 2026 is the sharpest reminder that a fab without an ecosystem is an incomplete asset. [16] The lesson is that America does not secure semiconductor leadership by building isolated fabs; it must make those fabs reproducible, expandable and supportable, and the measure of success is the second, third and fourth facility rather than the first.
Pillar 2: Geography Has Become Part of Semiconductor Strategy
Semiconductor geography was once discussed almost entirely at the international scale: Taiwan versus China, the United States versus Asia, Japan versus Korea. That scale remains decisive, and Miller’s framing of access today and trust tomorrow captures why. [20] But an additional geography is emerging inside the United States, in which Arizona, Texas, Idaho, New York and other states offer different combinations of skills, resources and industrial history, and in which the choice among them is a strategic decision for the companies that make it and for the nation whose resilience depends on it. The lesson is that the semiconductor map now exists at three scales simultaneously, global, national and interstate, and that policy which attends to only the first two will misunderstand where the physical infrastructure of AI is actually being decided.
Pillar 3: States Compete Most Effectively Through Specialization
Trying to duplicate TSMC Arizona fifty times would be neither economically realistic nor strategically necessary, and the agglomeration literature explains why: productivity rises with cluster size, so a dozen subscale copies would each be less productive than one large original. [24] A more plausible semiconductor federation allows regions to specialize. One state excels at frontier fabrication, another dominates memory, another leads packaging, another becomes indispensable in materials or photonics, another hosts the AI datacenters consuming the resulting processors. Texas’s Semiconductor Innovation Fund, deliberately spread across materials, components, photonics and workforce rather than concentrated on a single fab, is an early example of a state choosing its specialization consciously. [31,32,33,34] The lesson is that resilience can come from complementary capabilities rather than identical ones, and that the states most likely to win are those that understand which part of the system they can realistically own.
Pillar 4: Federal and State Industrial Policy Are Becoming Interdependent
Washington controls powerful tools: trade policy, export controls, federal research, taxation, national-security policy, defense procurement and semiconductor incentives that now include equity stakes as well as grants. [45,46,47] States control many of the conditions that determine whether those investments actually become functioning industrial ecosystems: sites, water, utilities, permitting, universities, workforce programs and local incentives. Neither level is sufficient independently. The CHIPS Act without Phoenix’s water institutions would have produced a fab with a supply problem; Phoenix’s water institutions without the CHIPS Act would have produced a well-planned city with no fab. The Erten-Stiglitz-Verhoogen evidence that federal money reinforced existing regional clusters rather than creating new ones is the empirical signature of this interdependence. [22] The lesson is that American semiconductor strategy is evolving into a multilayer system in which federal policy supplies strategic direction and capital while states perform much of the geographic execution, and that interaction is the essence of Foundry Federalism.
Pillar 5: Advanced Packaging Is the Hinge of the Geography
This pillar did not appear in the original outline, and it earns its place because the evidence of 2026 makes it unavoidable. Advanced packaging has become the binding constraint on AI accelerator supply, the step at which fabricated silicon becomes a usable processor, and the activity whose domestic absence most undermines the value of domestic fabrication. [12,16] The TSMC-Amkor ten-year agreement, Nvidia’s $1.5 billion prepayment, TSMC’s own planned Arizona packaging facilities and SK Hynix’s Indiana HBM packaging investment together represent the first serious attempt to close that gap, and they are concentrating in a small number of regions because packaging must sit close to the fabs and memory it integrates. [11,14,53] The lesson is that any state aspiring to a leading-edge role must plan for packaging and test from the beginning, and that the geography of packaging will do more to determine which clusters are complete than the geography of fabrication alone.
Pillar 6: Power, Water and People Are the Real Currency of Competition
Subsidies open the conversation, but the inputs that decide it are the ones that cannot be appropriated in a legislative session. Fab-grade electricity, with its demands for reliability, power quality and multi-decade price visibility, is scarcer than gross generation statistics suggest, and the ERCOT interconnection queue of more than 438 gigawatts shows how quickly datacenter demand can crowd the grid that fabs also need. [56,55] Water is governed less by rainfall than by the institutional capacity to run a reclamation loop, as the progression from 65 percent to 90 percent reuse in Phoenix demonstrates. [27,28] People are the slowest input of all: a technician can be trained in months, an engineer in years, a process culture in a decade, and the national shortfall estimates ranging from 67,000 to as many as 157,000 workers by 2030 mean that every new cluster is recruiting from a pool that every other cluster is also recruiting from. [51,53] The lesson is that states should measure their competitiveness in megawatts of firm power, gallons of reclaimed water and graduates per year, because those are the units in which site-selection teams now keep score.
Pillar 7: Policy Durability Is Itself an Input
A fab is a thirty-year asset, and the companies that build them discount every incentive by their confidence that its terms will hold. The 2025 and 2026 renegotiation of CHIPS awards, the introduction of equity conditions, the uncertainty over which recipients would be asked for stakes, and the scheduled expiration of Section 48D for construction beginning after 2026 have all raised the risk premium that corporate planners attach to federal commitments, even as the headline generosity of the credit rose from 25 to 35 percent. [45,48,49,50] New York’s twenty-year GREEN CHIPS term sheet and Phoenix’s hundred-year water plan are, by contrast, examples of durability designed into state and municipal policy from the start. [40,28] The lesson is that in a federation, the level of government that can credibly commit for the longest horizon will exert disproportionate influence over where long-lived assets are built, and that states have an underappreciated advantage in this respect precisely because their commitments are smaller, more specific and less exposed to national political reversal.
Pillar 8: Semiconductor Geography Will Shape the Five-Layer AI Economy
The location of semiconductor manufacturing is not a separate industrial-policy question from artificial intelligence; it is one of the physical foundations on which the AI economy stands. Layer 1 supplies the power and physical infrastructure. Layer 2 manufactures the processors, memory and networking silicon. Layer 3 turns those components into datacenter-scale computing systems. Layer 4 converts compute into models. Layer 5 converts models into applications and autonomous agents. Constraints or advantages created at one layer propagate through all the others: a packaging bottleneck in Arizona becomes an accelerator shortage for a hyperscaler, which becomes a training-run delay for a model developer, which becomes a product delay for an application company. The record results reported by Nvidia, TSMC and Micron in mid-2026 are the upper layers’ demand arriving at Layer 2, and the state-by-state scramble for fabs, packaging, power, water and talent is Layer 2’s demand arriving at Layer 1. [59,9,41] The lesson is that the geography of fabs eventually becomes the geography of computational capability, and therefore part of the geography of AI economic power.
Conclusion: From One Arizona to an American Semiconductor Federation
The story began with a possibility. On September 30, 2026, Reuters reported that TSMC was evaluating an investment in Texas beyond the extraordinary $265 billion expansion already planned for Arizona. The Texas possibility remained unfinalized when this paper was written, and it may develop, change or never proceed in the form currently being considered. [1] But its analytical importance does not depend on the final investment decision. What matters is that such a possibility can now be seriously contemplated by the world’s most important foundry, by the governors of two of the nation’s largest states, and by the investors, suppliers and customers who read the same report and immediately began asking what it meant for them.
A decade ago, America’s semiconductor question centered on whether advanced manufacturing could be restored domestically at all. The policy response concentrated on national capacity: bring fabrication back, reduce supply-chain dependence, support leading-edge manufacturing and rebuild capabilities that had migrated overseas over forty years of globalization. Arizona became the most visible physical manifestation of that strategy, and by the measures that matter, yield, customer commitment, supplier migration and successive expansion, it has succeeded beyond what most observers expected in 2020. [3,10,4]
Yet Arizona’s success creates the conditions for the next question, and the next question is geographic. If semiconductor reshoring works, where does the next factory go? And after that: where does the packaging go, where is the memory manufactured, who produces the wafers, who supplies the chemicals, where does optical networking concentrate, which universities train the engineers, which utilities can provide sufficiently reliable electricity, which regions possess water systems capable of supporting additional industrial capacity, and where are the customers building the next generation of AI datacenters? These questions cannot be answered entirely from Washington. They are questions about places, and in the American political economy, questions about places inevitably bring states into the equation.
Arizona can expand its leading-edge semiconductor cluster toward the twelve facilities now planned. Texas can leverage its existing semiconductor base, its supplier network, its design community, its generation capacity and its state industrial programs, whether or not TSMC ultimately joins Samsung and Texas Instruments there. Idaho can combine memory research and manufacturing. New York can pursue memory production at a scale and over a horizon that few regions in the world can match. Other states can develop roles in materials, equipment, packaging, defense microelectronics, power electronics, photonics or research. The strongest future semiconductor system may therefore not resemble one gigantic manufacturing location. It may resemble a federation.
That federation would still compete globally as the United States. Federal policy would continue shaping national-security priorities, research, taxation, trade, export controls and investment incentives, and the evidence of the CHIPS Act’s first years shows that federal money can measurably accelerate regional clusters. [22] But beneath that national strategy, states would compete for investment, specialize in different technologies, build universities and workforce systems, expand utilities and infrastructure, recruit the suppliers that constitute the hidden bill of materials, and develop complementary industrial capabilities. Competition would coexist with interdependence. Texas would not need Arizona to lose for Texas to succeed. Arizona would not need New York to manufacture the same products that Arizona manufactures. Idaho would not need to recreate Phoenix. The underlying economic objective would become more sophisticated than any single-state victory: build enough specialization that each region develops competitive depth, while maintaining enough integration that the regions collectively form a resilient national semiconductor system.
This is why the title Foundry Federalism fits the paper. Foundry anchors the argument in the physical reality of semiconductor manufacturing, in the 4.75 million gallons of water a day that a single fab consumes, in the hundreds of megawatts it draws, in the twelve thousand construction workers it requires and the thousands of engineers it must recruit. [27,8] Artificial intelligence may appear increasingly digital and autonomous at Layers 4 and 5, but every model ultimately rests on physical processors fabricated with extraordinarily demanding industrial processes in specific places governed by specific institutions. Federalism identifies the mechanism through which America’s semiconductor geography may now evolve. The federal government establishes national strategic objectives and supplies catalytic capital, while states compete and cooperate to determine where the actual fabs, packaging plants, supplier facilities, research laboratories, universities, power infrastructure and workforce ecosystems are built.
The term therefore describes something larger than competition between Arizona and Texas. It describes a possible transformation of American industrial organization, from a model in which industrial policy was a federal instrument applied to a national economy toward a model in which it is a layered system of national objectives and regional execution. The first chapter of the semiconductor revival asked whether America could manufacture advanced chips again. The second asked whether America could manufacture them at scale. The emerging third chapter asks how that capacity should be distributed across the country, and that question has no purely federal answer.
By 2030, the answer may no longer be found on a single semiconductor campus or inside one federal program. It may instead be visible on a map of the United States: advanced logic in one region, memory in another, packaging in another, materials somewhere else, research dispersed among several technology centers, and massive AI datacenters converting those chips into intelligence across yet another geography, with wafers, dies, modules and engineers moving between them along corridors that have themselves become national infrastructure. That would mark the transition from semiconductor reshoring to semiconductor regionalization, and eventually from isolated industrial projects to an integrated national production network.
That is Foundry Federalism.
Footnotes and Endnotes:
[1] Wen-Yee Lee and Fanny Potkin (Reuters). “TSMC evaluates potential Texas investment, sources say.” Reuters via Investing.com, September 30, 2026. https://www.investing.com/news/stock-market-news/tsmc-evaluates-potential-texas-investment-sources-say-4924436
[2] Investing.com / Yahoo Finance. “TSMC weighs investment in Texas to expand U.S. chip production, Reuters reports.” Yahoo Finance, September 30, 2026. https://finance.yahoo.com/technology/articles/tsmc-weighs-investment-texas-expand-021055343.html
[3] City of Phoenix, Community and Economic Development. “TSMC Announces Additional $100 Billion Investment In Arizona.” City of Phoenix Newsroom, July 16, 2026. https://www.phoenix.gov/newsroom/ced-news/tsmc-announces-additional–100-billion-investment-in-arizona.html
[4] Agence France-Presse / Yahoo Finance. “Taiwan chipmaker TSMC to invest another US$100 bn in Arizona fabs.” Yahoo Finance, July 16, 2026. https://finance.yahoo.com/technology/ai/articles/taiwan-chipmaker-tsmc-invest-another-065311994.html
[5] SiliconANGLE. “TSMC boosts Arizona fab investment by $100B after strong second quarter.” SiliconANGLE, July 16, 2026. https://siliconangle.com/2026/07/16/tsmc-boosts-arizona-fab-investment-100b-strong-second-quarter/
[6] Taiwan Semiconductor Manufacturing Company. “TSMC Announces Updates for TSMC Arizona (Form 6-K).” U.S. Securities and Exchange Commission, December 6, 2022. https://www.sec.gov/Archives/edgar/data/1046179/000156459022039051/tsm-6k_20221206.htm
[7] AZ Big Media. “TSMC will invest an additional $100 billion into Arizona operations.” AZ Big Media, August 27, 2026. https://azbigmedia.com/business/tsmc-will-invest-an-additional-100-billion-into-arizona-operations/
[8] Alaina Kwan (Arizona’s Family / AZFamily). “TSMC adds $100B to Arizona plans as leaders call it a chipmaking ‘gold rush’.” AZFamily, July 27, 2026. https://www.azfamily.com/2026/07/27/tsmc-adds-100b-arizona-plans-leaders-call-it-chipmaking-gold-rush/
[9] Taiwan Semiconductor Manufacturing Company. “TSMC August 2026 Revenue Report (Form 6-K).” U.S. Securities and Exchange Commission, September 10, 2026. https://www.sec.gov/Archives/edgar/data/0001046179/000104617926000658/tsm-revenue20260910.htm
[10] Arizona Technology Council. “Apple to buy more than 100 million chips from TSMC’s Arizona fab by year’s end.” Arizona Technology Council, February 2026. https://www.aztechcouncil.org/apple-chips-tsmc-fab/
[11] Amkor Technology, Inc. and TSMC. “TSMC and Amkor Technology Announce Long Term Partnership to Accelerate Advanced Packaging in the United States.” Business Wire / Amkor Investor Relations, June 16, 2026. https://ir.amkor.com/news-releases/news-release-details/tsmc-and-amkor-technology-announce-long-term-partnership
[12] TrendForce. “TSMC, Amkor Sign 10-Year Arizona Advanced Packaging Pact to Complete the U.S. Chip Supply Chain.” TrendForce News, June 17, 2026. https://www.trendforce.com/news/2026/06/17/news-tsmc-amkor-forge-10-year-arizona-advanced-packaging-partnership-to-complete-the-u-s-chip-supply-chain/
[13] Amkor Technology, Inc. “Amkor Technology Breaks Ground on New Semiconductor Advanced Packaging and Test Campus in Arizona; Expands Investment to $7 Billion.” Amkor Investor Relations, October 2025. https://ir.amkor.com/news-releases/news-release-details/amkor-technology-breaks-ground-new-semiconductor-advanced
[14] Amkor Technology, Inc. “Amkor Technology Announces Strategic Partnership with NVIDIA to Expand Advanced Packaging and Test for Next-Generation AI Infrastructure.” Business Wire / Amkor Investor Relations, July 23, 2026. https://ir.amkor.com/news-releases/news-release-details/amkor-technology-announces-strategic-partnership-nvidia-expand
[15] Reuters / Yahoo Finance. “Nvidia, Amkor strike $1.5 billion chip packaging deal.” Yahoo Finance, July 23, 2026. https://finance.yahoo.com/technology/ai/articles/nvidia-amkor-strike-1-5-214417511.html
[16] Willy C. Shih and PJ Lin (Harvard Business School). “Where the U.S.’s Chip Strategy Is Still Falling Short.” Harvard Business Review, April 28, 2026. https://hbr.org/2026/04/where-the-u-s-s-chip-strategy-is-still-falling-short
[17] Harvard Magazine. “Five Questions with Willy Shih.” Harvard Magazine, June 2026. https://www.harvardmagazine.com/five-questions/harvard-five-questions-willy-shih-semiconductors
[18] Fortune (interview with Willy C. Shih). “Wall Street thinks memory is AI’s golden ticket. Harvard’s chip expert warns: ‘Curves that just go to the sky with no end…never continue forever’.” Fortune, May 11, 2026. https://fortune.com/2026/05/11/ai-memory-chips-semiconductor-stock-boom-price-hikes-dram-shortage-hbm/
[19] Central News Agency, featuring Chris Miller (Fletcher School, Tufts University). “Culture, costs unlikely to be decisive for TSMC in Arizona, academic says.” Fletcher School, Tufts University (Russia and Eurasia Program site), 2024. https://sites.tufts.edu/fletcherrussia/culture-costs-unlikely-to-be-decisive-for-tsmc-in-arizona-academic-says/
[20] Carnegie Mellon Institute for Strategy & Technology. “Chips and Chokepoints: Chris Miller on the Geopolitics of the AI Supply Chain.” Carnegie Mellon University, March 2026. https://www.cmu.edu/cmist/news-archive/news/2026/march/chips-and-chokepoints-chris-miller-on-the-geopolitics-of-the-ai-supply-chain.html
[21] Center for Strategic and International Studies (quoting John VerWey, Pacific Northwest National Laboratory). “Can Semiconductor Reshoring Prime a U.S. Manufacturing Renaissance?.” CSIS, 2026. https://www.csis.org/analysis/can-semiconductor-reshoring-prime-us-manufacturing-renaissance
[22] Bilge Erten, Joseph E. Stiglitz and Eric Verhoogen. “Employment Impacts of the CHIPS Act.” Brookings Papers on Economic Activity, Fall 2025. https://www.brookings.edu/articles/employment-impacts-of-the-chips-act
[23] Columbia University School of International and Public Affairs. “New SIPA Research Finds that the CHIPS Act Created Jobs: A Q&A with Eric Verhoogen.” Columbia SIPA, February 2026. https://www.sipa.columbia.edu/news/new-sipa-research-finds-chips-act-created-jobs-qa-eric-verhoogen
[24] Enrico Moretti (University of California, Berkeley). “The Effect of High-Tech Clusters on the Productivity of Top Inventors.” American Economic Review, 2021 (working paper version). https://eml.berkeley.edu/~moretti/clusters.pdf
[25] Mark Muro and Bruce Katz (Brookings Institution). “The New ‘Cluster Moment’: How Regional Innovation Clusters Can Foster the Next Economy.” Brookings Metropolitan Policy Program, September 2010. https://www.brookings.edu/wp-content/uploads/2016/06/0921_clusters_muro_katz.pdf
[26] Arizona’s Family (AZFamily), quoting Sarah Porter (Kyl Center for Water Policy, Arizona State University). “TSMC Arizona expansion could make it Phoenix’s largest water customer.” AZFamily, July 23, 2026. https://www.azfamily.com/2026/07/23/tsmc-arizona-expansion-could-make-it-phoenixs-largest-water-customer/
[27] Arizona’s Family (AZFamily). “Arizona’s chip footprint grows, but TSMC’s water demand sparks questions.” AZFamily, August 24, 2026. https://www.azfamily.com/2026/08/24/arizonas-chip-footprint-grows-tsmcs-water-demand-sparks-questions/
[28] Data Center Dynamics. “TSMC breaks ground on water reclamation project in Phoenix, Arizona.” DCD, August 2026. https://www.datacenterdynamics.com/en/news/tsmc-breaks-ground-on-water-reclamation-project-in-phoenix-arizona/
[29] American Bar Association, Section of Environment, Energy, and Resources (quoting Sarah Porter). “A Tale of Two Shortages: Reconciling Demand for Water and Microchips in Arizona.” ABA Water Resources Committee Newsletter, 2023. https://www.americanbar.org/groups/environment_energy_resources/resources/newsletters/water/reconciling-demand-water-microchips-arizona/
[30] Grist. “Arizona’s water is drying up. That’s not stopping the data center rush.” Grist, March 2026. https://grist.org/technology/arizona-water-data-centers-semiconducters/
[31] Office of the Texas Governor. “Governor Abbott Announces Texas Semiconductor Innovation Fund Grant To Soulbrain.” gov.texas.gov, June 4, 2026. https://gov.texas.gov/news/post/governor-abbott-announces-texas-semiconductor-innovation-fund-grant-to-soulbrain
[32] Semiconductor Today. “Texas Semiconductor Innovation Fund grants Coherent $14m to accelerate scaled production of InP wafers.” Semiconductor Today, February 9, 2026. https://www.semiconductor-today.com/news_items/2026/feb/coherent-090226.shtml
[33] Applied Optoelectronics, Inc. “AOI Awarded $20.9M Texas Semiconductor Innovation Fund Grant to Advance Manufacturing in Sugar Land.” AOI Newsroom, April 29, 2026. https://newsroom.ao-inc.com/news-releases/aoi-awarded-20-9m-texas-semiconductor-innovation-fund-grant-to-advance-manufacturing-in-sugar-land/
[34] CBS Austin. “Georgetown semiconductor supplier to get $3.9 million state grant for expansion.” CBS Austin, June 19, 2026. https://cbsaustin.com/news/local/georgetown-semiconductor-supplier-to-get-39-million-state-grant-for-expansion
[35] National Institute of Standards and Technology, CHIPS for America. “Samsung Electronics (Texas): Project Summary.” NIST CHIPS Program Office, updated September 2026. https://www.nist.gov/chips/samsung-electronics-texas-taylor
[36] Seoul Economic Daily. “Samsung’s Texas Fab Starts Mass Production on Tesla Orders.” Seoul Economic Daily (English), September 15, 2026. https://en.sedaily.com/finance/2026/09/15/samsung-starts-chip-output-at-texas-plant-on-tesla-orders
[37] Electrek. “Samsung’s Texas fab enters production for Tesla’s AI5 chip on 2nm.” Electrek, July 13, 2026. https://electrek.co/2026/07/13/samsung-taylor-fab-tesla-ai5-chip-2nm/
[38] Texas Instruments Incorporated. “Texas Instruments plans to invest more than $60 billion to manufacture billions of foundational semiconductors in the U.S..” PR Newswire, June 18, 2025. https://www.prnewswire.com/news-releases/texas-instruments-plans-to-invest-more-than-60-billion-to-manufacture-billions-of-foundational-semiconductors-in-the-us-302485156.html
[39] Tom’s Hardware. “New Texas Instruments fab will pump out tens of millions of chips per day; first 300mm fab starts production after $60 billion investment.” Tom’s Hardware, December 18, 2025. https://www.tomshardware.com/tech-industry/semiconductors/new-texas-instruments-fab-will-pump-out-tens-of-millions-of-chips-per-day-first-300mm-fab-starts-production-after-usd60-billion-investment
[40] Micron Technology, Inc. “Quarterly Report on Form 10-Q for the quarter ended May 28, 2026.” U.S. Securities and Exchange Commission, June 2026. https://www.sec.gov/Archives/edgar/data/0000723125/000072312526000015/mu-20260528.htm
[41] Micron Technology, Inc. “Micron Technology, Inc. Reports Record Results for the Third Quarter of Fiscal 2026 (Form 8-K, Exhibit 99.1).” U.S. Securities and Exchange Commission, June 24, 2026. https://www.sec.gov/Archives/edgar/data/0000723125/000072312526000013/a2026q3ex991-pressrelease.htm
[42] Micron Technology, Inc. (GlobeNewswire). “Micron Accelerates U.S. Investments, Pours First Concrete at New York Fab.” GlobeNewswire via The Manila Times, July 9, 2026. https://www.manilatimes.net/2026/07/09/tmt-newswire/globenewswire/micron-accelerates-us-investments-pours-first-concrete-at-new-york-fab/2381646
[43] Tom’s Hardware. “Micron lifts U.S. spending to $250 billion; company takes $500 million position in America’s only 300 mm wafer plant.” Tom’s Hardware, July 10, 2026. https://www.tomshardware.com/tech-industry/semiconductors/micron-takes-a-500-million-position-in-americas-only-300mm-wafer-plant
[44] Naga Chandrasekaran (Intel Corporation). “Ohio One Construction Timeline Update.” Intel Newsroom, February 28, 2025. https://www.intel.com/content/www/us/en/newsroom/news/corporate/ohio-one-construction-timeline-update.html
[45] Congressional Research Service. “Semiconductor Fabrication Facilities Funded by the CHIPS Act: Project Status and Considerations for Congress (R49031).” U.S. Congress, July 2026. https://www.congress.gov/crs-product/R49031
[46] CNBC. “Lutnick says Intel has to give government equity in return for CHIPS Act funds.” CNBC, August 19, 2025. https://www.cnbc.com/2025/08/19/lutnick-intel-stock-chips-trump.html
[47] Fox Business. “Commerce Dept to demand equity from 7 tech companies as a condition to tap into millions in federal funding.” Fox Business, July 2026. https://www.foxbusiness.com/politics/commerce-dept-demand-equity-from-tech-companies-condition-tap-millions-federal-funding
[48] Engadget. “Trump administration official says some CHIPS Act companies won’t need to give up equity.” Engadget, August 2025. https://www.engadget.com/big-tech/trump-administration-official-says-some-chips-act-companies-wont-need-to-give-up-equity-130041299.html
[49] Legal Information Institute, Cornell Law School. “26 U.S. Code § 48D: Advanced manufacturing investment credit.” Cornell LII, current through Public Law 119-21. https://www.law.cornell.edu/uscode/text/26/48D
[50] Semiconductor Industry Association. “Extend the 35% Advanced Manufacturing Investment Credit (policy two-pager).” SIA, May 2026. https://www.semiconductors.org/wp-content/uploads/2026/05/SIA-48D-2-Pager.pdf
[51] Semiconductor Industry Association and Oxford Economics. “America Faces Significant Shortage of Tech Workers in Semiconductor Industry and Throughout U.S. Economy.” SIA, July 25, 2023. https://www.semiconductors.org/america-faces-significant-shortage-of-tech-workers-in-semiconductor-industry-and-throughout-u-s-economy/
[52] Harvard Business School, Institute for Business in Global Society. “Shortage of semiconductor workers could impact climate firms.” HBS BiGS, 2026. https://www.hbs.edu/bigs/shortage-of-semiconductor-workers-could-impact-climate-firms
[53] Tech Times. “US Chip Fabs Face 157,000-Worker Shortfall as AI Labs Drain Same Talent Pool.” Tech Times, September 20, 2026. https://www.techtimes.com/articles/327782/20260920/us-chip-fabs-face-157000-worker-shortfall-ai-labs-drain-same-talent-pool.htm
[54] WebProNews (reporting CNBC interview with Jon Taylor, Samsung Austin Semiconductor). “America’s Chip Factories Rise, But the Workers to Run Them Are Nowhere in Sight.” WebProNews, September 2026. https://www.webpronews.com/americas-chip-factories-rise-but-the-workers-to-run-them-are-nowhere-in-sight
[55] Robert Walton (Utility Dive). “Facing an estimated 474 GW of interconnection requests, Texas hits pause on data centers.” Utility Dive, August 5, 2026. https://www.utilitydive.com/news/texas-hits-pause-data-center-interconnections/827046/
[56] Electric Reliability Council of Texas. “ERCOT Trending Topics: ERCOT’s New Batch Connection Process for Large Electricity Users.” ERCOT, June 18, 2026. https://www.ercot.com/files/docs/2026/06/18/ERCOT-Trending-Topic-New-Batch-Connection-Process-for-Large-Electricity-Users.pdf
[57] Latitude Media. “ERCOT’s large load queue has nearly quadrupled in a single year.” Latitude Media, February 2026. https://www.latitudemedia.com/news/ercots-large-load-queue-has-nearly-quadrupled-in-a-single-year/
[58] International Energy Agency. “Energy and AI: Executive Summary.” IEA World Energy Outlook Special Report, April 2025. https://www.iea.org/reports/energy-and-ai/executive-summary
[59] NVIDIA Corporation. “NVIDIA Announces Financial Results for Second Quarter Fiscal 2027 (Form 8-K).” U.S. Securities and Exchange Commission, August 26, 2026. https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000073/q2fy27pr.htm
[60] Joseph Parilla, Mark Muro and Xavier de Souza Briggs (Brookings Institution). “In Central New York, a test of the CHIPS and Science Act’s promise for economic revitalization.” Brookings, 2023. https://www.brookings.edu/articles/in-central-new-york-a-test-of-the-chips-and-science-acts-promise-for-economic-revitalization/
[61] Semiconductor Today. “Arizona State University, Rensselaer Polytechnic Institute and Stanford University to join Applied Materials’ EPIC Center in Silicon Valley.” Semiconductor Today, May 12, 2026. https://www.semiconductor-today.com/news_items/2026/may/appliedmaterials-asu-rpi-stanford-120526.shtml
[62] Stanford Emerging Technology Review. “Semiconductors (2026 edition).” Stanford University, 2026. https://setr.stanford.edu/technology/semiconductors/2026
[63] Investing.com / Yahoo Finance. “Nvidia Signed a $1.5 Billion Deal With This Chip Stock. Then the Stock Crashed 25% (reporting Amkor Q2 2026 results).” Yahoo Finance, August 2026. https://finance.yahoo.com/markets/stocks/articles/nvidia-signed-1-5-billion-181658818.html
[64] Texas Instruments Incorporated via TechPowerUp. “Texas Instruments Begins Production at Its 300 mm Semiconductor Fab in Sherman, Texas.” TechPowerUp, December 2025. https://www.techpowerup.com/344190/texas-instruments-begins-production-at-its-300-mm-semiconductor-fab-in-sherman-texas
