Introduction: The AI Election Is Moving Underground
On August 18, 2026, less than three months before Americans vote in the November midterm elections, Pennsylvania offered a remarkable illustration of how quickly the politics of artificial intelligence infrastructure has changed.
Governor Josh Shapiro signed Executive Order 2026-05, imposing what his office called the nation’s strictest guardrails on AI datacenter development in a state that, only a year earlier, had been aggressively courting some of the largest technology investments in the country. The order directs the Pennsylvania Department of Environmental Protection to review permit applications only when developers have made a legally binding commitment to meet the Governor’s Responsible Infrastructure Development — GRID — Requirements and have first received local approval. It removes all AI datacenter proposals from the state’s Fast Track permitting program, prohibits the use of nondisclosure agreements between state agencies and datacenter developers, and demands strict standards on energy affordability, environmental protection, workforce development, transparency, and community engagement [1]. Standing in Harrisburg, Shapiro was blunt about the mechanism at the heart of the order:
“If the local community doesn’t approve a project, the state won’t approve it either.”
— Governor Josh Shapiro, Commonwealth of Pennsylvania [2]
The scale of what the order governs is itself revealing. Pennsylvania’s environmental regulators reported that they had become aware of an unprecedented number of datacenter proposals — more than 100 projects appearing in publicly sourced databases, 58 projects engaged with the Department at some level of formality, fifteen holding at least one permit application, and only five having received all permits necessary for a first phase of development [1]. The gap between speculative announcement and permitted reality had become, in the administration’s telling, a problem in its own right: a landscape of paper gigawatts competing for land options, interconnection positions, water allocations, and political attention.
The executive order goes considerably further than procedure. According to reporting by the Philadelphia Inquirer, it mandates that datacenter projects bring their own electricity generation and pay all costs associated with their increased energy usage; requires community-benefit agreements that include local hiring, training, and developer investments in schools or infrastructure; instructs the Department of Environmental Protection to publish a publicly accessible map of permitting information for every proposed datacenter project; and explicitly grounds the entire framework in Pennsylvania’s distinctive constitutional guarantee of clean air, pure water, and environmental preservation [3]. The Department’s secretary, Jessica Shirley, described the bargain being offered to developers in the plainest possible terms — permits will be reviewed, but not issued, until all local land-use approvals are in hand, and only for projects that commit to
“the highest standards for environmental protection, including strict water conservation requirements.”
— Jessica Shirley, Secretary, Pennsylvania Department of Environmental Protection [4]
For a governor who had announced barely a year earlier that Pennsylvania was “all in” on artificial intelligence — who had stood beside Amazon executives to celebrate a $20 billion datacenter commitment — the political reversal was unmistakable, and his opponents said so loudly. The Republican-controlled State Senate had declined to act on GRID legislation that passed the House, and Shapiro framed executive action as the only remaining instrument:
“The absence of legislative approval has left me with no other option.”
— Governor Josh Shapiro, remarks in Harrisburg, August 18, 2026 [5]
The political argument in Pennsylvania is therefore no longer about whether the Commonwealth wants artificial intelligence investment. It has become an argument about what kind of AI industrialization Pennsylvania is willing to accept, who will pay for it, where its electricity will come from, and who holds the veto.
The disagreement had already entered electoral politics well before the signing ceremony. Reporting by the Associated Press on the same day described datacenter development as an increasingly combustible issue in several 2026 gubernatorial contests, most visibly Pennsylvania and Texas. In Pennsylvania, Shapiro’s Republican challenger, State Treasurer Stacy Garrity, has attacked the governor’s record from the opposite direction — arguing that his early enthusiasm created the very backlash he now claims to be managing, saying Shapiro
“lit the fuse on the chaos we are seeing in community after community.”
— Stacy Garrity, Pennsylvania State Treasurer and Republican nominee for Governor [6]
In Texas, the same Associated Press reporting described Democratic challenger Gina Hinojosa releasing television advertising in rural markets accusing Governor Greg Abbott of selling out rural communities to datacenter companies, seeking to exploit discontent in Republican strongholds over datacenters’ perceived threat to rural life, ranchland, and dwindling water supplies — even as Abbott himself, once a recruiter of the industry, has pivoted toward a regime of audits, disclosure requirements, and cost-responsibility rules examined in detail in Section 4 [6]. These are descriptions of publicly reported positions rather than an assessment of either party or candidate; what matters for this paper is the structural fact both parties now acknowledge: gigawatt computing has become large enough to require a political settlement.
What makes the Pennsylvania story particularly significant, however, is what happened seven days before Shapiro’s August 18 order.
On August 11, 2026, Alpha Compute Corporation — an AI Confidential Compute and GPU-as-a-service company listed on Nasdaq — announced that it had signed a binding term sheet giving its subsidiary an exclusive option to acquire not merely a parcel of northern Pennsylvania land on which to construct an AI datacenter, but the mineral, surface, and pore-space assets beneath and around it, for a base purchase price of $55 million. The deal, first reported by Reuters, includes undeveloped Marcellus shale gas rights covering approximately 1,800 oil and gas mineral acres. The initial campus is envisioned at approximately 200 megawatts of behind-the-meter, gas-fired power and datacenter capacity, with potential expansion to roughly one gigawatt [7]. A third-party evaluation cited by the company estimated that gas produced from the property’s Marcellus formation could support 200 megawatts of continuous generation for ten years at an all-in delivered electricity cost of approximately $0.0585 per kilowatt-hour — a figure the company presented as sitting below prevailing commercial and industrial rates in the PJM market — while the county’s chief assessor reportedly estimated that a project at full scale could add roughly $2.08 billion in taxable assessed value and $33.4 million in combined annual tax impact across the county, municipalities, and school district [8].
Rather than waiting years for sufficient electricity to arrive through the conventional utility grid, the developer is attempting something conceptually different: placing computation directly on top of the energy resource and obtaining ownership-level control over the fuel required to produce its electricity. The transaction remains subject to due diligence, financing, permitting, definitive agreements, and — after August 18 — the full weight of Pennsylvania’s new GRID Requirements. It may close; it may not. But as a signal of where the industry’s frontier of control now sits, it deserves far more attention than its relatively modest dollar value might suggest.
For most of the cloud-computing era, technology companies regarded electricity as an input purchased after a datacenter location had been selected. A hyperscaler chose land, fiber access, tax treatment, and proximity to customers; utilities supplied the electricity. Energy infrastructure existed outside the conceptual boundaries of the computing company.
Artificial intelligence is beginning to invert that relationship.
Increasingly, the question is no longer:
Where should we put the datacenter, and how do we obtain electricity for it?
It is becoming:
Where can we control electricity — and therefore, where should we put the datacenter?
This reversal is visible far beyond one Pennsylvania project, and at every scale of the industry. In May 2026, UGI Energy Services and Prime Data Centers announced a strategic partnership to develop major natural-gas infrastructure supporting a northern Pennsylvania AI and high-performance-computing campus, including dedicated pipeline facilities and an on-site gas-fired generation complex; UGI said expected gas demand associated with the project could exceed 100,000 dekatherms per day within three to five years, with planned midstream investment in excess of $100 million [9].
In West Texas, Chevron and Microsoft have moved even further toward the fusion of energy and computing. Their Project Kilby is expected to deliver approximately 2.67 gigawatts of dedicated capacity under a 20-year power purchase agreement for a Microsoft-operated datacenter, with the majority of generation coming from large GE Vernova gas turbines fed by Permian Basin natural gas. The project effectively places an oil supermajor, a merchant power plant, an activist-investor energy venture, and a hyperscale computing operation inside one coordinated industrial architecture [10].
In Louisiana, Meta’s expansion of its Hyperion campus in Richland Parish toward five gigawatts of compute capacity — at a total investment the company now puts above $50 billion — is welded to an energy program that includes seven new combined-cycle natural gas plants, grid-scale battery storage, nuclear uprates, renewable procurement of up to 2.5 gigawatts, and hundreds of miles of new high-voltage transmission [11].
And beneath the surface does not necessarily mean fossil fuel. Google and Fervo Energy signed a framework agreement in March 2026 that could support as much as three gigawatts of enhanced geothermal power for Google’s datacenters through 2033 [12]. Fervo’s technology extracts heat from deep underground formations, turning another form of geology into an input for the digital economy. In June 2026, Fervo, NVIDIA, and the Department of Energy’s Pacific Northwest National Laboratory announced they would jointly develop EGS-Twin, an AI-driven digital-twin platform for modeling underground geothermal reservoirs — creating a remarkable feedback loop in which artificial intelligence helps extract subsurface energy that can eventually supply more artificial intelligence [13].
Even the capital markets have begun pricing the shift. On the very day Shapiro signed his order — August 18, 2026 — Reuters reported that private-equity firm KKR had made an approximately $9 billion takeover offer for UGI Corporation, the Pennsylvania-based natural gas and electricity distributor already positioning its pipelines, storage, and midstream capabilities around growing datacenter demand, at a 21 percent premium to the prior day’s close [14]. A regulated gas network that spent decades being valued as a stable dividend payer suddenly commands a control premium because of what may be plugged into it.
The cloud, in other words, is discovering geology.
Why I Chose the Term “Subsurface Compute”
I chose Subsurface Compute because the term captures a structural transformation that existing phrases — behind-the-meter generation, co-located power, bring your own power — describe only partially. Those expressions explain where electricity is generated. Subsurface Compute asks a deeper question: how far upstream must an AI company travel in order to secure the physical resources underlying computation?
If electricity scarcity becomes one of the principal constraints on AI expansion — and the evidence assembled in this paper suggests it already has — then control over computing capacity increasingly begins before electricity exists. It can begin with acreage, mineral rights, gas formations, pipeline capacity, geothermal reservoirs, drilling rights, turbine reservations, underground storage, and eventually, perhaps, pore-space rights for carbon sequestration. A parcel of rural land may therefore be valuable not merely because a datacenter can be constructed on top of it, but because an enormous energy asset exists beneath it. The Alpha Compute term sheet makes this literal: the acquisition bundles surface, mineral, and pore-space estates into a single AI infrastructure position [8].
The emerging chain becomes:
Geology → Fuel or Heat → Generation → Electricity → Datacenter → Accelerator → Model → Application → Agent
That is why Subsurface Compute fits naturally into the Five-Layer AI Economy framework that organizes this paper. Layer One — Energy — is no longer simply an external utility supplying Layers Two through Five. Layer One can increasingly determine the ownership structure, geography, financing, speed, and competitive position of the entire stack.
I also chose the term because it deliberately challenges the imagery surrounding artificial intelligence. AI is usually represented by clouds, satellites, networks, algorithms, and intangible intelligence. The vocabulary of the field points upward — toward abstraction, cognition, and eventually superintelligence. Yet the next phase of AI industrialization may depend increasingly on something profoundly tangible: rock formations, wells, pipes, turbines, transmission corridors, aquifers, and mineral estates. The International Energy Agency’s executive director, Fatih Birol, put the underlying dependence in a single sentence when the IEA published its landmark Energy and AI analysis:
“AI is one of the biggest stories in the energy world today.”
— Dr. Fatih Birol, Executive Director, International Energy Agency [15]
The more sophisticated artificial intelligence becomes, the more physical its upstream economy may become. That paradox — the immaterial product with the most material supply chain in the modern economy — is the subject of everything that follows.
A note on method. This paper is an argumentative synthesis rather than an econometric exercise. It draws on the primary record through August 18, 2026: corporate announcements and securities filings, state executive orders and regulatory dockets, quarterly earnings disclosures through Q2 2026, and analyses published by the International Energy Agency, the U.S. Department of Energy and its national laboratories, university research centers, and the financial and trade press. Where the paper describes candidates’ positions in the 2026 election cycle, it does so descriptively, as evidence of the issue’s political emergence, and endorses no party, candidate, or project. Where transactions remain proposed rather than closed — Alpha Compute’s acquisition, KKR’s bid for UGI, Chevron’s final investment decision on Kilby — the paper treats them as signals of strategic direction, not as accomplished facts.
Section 1: When the Cloud Meets the Wellhead
1.1 From Electricity Customer to Energy Developer
To understand how unusual the present moment is, it helps to reconstruct the model it is replacing.
For roughly two decades, the datacenter industry operated on a clean division of labor. Technology companies concentrated their capital and expertise on the things they understood — servers, networking, software, buildings, and the site-selection arithmetic of fiber routes, latency zones, tax abatements, and land prices. Electricity was abundant, cheap, and, above all, someone else’s problem. Utilities planned generation; regional transmission organizations planned wires; regulators allocated costs. A hyperscaler’s energy strategy consisted principally of negotiating rates, signing renewable power purchase agreements for accounting and sustainability purposes, and occasionally lobbying for favorable tariff treatment. The corporate boundary between the technology industry and the energy industry was one of the most stable boundaries in American capitalism.
That stability rested on a macro condition that has now dissolved: flat electricity demand. For nearly two decades after the mid-2000s, U.S. electricity consumption barely grew, and the grid’s spare capacity quietly absorbed the entire buildout of the first cloud era. Artificial intelligence ended that era with startling speed. Researchers convened by the MIT Energy Initiative noted that after decades of flat demand, computing centers now consume approximately 4 percent of U.S. electricity, with some projections suggesting the figure could rise to 12–15 percent by 2030, driven largely by AI applications [16]. The Department of Energy’s Lawrence Berkeley National Laboratory, in the industry’s standard-setting assessment, found that datacenters consumed about 4.4 percent of total U.S. electricity in 2023 and projected consumption of approximately 6.7 to 12 percent of total U.S. electricity by 2028, with total usage climbing from 176 terawatt-hours in 2023 to an estimated 325 to 580 terawatt-hours by 2028 [17]. Announcing the report, then-Energy Secretary Jennifer Granholm framed the demand surge as a national industrial phenomenon rather than a sectoral quirk:
“This industrial renaissance has created greater demand on our domestic energy supply.”
— Jennifer M. Granholm, U.S. Secretary of Energy [17]
Globally, the International Energy Agency projects that electricity demand from datacenters will more than double by 2030 to around 945 terawatt-hours — slightly more than the entire electricity consumption of Japan today — with AI-optimized facilities more than quadrupling their draw, the United States accounting for by far the largest share of the increase, and American datacenters on course to consume more electricity than the production of aluminum, steel, cement, chemicals, and all other energy-intensive goods combined by the end of the decade [18]. In its April 2026 update, the IEA reported that datacenter electricity demand grew 17 percent in 2025, that consumption at AI-focused facilities surged roughly 50 percent, and that the small-modular-reactor pipeline associated with datacenter operators had nearly doubled in a single year, from 25 to 45 gigawatts [19]. Dr. Birol captured the two-sided character of the moment in the 2026 update:
“While AI is still an energy taker, it is also becoming an energy maker.”
— Dr. Fatih Birol, Executive Director, International Energy Agency [19]
The consequence of demand growing this fast is not merely higher bills. It is queue physics. Gigawatt-scale loads can now appear on a utility’s doorstep faster than transmission systems and generating fleets can possibly be expanded to serve them. Bank of America analysts estimated in July 2026 that the United States will need more than 230 gigawatts of new generating capacity over the next five years while regulated utilities are on track to add only about 93 gigawatts of accredited supply — a gap of more than 100 gigawatts — with datacenters alone potentially adding roughly 125 gigawatts of load and large gas turbines effectively sold out through 2030 [20]. When the thing you need most cannot be bought at any price on the timeline you need it, rational firms stop buying it and start building it — and when building it requires fuel, they start securing the fuel.
The central thesis of this section can therefore be stated simply:
AI’s electricity problem has become sufficiently large that computing companies are crossing the traditional boundary separating the technology industry from the energy industry — and the crossing is happening upstream, link by link, toward the resource itself.
1.2 Alpha Compute and the Marcellus Proof Point
Every structural argument needs a case that makes the abstraction concrete. For Subsurface Compute, that case is Alpha Compute’s Pennsylvania term sheet — not because it is the largest transaction in this paper (it is by far the smallest), but because it is the purest.
Consider precisely what the company proposes to buy. According to the Reuters report and the company’s own disclosure, the binding term sheet grants Alpha Compute Management, LLC an exclusive option to acquire, for a base price of $55 million with a $3 million deposit credited at closing: unleased Marcellus gas rights covering approximately 1,800 oil and gas mineral acres carrying a 100 percent net revenue interest, subject to title confirmation; the associated surface estate on which a greenfield 200-megawatt behind-the-meter datacenter campus would be constructed; and the pore-space rights beneath it, while excluding existing leasehold rights in the deeper Utica and other formations [7] [8]. The planned structure is vertically closed: gas produced from the property fuels on-site generation; on-site generation powers the campus; and Alpha Compute itself serves as the primary capacity and power offtaker for its GPU-as-a-service operations.
The importance of the transaction is not merely that a company plans to build a gas-powered datacenter — dozens of developers now propose that. It is that land and mineral rights are being assembled as components of an AI infrastructure strategy. The deal fuses four estates that American commerce has traditionally traded separately:
Land ownership + Mineral ownership + Power generation + Computing capacity
Each element changes the meaning of the others. The mineral estate converts the land from a construction site into a fuel reserve. The generation plan converts the fuel reserve from a commodity royalty stream into a dedicated cost hedge — the third-party estimate of roughly 5.85 cents per kilowatt-hour, if validated, would function as a decade-long fixed input price for the scarcest input in AI [8]. And the computing offtake converts the entire energy complex from an infrastructure project into the substrate of a technology company’s product margin.
Three caveats keep the case honest. First, the transaction is an option, not a closing; it remains subject to diligence, financing, permitting, and definitive documentation, and the company itself notes that no power or datacenter capacity currently operates at the site [8]. Second, the economics rest on a third-party reserve evaluation whose assumptions — decline curves, drilling costs, gas quality, deliverability — require validation, as the company acknowledges. Third, the project must now clear Pennsylvania’s GRID regime, including local approval, adopted one week after the announcement; the sequencing of those two events is itself part of this paper’s story. But even as a proposal, the term sheet establishes the proof point: the question “what lies beneath the acreage?” has formally entered datacenter site selection.
It also invites a forward-looking question this paper returns to in Section 5: will future datacenter acquisitions increasingly resemble energy-development transactions — with reserve reports, title opinions, division orders, and royalty schedules — rather than conventional commercial-real-estate transactions with appraisals and environmental Phase I reports? The Alpha Compute documents read far more like the former than the latter, and that stylistic fact is a leading indicator.
1.3 The UGI–Prime Model: Pipeline Meets GPU
If Alpha Compute represents resource ownership, the UGI–Prime partnership represents the next progression outward from the wellhead: infrastructure integration — the deliberate coupling of midstream gas assets with hyperscale computing.
On May 6, 2026, UGI Energy Services and Prime Data Centers announced a strategic partnership to develop natural-gas supply infrastructure in Pennsylvania’s northern tier for a proposed on-site gas-fired generation complex serving one of the state’s largest planned AI and high-performance-computing campuses. The structure is instructive: UGI will sell Prime a portion of its own property for the generation facility while retaining approximately 15 billion cubic feet of underground storage capacity and related oil and gas rights; it will develop a dedicated pipeline and associated facilities, with planned midstream investment exceeding $100 million; and the site sits atop redundant supply pathways — locally produced Marcellus gas, the Eastern Gas Transmission and Storage system, the Tennessee Gas Pipeline, and UGI’s own distribution network [9] [21]. Prime’s expected demand — more than 100,000 dekatherms per day within three to five years, roughly 100 million cubic feet daily — would by itself constitute one of the region’s largest new industrial gas loads [9]. UGI Energy Services president Joseph Hartz described the project as
“a strong fit for our midstream capabilities.”
— Joseph Hartz, President, UGI Energy Services [21]
while Prime’s chief executive, Nicholas Laag, explained the site-selection logic in terms that could serve as a one-sentence definition of Subsurface Compute geography:
“A rare combination of direct Marcellus Shale access, robust pipeline infrastructure.”
— Nicholas Laag, Chief Executive Officer, Prime Data Centers [22]
What the partnership illustrates is that pipeline companies and datacenter developers are becoming natural counterparties. Dedicated gas infrastructure can remove or dramatically reduce the single largest uncertainty facing AI developments: the multiyear wait for large utility interconnections. A pipeline lateral and an on-site generation complex can be permitted and constructed on timelines the developer substantially controls, fed by a resource base whose deliverability is contractually secured, with underground storage providing the buffer that the electric grid would otherwise supply.
It is useful to formalize the physical dependency this creates. Call it the Fuel-to-Compute Chain:
| Link | Asset | Who Traditionally Owned It | Who Increasingly Controls It |
| 1 | Reservoir / formation | Mineral owners, E&P companies | AI developers (Alpha Compute), integrated energy majors |
| 2 | Well | E&P companies | Energy majors, project ventures |
| 3 | Pipeline / storage | Midstream companies | Midstream–datacenter partnerships (UGI–Prime) |
| 4 | Turbine / power plant | Utilities, IPPs | Hyperscaler-dedicated ventures (Energy Forge One / Kilby) |
| 5 | Substation & campus grid | Utilities | Datacenter developers |
| 6 | Server rack / GPU | Cloud companies | Cloud companies |
| 7 | Model | AI labs | AI labs |
| 8 | Application / agent | Software companies | Software companies |
Every link in the upper half of this table represents infrastructure that must exist before a frontier model can train or an AI agent can perform useful work. The strategic novelty of 2025–2026 is that ownership and control are migrating upward through the table — from links 6–8, where technology companies have always lived, into links 1–5, where they never previously ventured.
1.4 Why Grid Waiting Changes Corporate Boundaries
Economists since Ronald Coase have understood that the boundary of the firm — what a company chooses to own rather than buy — is set by transaction costs, and that firms integrate vertically when markets fail to deliver critical inputs reliably. The AI electricity crunch is a textbook market failure of exactly this kind, and the interconnection queue is its clearest measurement.
The numbers involved have become almost surreal. In Texas alone, ERCOT’s large-load interconnection queue surged to 474 gigawatts by mid-2026 — approximately 90 percent of it datacenters — a volume Governor Abbott himself noted exceeds five times the state’s all-time record peak demand [23]. Even discounting heavily for speculative and duplicative requests, the queue represents demand that no utility planning process, transmission buildout, or turbine supply chain can serve on the timelines AI developers require. BloombergNEF, meanwhile, has tracked roughly 100 to 126 gigawatts of planned on-site gas-burning capacity intended specifically to power datacenters — equivalent to 18 percent of the entire existing U.S. gas fleet — precisely because developers no longer trust the grid to arrive on time [24] [25].
When electricity was abundant, owning gas infrastructure would have seemed pointlessly vertical — a distraction from a technology company’s genuine comparative advantage, punished by investors as capital indiscipline. When electricity becomes the binding constraint, the calculus inverts. Consider the asymmetry of losses. A hyperscaler that overpays modestly for owned or dedicated generation loses basis points of margin. A hyperscaler whose multibillion-dollar GPU fleet sits energized-but-idle — or worse, un-energized in a warehouse — while an interconnection queue grinds forward loses the market itself: model-training windows, customer commitments, and competitive position against rivals who solved power first. Depreciating accelerators are perhaps the fastest-melting capital asset in industrial history; every quarter of delay converts state-of-the-art silicon into last-generation silicon. Against that loss function, securing upstream energy assets is not empire-building. It is insurance — and cheap insurance at that.
This is why the corporate-boundary question now runs through every hyperscaler earnings call. In the Q2 2026 reporting season, the four largest hyperscalers — Amazon, Microsoft, Alphabet, and Meta — collectively guided toward roughly $725 to $760 billion of 2026 capital expenditure, up approximately 77 percent from about $410 billion in 2025 [60], with executives explicitly describing a strategy of committing early to long-lived assets — land, datacenter shells, and power — while deferring chip purchases until demand is visible [26] [27] [28]. Power, in other words, has moved into the same category as land: the thing you secure first, years ahead, because it cannot be conjured later. Amazon chief executive Andy Jassy reported AWS growth of 36.7 percent — its fastest in eighteen quarters — alongside a roughly $220 billion capex trajectory, a pairing that only makes sense if the energized capacity to serve that growth has been locked down far in advance [28].
The deeper point is Coasean: scarcity redraws firms. The interconnection queue is doing to the technology industry what unreliable component markets did to the early automobile industry — pulling the assembler backward into its own supply chain. The difference is that the supply chain of intelligence terminates not in a parts supplier but in the earth’s crust.
1.5 Subsurface Compute and the Five-Layer AI Economy
The framework that organizes the remainder of this paper is the Five-Layer AI Economy, and it is worth mapping the concept onto it explicitly:
| Layer | Domain | Contents | Subsurface Compute’s Effect |
| Layer 1 | Energy | Wells, pipelines, geothermal reservoirs, turbines, generation, storage, transmission | Transforms from external utility into owned/contracted strategic asset |
| Layer 2 | Chips | NVIDIA, AMD, custom accelerators requiring continuous high-density power | Utilization — and therefore ROI — becomes hostage to Layer 1 |
| Layer 3 | Datacenters | The physical point where energy becomes computational capacity | Site selection migrates from fiber geography to fuel geography |
| Layer 4 | Models | Training and inference whose economics depend on available compute | Training cost curves inherit Layer 1 cost curves |
| Layer 5 | Applications & Agents | The ultimate economic output | Retail price of intelligence embeds the price of molecules and heat |
Two observations follow from the table. First, the layers are not merely stacked; they are load-bearing. A constraint at Layer 1 propagates upward with almost no attenuation, because the intermediate layers — chips, datacenters, models — have essentially no ability to substitute away from electricity. Efficiency gains are real and continuing; the IEA finds power consumption per AI task declining at an unprecedented rate [19]. But at the system level, efficiency has so far functioned as a demand accelerant rather than a demand ceiling — the classic Jevons dynamic — because cheaper computation expands the set of economically viable AI applications faster than it shrinks the energy bill of existing ones.
Second, ownership at Layer 1 changes bargaining power everywhere above it. A model developer renting compute from a cloud whose energy costs float with wholesale power markets holds a different competitive position than one whose compute rests on a twenty-year fixed-price gas or geothermal contract — or on 1,800 mineral acres. The central insight of the section, and in many ways of the paper, is this:
The Five-Layer AI Economy may look digital from the top, but its bottleneck increasingly originates at the bottom — and the bottom is, with growing literalness, underground.
Section 2: The Economics of Owning the Molecule Before Buying the Electron
2.1 From Real Estate to Energy Estate
Traditional datacenter land valuation is a well-understood discipline. Appraisers and site-selection consultants price acreage against a familiar checklist: zoning and entitlement risk; proximity to long-haul and metro fiber; latency to population centers and cloud regions; construction logistics; tax abatements and sales-tax exemptions on equipment; and — the item that has migrated from the bottom of the checklist to the top — access to transmission capacity and a credible interconnection position. For twenty years, those variables fully described the asset. The dirt was a platform; its value was its surface.
Subsurface Compute introduces a variable that the commercial-real-estate tradition has no standard field for:
What lies beneath the acreage?
A strategically located mineral estate could theoretically provide both the site and a substantial share of the site’s future energy supply — and, through pore space, potentially its future emissions-disposal capacity as well. When Alpha Compute’s term sheet prices 1,800 mineral acres, a surface estate, and pore-space rights as a single $55 million AI-infrastructure position [8], it is implicitly asserting that these estates are worth more fused than separate — that there is a conglomerate premium in stacking fuel, land, and compute demand on one title. The county assessor’s reported estimate that the completed project could add $2.08 billion of assessed value [8] gives a sense of the multiplier at stake: the underground asset is a rounding error against the surface asset it enables, which is precisely why controlling it is so cheap relative to its option value.
This is the quiet revaluation now underway across energy-rich rural America. Land above productive shale, adjacent to trunk pipelines, near underappreciated substations, or atop high-temperature geothermal gradients is beginning to carry a shadow price set not by agriculture, timber, or even conventional energy royalties, but by its potential to host manufactured intelligence. The real-estate category is dissolving into something better described as an energy estate.
2.2 Mineral Rights as an AI Infrastructure Asset
To non-American readers — and to many technology executives — the pivotal legal fact enabling this entire strategy is unfamiliar: in the United States, unlike most of the world, subsurface minerals are typically privately owned and severable from the surface. A landowner can sell the farm and keep the gas; sell the gas and keep the farm; lease the minerals to one company and the surface to another. Over more than a century of drilling history, this has produced the famous “split estate”: millions of parcels where surface ownership and mineral ownership diverged generations ago, tracked through county courthouse records, fractionalized among heirs, and traded in a specialized market of landmen, title attorneys, and mineral-acquisition funds.
For most of that history, the split estate was a problem for datacenter developers only in the negative sense — a title-diligence item, a risk that someone else’s drilling rights might encumber the surface. Subsurface Compute turns the split estate into an opportunity set. Unleased mineral acreage over a proven formation is, functionally, an unexercised call option on fuel; a developer who acquires it alongside the surface acquires the right to convert geology into electrons at a timing and price of its own choosing. Alpha Compute’s disclosure is explicit on this point — the Marcellus rights it seeks are unleased, carrying a 100 percent net revenue interest, meaning no intervening lessee or royalty chain dilutes the economics [8].
The practical consequence is that AI-campus due diligence is expanding beyond its traditional scope. The conventional checklist —
- transmission capacity and queue position,
- water availability,
- fiber connectivity,
- zoning and entitlements,
- taxes and incentives —
now increasingly extends to a second, geological checklist:
- mineral ownership and severance history,
- gas-production potential and reserve quality,
- pipeline access and takeaway capacity,
- drilling restrictions, setbacks, and unitization rules,
- geothermal gradient and reservoir characteristics,
- environmental liabilities (legacy wells, subsidence, induced seismicity),
- underground storage and pore-space rights.
Law firms, title companies, and reserve engineers who spent their careers serving exploration-and-production clients are discovering a new client class. The datacenter industry, which once hired fiber-route consultants, is beginning to hire petroleum geologists.
2.3 Behind-the-Meter Generation: Why the Electron Wants to Be Born On-Site
The economic logic of co-located, behind-the-meter generation deserves careful statement, because it is the hinge between resource ownership and computing capacity.
Electricity produced adjacent to the datacenter offers four distinct advantages. First, time: an on-site plant’s schedule is governed by equipment procurement and air permitting rather than by a regional interconnection queue measured in half-decades. Second, cost transparency: fuel plus conversion plus fixed charges, without the accumulating layers of transmission tariffs, congestion, capacity-market volatility, and rate-case risk embedded in delivered grid power. Third, reliability engineering under the developer’s own control: redundancy, storage, and maintenance regimes designed for a single customer’s uptime requirements. Fourth, and increasingly decisive, political insulation: a campus that manifestly brings its own generation is far easier to defend before regulators and voters worried about residential rate impacts — indeed, Pennsylvania’s GRID order and Governor Abbott’s Texas directives effectively make bring-your-own-power a condition of civic legitimacy [3] [29].
None of this makes behind-the-meter generation free of constraint. Such projects remain fully subject to environmental permitting, fuel-supply risk, turbine availability (with large frames effectively sold out through 2030 [20]), reliability engineering without a grid backstop, and — as Section 5 examines — a serious emissions ledger. Nor does self-supply spare the neighbors entirely: analyses of the behind-the-meter boom warn that on-site plants still compete for the same turbines, gas supply, and skilled labor as grid resources, and can raise system costs through channels other than direct demand [24].
Chevron’s Project Kilby with Microsoft is the corporate-scale demonstration of the model. Under the 20-year power purchase agreement announced June 22, 2026, Chevron’s wholly owned subsidiary Energy Forge One — partnered with Joulent, the energy venture of investment firm Engine No. 1 — will develop approximately 2.67 gigawatts of dedicated, phased, modular gas-fired capacity on more than 2,000 acres in Reeves County, in the heart of the Permian Basin, with first power expected in 2028, most generation from large GE Vernova turbines supplemented by Caterpillar’s Solar Turbines units, brackish rather than fresh water for plant operations, and selective catalytic reduction for NOx control [10] [30]. The fuel comes from Chevron’s own Permian production — including gas that regional pipeline constraints would otherwise strand or depress in price, a dynamic the company presents as a structural cost advantage [31]. Chevron’s president of New Energies, Jeff Gustavson, stated the strategic premise directly:
“Abundant, affordable, reliable energy is essential to fueling that transformation.”
— Jeff Gustavson, President of New Energies, Chevron [10]
Analysts have sketched the project’s financial anatomy: an estimated capital outlay of roughly $9 billion, largely project-financed, targeting developer returns in the mid-teens, implying a power price to Microsoft in the neighborhood of $150 per megawatt-hour, with Microsoft expected to add approximately two gigawatts of datacenter capacity at the site over five to seven years [32]. Note what that price implies: Microsoft is plausibly paying a premium over average grid power — and doing so gladly — because what it is buying is not energy in the abstract but firm, dedicated, schedulable energy at a location and timeline of its choosing. Scarcity has made delivery certainty itself the product.
Months before the definitive agreement, when the parties announced their exclusivity arrangement, their joint statement already described the template in general terms — an approach
“bringing energy supply closer to demand through co-located, behind-the-meter generation.”
— Joint statement of Chevron, Microsoft, and Engine No. 1 [33]
That sentence, issued by an oil supermajor, a hyperscaler, and an activist investment firm speaking in one voice, is as concise a manifesto for Subsurface Compute’s middle layers as the record offers.
2.4 The New Power Hedge
Step back from individual projects and consider the risk book of a large AI company in 2026. It is simultaneously exposed to:
Electricity price risk + Interconnection/timing risk + GPU depreciation risk + Construction-delay risk + Model-demand risk
These exposures interact viciously. Interconnection delay converts directly into GPU depreciation loss; electricity price spikes flow directly into inference margins; construction delay compounds both. Traditional financial hedges cover almost none of this, because the underlying risk is physical delivery, not price alone. Wholesale power can be hedged for months or a few years at liquid hubs; it cannot be hedged for fifteen years at a specific rural substation that does not yet exist.
Controlling part of the energy supply chain hedges several of these risks at once — while, honesty requires noting, introducing new ones. An owned or dedicated gas position fixes fuel cost and delivery but adds commodity-basis risk, operational risk, reserve risk, and a long-lived emissions liability. A geothermal framework fixes carbon and fuel cost but adds subsurface-performance risk. The portfolio result is that the modern AI campus increasingly resembles a hybrid instrument: part technology plant, part merchant power project, part energy trading position. Its owners must now think like utilities about reliability, like producers about depletion, and like traders about basis — while still thinking like technologists about tokens per second per watt.
This blended identity also explains a subtle shift in disclosure. Hyperscaler earnings commentary in Q2 2026 dwelt on power procurement, useful-life assumptions, and the sequencing of long-lived versus short-lived assets to a degree that would have been unimaginable in a technology earnings call five years ago [27] [28]. The market is learning to read cloud companies partly as energy companies, because that is partly what they have become.
2.5 Financing Subsurface Compute: Wall Street Discovers the Basement
The final piece of the economic architecture is capital formation, and here 2026 has supplied its own punctuation mark.
The layers of a Subsurface Compute project map onto distinct pools of capital with distinct costs and risk appetites. Regulated pipeline and distribution assets attract infrastructure funds and pension capital at low required returns; merchant generation under long-dated hyperscaler PPAs attracts project finance, as the Kilby structure illustrates [32]; the datacenter shell attracts real-estate and digital-infrastructure credit; the GPUs attract corporate balance sheets and, increasingly, vendor and private-credit structures; upstream reserves attract energy private equity. A fully integrated project is therefore not one financing but a stack of financings, each secured by a different layer of the Fuel-to-Compute Chain. This is precisely how railroads, electrification, and telecommunications were financed in their own buildout eras — and the aggregate sums now involved are of that magnitude: Goldman Sachs projects combined capital expenditure for the four largest hyperscalers of $5.3 trillion between fiscal 2025 and 2030, with baseline aggregate estimates reaching $7.6 trillion across compute, datacenters, and power through 2031 [26].
A particularly timely signal arrived on August 18, 2026 — the same day as Shapiro’s executive order. Reuters, citing a Wall Street Journal report, disclosed that KKR had offered approximately $9 billion, or $42.50 per share, to acquire UGI Corporation — the Pennsylvania-based natural gas and electricity distributor whose Energy Services arm had, three months earlier, announced the Prime Data Centers pipeline partnership. The offer represented a 21.1 percent premium; UGI shares jumped more than 12 percent, trading was briefly halted for volatility, and coverage universally framed the bid within the surge of AI-driven electricity demand that has put reliable gas infrastructure “in greater focus” [14] [34]. The proposal may or may not result in a completed transaction, and no inference is drawn here about its merits. But its interpretive value is unambiguous: a diversified, unglamorous gas utility — propane distribution, local pipes, 15 Bcf of storage in the northern tier — suddenly commands a control premium from one of the world’s most sophisticated infrastructure investors, at the exact moment its assets sit between the Marcellus and the datacenter boom.
The larger question the bid poses is the one this section has been building toward:
Will AI investment cause Wall Street to systematically revalue pipelines, gas storage, geothermal acreage, turbine order books, and energy-rich land according to their proximity to future compute?
The early evidence — Fervo’s $2.2 billion IPO as the largest renewable-energy listing in history on the strength of a hyperscaler framework agreement [35], the KKR–UGI premium, the mid-teens returns targeted on hyperscaler-anchored generation [32] — suggests the revaluation has already begun. Proximity to compute is becoming a pricing factor for energy assets in the way proximity to ports once priced industrial land.
Section 3: America’s New Geological Map of Artificial Intelligence
If Subsurface Compute is real, it should be visible on a map — and it is. The four regional studies that follow are not a survey of all American datacenter geography; they are chosen because each demonstrates a distinct variant of the same underlying logic: computation migrating toward controllable subsurface energy.
3.1 Pennsylvania: Marcellus Meets Machine Intelligence
Pennsylvania is the central geographic case study of this paper because it concentrates, within a single state, every force the concept describes — and every force resisting it.
The Commonwealth’s appeal to AI developers is a stacked inheritance. It sits atop the Marcellus, among the largest natural-gas formations on earth, giving its northern tier what Prime Data Centers’ CEO called direct shale access with built-in supply security [22]. It possesses a dense legacy of pipeline infrastructure — Eastern Gas Transmission and Storage, Tennessee Gas Pipeline, UGI’s network with its underground storage — laid down across a century of energy production [9]. It offers legacy industrial sites with existing grid connections, water access, and brownfield entitlements; substantial nuclear generation; proximity to East Coast population centers and the Northern Virginia cloud complex; abundant rural land; and membership in PJM, the nation’s largest wholesale market. This is why more than 100 projects have appeared in public databases and why Amazon alone committed $20 billion to Bucks and Luzerne County campuses [1] [3].
But those advantages now collide with three countervailing pressures, each on display in the events of August 2026. The first is community opposition, which Shapiro’s order institutionalizes by conditioning state permits on local approval — effectively a distributed veto [3]. The second is electricity affordability: PJM capacity prices and residential bills have become the vocabulary through which ordinary Pennsylvanians experience the AI boom, and the GRID Requirements’ insistence that projects bring their own generation and pay all associated energy costs is a direct response [1] [3]. The third is environmental accounting: a July 2026 report by the Environmental Integrity Project found datacenter developers in Pennsylvania planning to rely on at least seven new gas-fired plants emitting roughly 68 million tons of CO₂-equivalent annually — part of a national pipeline of at least 74 such plants whose combined emissions would rival Australia’s — with the largest single behind-the-meter site being the redevelopment of the former Homer City coal station [36].
Pennsylvania is therefore almost an ideal laboratory: the state where the resource, the infrastructure, the capital, the backlash, and the regulatory response are all maximal simultaneously. Alpha Compute’s mineral-rights proposal and Shapiro’s GRID order, arriving seven days apart, are best understood as the thesis and antithesis of a single dialectic whose synthesis — responsible, locally consented, self-powered subsurface development, or stalemate — the next several years will reveal.
3.2 West Texas: The Permian-to-Processor Corridor
Texas presents the second variant: not resource acquisition by computing companies, but resource integration by energy companies moving downstream to meet compute halfway.
Project Kilby demonstrates why abundant natural gas can reorganize datacenter geography. Rather than forcing every gigawatt of new AI demand through an already strained transmission network, developers can select energy-rich locations where generation and computing are developed together as a single industrial complex. The Reeves County site sits amid Chevron’s Permian production, where associated gas frequently exceeds pipeline takeaway capacity, depressing local prices and forcing flaring — meaning the project monetizes molecules that were, at the margin, nearly valueless where they stood [31]. Kilby’s 2.67 gigawatts, delivered under a 20-year PPA with phased GE Vernova and Solar Turbines capacity, brackish-water cooling, and mid-teen target returns, positions it among the largest co-located gas-and-datacenter developments in the country [10] [30] [32]. Chevron frames the venture explicitly as diversification — cash flow less exposed to oil and gas price cycles — and as differentiation, with Gustavson noting that while peers discuss such projects, Chevron is executing one; Texas, by BloombergNEF’s count, leads the nation with roughly 33 gigawatts of planned datacenter power projects [37].
Fermi America’s Project Matador near Amarillo provides the maximal variation on the theme: a self-powered “HyperGrid” campus on 5,236 acres leased for 99 years from the Texas Tech University System, planned at up to 11 gigawatts across natural gas, nuclear (including four proposed AP1000 reactors), solar, and battery storage, with approximately 6 gigawatts of gas generation already through preliminary state environmental approval, more than $1.5 billion invested to date, and first power targeted for 2026 [38] [39]. The company’s own siting rationale reads like a Subsurface Compute checklist: adjacency to one of the largest known natural-gas fields in the United States, a high-radiance solar corridor, and positioning for advanced nuclear [40]. Whatever one makes of the project’s ambitions — and its financing and tenant commitments have fluctuated — its premise is the premise of this paper stated at industrial-park scale: the grid cannot grow fast enough, so the campus must contain its own energy system, anchored to the resources beneath and around it.
Together, Kilby and Matador sketch what might be called the Permian-to-Processor Corridor: a West Texas geography where the stranded-gas discount, cheap land, permissive siting, and ferocious solar resource combine to make electrons cheaper to manufacture beside the wellhead than to import across the wires.
3.3 Louisiana: AI Creates a Gas-and-Compute Industrial Cluster
Louisiana demonstrates the third variant: a single hyperscale project large enough to reorganize an entire state’s generating portfolio — Subsurface Compute executed through the utility rather than around it.
Meta’s Hyperion campus in Richland Parish began, in December 2024, as a $10 billion, two-gigawatt project. By October 2025 the announced investment had reached $27 billion; on July 13, 2026, Meta announced expansion to five gigawatts of compute capacity across nearly ten million square feet at a total investment exceeding $50 billion — the company’s largest datacenter worldwide and one of the largest ever built, with peak construction employment above 7,500, some 1,000 permanent roles, more than $1.6 billion already contracted to local businesses, and over $1 billion in local road, water, and wastewater improvements [11] [41]. Meta’s vice president of datacenters, Rachel Peterson, framed the undertaking in deliberately civic terms:
“It’s about building alongside the community.”
— Rachel Peterson, Vice President of Data Centers, Meta [11]
The energy program attached to the campus is where the concept lives. Meta’s agreements with Entergy Louisiana fund seven new combined-cycle natural-gas generating plants totaling more than 5.2 gigawatts, grid-scale batteries at three sites, roughly 240 miles of new 500-kV transmission, nuclear uprates, additional purchased power, a commitment to help fund up to 2.5 gigawatts of clean and renewable energy, a memorandum of understanding on future nuclear development, $215 million for bill-assistance and efficiency programs, and — by Entergy’s estimate — approximately $2.65 billion in benefits to the broader customer base over twenty years, with Meta stating that it pays the full costs of energy, water, and related infrastructure so those expenses are not passed to consumers [41] [42] [43]. Entergy Louisiana’s president was categorical:
“Meta’s commitment isn’t just transforming north Louisiana, it’s directly benefiting every Entergy Louisiana customer.”
— Phillip May, President and CEO, Entergy Louisiana [43]
That claim is genuinely contested, and the contest matters for the national debate. Consumer and environmental intervenors — including the Alliance for Affordable Energy — have challenged the confidentiality of Meta’s job and power commitments, questioned whether Entergy’s 7.5 gigawatts of approved-or-pending gas capacity (more than double Meta’s stated need) is fully justified, and won procedural rulings requiring greater disclosure, even as state legislation curtailed the administrative review [44]. Louisiana thus previews the accountability questions that follow whenever a single private computing project becomes the organizing fact of a state’s resource plan: Who verifies the load forecast? Who audits the ratepayer-protection math? Who bears the risk if five gigawatts of demand arrives late, smaller, or never — a stranded-asset scenario that analysts at RMI note echoes the utility overbuilds of the 1970s [45]?
The lesson of Louisiana is double-edged: a hyperscale AI cluster can be large enough to modernize a regional energy system — and large enough that its assumptions, if wrong, become everyone’s problem.
3.4 Utah, Nevada, and the Geothermal Variant: Subsurface Without the Carbon
Subsurface Compute must not be mistaken for a synonym for natural gas, and the Google–Fervo relationship is the decisive counterexample.
Enhanced geothermal systems (EGS) apply the directional-drilling and fracturing toolkit of the shale revolution to hot dry rock, engineering artificial reservoirs where water can be circulated through fractures at depths approaching 10,000 feet and temperatures above 500°F, returning to the surface to drive turbines around the clock. Fervo’s 3.5-megawatt Project Red pilot in Nevada — the world’s first corporate EGS agreement, with Google — proved the technology in 2023 and began serving the grid supplying Google’s Nevada datacenters; Google has since called out the Department of Energy’s finding that geothermal could ultimately provide up to 120 gigawatts of reliable, flexible U.S. capacity by 2050 [46] [47]. Fervo’s Cape Station project in Beaver County, Utah — 500 megawatts under construction, permitted toward 2,000 — anchors the scale-up, with Phase I’s roughly 100 megawatts on track for first power in Q4 2026 [35].
The March 2026 Geothermal Framework Agreement transforms the relationship from pilot to pipeline: Google receives a right of first refusal over up to three gigawatts of electricity from designated new Fervo projects through 2033, with at least one gigawatt of projects to be proposed within the first two years [12] [48]. Two months later, Fervo completed a $2.2 billion IPO — the largest renewable-energy listing in history — and secured $421 million of non-recourse project financing for Cape Station Phase I, capital-markets validation that hyperscaler demand can now underwrite an entirely new category of subsurface asset [35].
The feedback loop closed in June 2026, when Fervo, NVIDIA, and Pacific Northwest National Laboratory announced EGS-Twin: a digital-twin platform that trains AI models on Fervo’s Nevada and Utah field data using NVIDIA infrastructure and DOE supercomputing, integrating physics-based simulation with AI forecasting inside NVIDIA’s Omniverse libraries to give operators near-real-time insight into reservoir behavior, targeted for deployment by 2029 and ultimately available to any geothermal operator [13] [49]. Fervo’s co-founder and chief technology officer described the ambition plainly:
“Digital twins will expedite the learning curve for geothermal development.”
— Jack Norbeck, Co-Founder and CTO, Fervo Energy [49]
The loop deserves emphasis because it is the concept’s purest expression: AI modeling the subsurface, to extract subsurface energy, to power more AI. Nor is geothermal the only low-carbon branch. The same hyperscaler procurement muscle is pulling forward small modular reactors — a datacenter-linked SMR pipeline the IEA now sizes at 45 gigawatts [19] — and Meta’s Louisiana program bundles nuclear uprates and 2.5 gigawatts of renewables into its gas-heavy core [41]. Subsurface Compute is therefore technological rather than ideological: its unifying question is simply what underground or firm resource can deliver reliable energy for computation, where, and how fast — and gas, heat, uranium, and stored electrons are all admissible answers with different clocks and different carbon ledgers.
3.5 From the Datacenter Map to the Geological Map
Northern Virginia became the datacenter capital of the world for reasons that were fundamentally informational: MAE-East and the earliest internet exchange points, dense fiber, federal demand, cloud-ecosystem gravity, and a generation of accumulated expertise. Its geography was the geography of the network.
The regional evidence assembled above suggests the next generation of AI campuses is being sited by a different logic. The emerging clusters form around gas basins (the Marcellus, the Permian), existing and new nuclear stations, geothermal fields of the interior West, high-capacity substations and energy corridors, and utilities willing to build — a geography of energy, not information. Even Virginia’s own continued growth now runs through energy policy, as Section 4 shows, with the state’s regulators redesigning rates specifically to govern it.
The proposition, stated as a map-making rule: America’s future AI map will increasingly resemble its energy map — and, one layer further down, its geological map. Where the crust holds recoverable fuel or heat, and the surface holds a community willing to consent, intelligence factories will follow. Where either is absent, fiber alone will no longer suffice.
Section 4: Subsurface Compute Enters the 2026 Ballot Box
Energy has always been political; computing, until now, largely was not. The 2026 election cycle is the moment those two conditions merged. This section examines three state models — Pennsylvania’s guardrails, Texas’s conditional admission, Virginia’s rate design — and argues that together they constitute an emerging American social contract for gigawatt computing.
4.1 Pennsylvania: Growth Meets Guardrails
Return to the opening anecdote with the benefit of the intervening analysis. Governor Shapiro’s August 18 action is important precisely because Pennsylvania possesses extraordinary energy advantages — the political system of the best-endowed state is the natural bellwether for how democratic institutions will price those endowments.
The mechanics of Executive Order 2026-05 amount to a permitting philosophy inversion. Where the state’s earlier posture treated speed as the deliverable — the Fast Track program the administration once touted as its answer to red tape — the new posture treats conditionality as the deliverable. Permits will be reviewed only for projects with legally binding GRID commitments across energy affordability, environmental protection, workforce and economic development, transparency, and community engagement; reviewed but not issued until local land-use approvals are complete; stripped of expedited treatment; and barred from confidentiality agreements with state agencies — with the administration explicitly framing the order as a tool to “weed out speculative proposals and hold bad actors accountable” in a landscape where only five of more than a hundred rumored projects hold complete first-phase permits [1] [4]. The order’s requirements that projects bring their own generation and fund community-benefit agreements convert Subsurface Compute from a corporate strategy into a regulatory expectation [3].
The politics are frankly competitive, and the paper presents them descriptively. Shapiro — a first-term Democrat, a reported 2028 presidential prospect, and formerly the industry’s most prominent state-level champion — now argues that only enforceable standards can preserve public consent for growth [3] [5]. His Republican challenger, Treasurer Stacy Garrity, has advocated pausing new datacenter development and attacks the governor as the author of the disorder his order claims to cure [6]. Senate Republicans counter that they acted on transparency through the budget and accuse the governor of erasing his own record [5]. The analytically important fact is what no one is arguing: no significant Pennsylvania actor now defends unconditional, expedited, ratepayer-socialized datacenter growth. The argument is exclusively about which conditions, imposed by whom. That is what the arrival of a social contract looks like.
4.2 Texas: From Recruitment to Conditional Admission
Texas represents the same transition executed through a different institutional grammar — executive directives to regulators rather than permitting orders — and by a Republican governor, which is precisely what makes the development bipartisan in character.
The chronology is compressed and telling. In February and March 2026, the Public Utility Commission adopted rules setting minimum information standards for large-load forecasting and requiring approval before large loads could net-meter with existing generation resources [23]. On June 10, Governor Abbott directed the PUC to ensure datacenter interconnections reduce residential bills, to require datacenters to pay all electric infrastructure costs they cause, and to identify further ratepayer safeguards [29] [50]. He simultaneously called for legislation requiring new facilities to add generation to the grid, fund their own interconnection, use closed-loop or similarly water-efficient cooling, file annual electricity and water consumption reports, and meet statewide community-impact standards [51]. Then, on August 3, confronting an ERCOT large-load queue of 474 gigawatts — about 90 percent datacenters, against a record system peak of 91,089 megawatts — Abbott ordered a comprehensive audit of every project in the queue, with disclosure of ownership, financing, water use, incentives, and community impacts, and paused interconnection advancement pending completion. His arithmetic was the argument:
“That is more than five times Texas’ record peak electricity demand for ERCOT.”
— Governor Greg Abbott, letter to the PUCT and ERCOT, August 3, 2026 [23]
Any project failing the audit, the governor warned, would be denied grid access; at least one developer, unable to meet the standards, has already withdrawn a project, while hyperscale operators including QTS, MARA, and Montera have publicly committed to the framework — several noting that they already bring their own power, curtail during grid stress, and fund their own interconnection [52] [53]. Meanwhile, Democratic challenger Gina Hinojosa campaigns on the industry’s effects on rural land, water, and power, framing Abbott as the industry’s ally rather than its regulator [6]. Again the positions are presented as evidence, not endorsement; and again the convergence is the finding. A Republican governor of the nation’s most development-friendly state and his Democratic challenger disagree about blame and stringency — but both now operate inside the same premise:
Gigawatt computing has become large enough to require a political social contract.
4.3 Virginia: Who Pays for the AI Grid?
Virginia — the incumbent capital of the datacenter world — contributes the third model: neither guardrails nor audits, but rate design.
In November 2025, the State Corporation Commission’s final order in Dominion Energy’s biennial review created the GS-5 rate class for customers demanding 25 megawatts or more, effective January 1, 2027 — a threshold that captures most of the roughly 450 datacenters Dominion already serves. GS-5 customers must sign 14-year contracts (with a load-ramp period of up to four years); pay minimum monthly charges equal to at least 85 percent of their contracted transmission and distribution demand and 60 percent of generation demand regardless of actual usage; post collateral reported at $1.5 million per megawatt; give three years’ notice of demand reductions; and pay exit fees covering remaining minimum charges on default [54] [55] [56]. The Commission’s stated purpose is to insulate ordinary ratepayers from the costs of infrastructure built for hyperscale growth [54]. Virginia has also enacted a datacenter electricity-consumption tax, and the SCC has separately ordered Dominion to directly assign to large-load customers the transmission upgrades driven solely by them [57].
The design is best read as a regulator writing down, in enforceable numbers, exactly which risks it fears: the 14-year term prices tenant-flight risk; the 85 percent minimum prices reservation-without-use risk; the collateral prices stranded-cost risk. Critics argue it does not go far enough — the Piedmont Environmental Council’s expert testimony found 61 percent of capital costs would remain unrecovered at the end of the contract term, and its president was blunt about the residual burden:
“The SCC is asking their constituents to continue to subsidize the energy needs.”
— Chris Miller, President, Piedmont Environmental Council [55]
For this paper’s argument, Virginia’s contribution is a negative theorem: energy abundance alone does not decide AI geography — rate design now co-decides it. A developer comparing a Virginia grid connection under GS-5’s minimum-payment and collateral regime against a Pennsylvania self-supply project under GRID, or a Texas behind-the-meter campus under Abbott’s audit standards, is no longer comparing electron prices. It is comparing institutional architectures for allocating risk between computing capital and the public. Ohio and Oregon have adopted kindred minimum-payment structures [56]; Georgia’s regulators have approved nearly ten gigawatts of capacity against a large-load pipeline [45]. The state laboratory is running the experiment in parallel across a dozen jurisdictions.
4.4 The Emerging State Bargain
Synthesizing the three models, the outline of a common settlement becomes visible. States are increasingly telling AI developers, in effect:
You can have the land. You can have the permits. You can build the campus. You may receive economic-development assistance.
But you must increasingly demonstrate:
| Demand | Pennsylvania (GRID / EO 2026-05) | Texas (Abbott directives / audit) | Virginia (GS-5 / SCC) |
| Your Power | Bring your own generation [3] | Add generation to the grid; on-site supply plans disclosed [23] [51] | 60% minimum generation-demand payments [54] |
| Your Infrastructure | Pay all costs of increased energy usage [3] | Fund full interconnection and infrastructure costs [29] | 85% minimum T&D payments; direct assignment [54] [57] |
| Your Water Plan | Strict water-conservation standards [4] | Closed-loop cooling; annual water reporting [51] | (Handled via local siting) |
| Your Community Benefit | Binding community-benefit agreements; local approval [3] | Community-impact standards; neighborhood protections [51] | Local land-use and tax instruments |
| Your Transparency | No NDAs; public permitting map [1] [3] | Ownership, incentive, and consumption disclosure under audit [52] | Registration; SCC oversight of interconnection process [56] |
| Your Financial Commitment | Legally binding GRID commitments [1] | Deposits and security under SB 6 framework [51] | 14-year contracts; $1.5M/MW collateral; exit fees [55] [56] |
The convergence across a Democratic governor, a Republican governor, and an independent regulatory commission — three institutions with no incentive to coordinate — is the strongest available evidence that the bargain is structural rather than partisan. It is the political counterpart of the corporate strategy this paper describes: as companies internalize their energy supply, states internalize their oversight of it.
4.5 From Energy Policy to AI Industrial Policy
The final subsection states the section’s theorem. Governors, legislators, and utility commissioners have become — mostly without intending it — participants in the Five-Layer AI Economy. State decisions about gas pipelines and permitting (Layer 1), utility rates and interconnection queues (Layer 1–3 interface), water allocations, tax exemptions, and transmission investment now materially influence where GPUs are installed (Layer 2–3), which models can be trained economically (Layer 4), and ultimately whose applications and agents reach the market at what cost (Layer 5). The federal layer compounds the point from the opposite direction: Executive Order 14318 of July 2025 directs federal agencies to accelerate permitting and financing for datacenters and the energy infrastructure powering them [58] — meaning American AI development now proceeds inside a two-level game in which Washington accelerates while states condition.
Therefore:
Energy policy is becoming AI policy — even when the legislation never mentions a model, an algorithm, or a semiconductor. A commissioner setting a minimum-demand charge is setting a floor under the cost of inference. A governor auditing an interconnection queue is allocating the national stock of trainable compute. A county board voting on a land-use permit is voting, at several removes, on the price of intelligence.
Section 5: From Hyperscaler to Energy Industrialist
5.1 The Changing Identity of the Technology Company
Assemble the record of the past eighteen months in one paragraph and its meaning becomes hard to miss. Microsoft has bound itself to an oil supermajor for twenty years of dedicated gas-fired power in the Permian [10]. Meta is financing seven gas plants, batteries, transmission, nuclear uprates, and 2.5 gigawatts of renewables through a state utility to feed a single campus [41] [42]. Google holds a right of first refusal over three gigawatts of engineered geothermal through 2033 and helped midwife the largest renewable-energy IPO in history [12] [35]. Amazon, Microsoft, Alphabet, and Meta collectively guided toward roughly three-quarters of a trillion dollars of 2026 capital expenditure, describing power and shells — not chips — as the assets they commit to earliest [26] [27]. And an AI infrastructure startup has proposed buying 1,800 mineral acres of the Marcellus outright [7].
Technology companies are not, in any formal sense, becoming oil-and-gas companies or utilities; they hold few operating licenses, employ few roughnecks, and mostly structure their energy exposure through subsidiaries, ventures, and long-dated contracts. But they are unmistakably becoming energy architects: entities that specify, finance, sequence, and increasingly co-own the generation, fuel, and delivery systems their computing requires. The identity shift is visible even in personnel and disclosure — heads of datacenter energy sitting beside CEOs at state announcements, earnings calls parsing useful lives, turbine queues, and power procurement with the fluency once reserved for chip roadmaps [26] [27].
Academic observers have been describing the underlying collision for two years. The director of the MIT Energy Initiative, opening the institute’s symposium on AI and energy, framed the stakes in economy-wide terms:
“We’re at a cusp of potentially gigantic change throughout the economy.”
— Professor William H. Green, Director, MIT Energy Initiative [16]
And Boston University’s Ayşe Coşkun, who has studied the grid–datacenter relationship for over a decade, locates the discontinuity precisely in the training era’s arrival — grids that once planned comfortably around datacenter growth were unprepared for what came after, because, as she puts it with disarming simplicity:
“AI consumes a lot of energy.”
— Professor Ayşe Coşkun, Boston University College of Engineering [59]
5.2 The Hyperscaler–Energy Company Convergence
The historical arrangement of the two industries was a one-way pipe:
Energy Company → Utility → Datacenter → Cloud Company
Each arrow was a market interface: the producer sold to the utility, the utility sold delivered power to the datacenter, the datacenter sold capacity to the cloud. Information, risk, and capital largely stopped at each boundary.
The emerging arrangement is a mesh of co-investment and mutual dependency:
Energy Company ↔ Hyperscaler ↔ Datacenter Developer ↔ Infrastructure Fund
Kilby is the template: Chevron supplies fuel and builds through Energy Forge One; Engine No. 1’s Joulent holds a 50 percent equity option; Microsoft anchors demand for twenty years; project finance intermediates [32]. UGI–Prime runs the same pattern at midstream scale — asset sales, retained storage, dedicated infrastructure, shared development [9]. Meta–Entergy runs it through a regulated utility, with the hyperscaler funding the rate base’s expansion [42]. Google–Fervo runs it through the capital markets themselves, with a framework agreement functioning as the collateral for an IPO [35] [48]. In each case, firms that once met only at a tariff now jointly design generation portfolios, share development risk, and co-sponsor infrastructure whose life span — twenty years to half a century — vastly exceeds any technology product cycle. That mismatch of clocks, a two-year chip cadence married to forty-year steel in the ground, is the defining managerial tension of the convergence, and the “commit early to power, decide late on chips” doctrine articulated across Q2 2026 earnings calls is the industry’s first systematic answer to it [27].
5.3 The Next AI M&A Frontier
If the convergence continues, transaction flow should move upstream — and 2026 offers the first confirmations. The strategic asset classes to watch:
- Pipeline operators and gas distributors, whose networks connect basins to campuses — the KKR bid for UGI, at a 21 percent premium on the very day of Pennsylvania’s regulatory tightening, is the signature datapoint [14] [34];
- Gas storage, the physical buffer that substitutes for grid flexibility (UGI’s retained 15 Bcf in the Prime deal shows storage being deliberately kept while land is sold [9]);
- Generation developers and independent power producers with turbine positions, since large frames are effectively sold out through 2030 [20];
- Geothermal companies, now IPO-viable on hyperscaler frameworks [35];
- Turbine and electrical-equipment manufacturers — GE Vernova, Caterpillar’s Solar Turbines, transformer and switchgear makers — whose order books have become the physical rate-limiter of the buildout [10];
- Transmission developers, wherever grid delivery remains the chosen path;
- Energy-rich land and mineral portfolios, the Alpha Compute category — arguably the earliest-stage and highest-optionality asset class of all [7] [8].
The pattern to anticipate is not necessarily hyperscalers buying oil companies; balance-sheet logic and regulatory scrutiny argue against it. It is infrastructure capital — KKR’s bid follows its acquisition of EDF’s North American power business earlier in 2026 [34] — consolidating the middle links of the Fuel-to-Compute Chain and leasing certainty upward to compute, while computing firms take targeted resource positions (mineral acres, pore space, frameworks, offtakes) at the ends of the chain. The datacenter transaction of 2030 may open with a reserve report and close with a service-level agreement.
5.4 The Carbon Contradiction
Subsurface Compute also creates one of AI’s largest policy tensions, and intellectual honesty requires treating it as a genuine dilemma rather than a talking point for either side.
The tension is arithmetic. The same technology companies that spent a decade building the corporate clean-energy market — and that still account for roughly 40 percent of corporate renewable PPAs [19] — now require firm power immediately, at gigawatt scale, in specific places. The fastest firm answer available in 2026 is natural gas, and the aggregate consequences are no longer hypothetical. A Bloomberg analysis published August 18, 2026, found that 99 proposed gas plants intended to power datacenters, drawn from BloombergNEF’s tracking of some 126 gigawatts of planned on-site capacity, would emit about 318 million metric tons of CO₂ annually at standard utilization — enough to lift total U.S. power-sector emissions by roughly 20 percent, and by as much as a third if run flat out [25]. The Environmental Integrity Project counts at least 74 datacenter-serving gas plants nationally, nearly half in Texas and twenty across the Ohio River Valley states including Pennsylvania [36]. The head of one utility watchdog compressed the industry’s revealed preference into a sentence:
“There is immense, immense pressure on the whole sector to get power.”
— David Pomerantz, Executive Director, Energy and Policy Institute [25]
Each escape route from the contradiction carries its own constraint, and a fair accounting lists them symmetrically. Natural gas delivers speed and firmness but locks in decades of emissions and pipeline infrastructure — mitigations such as brackish-water cooling and NOx controls at Kilby [30], or the pore-space rights in the Alpha Compute package that at least gesture toward future carbon storage [8], address local impacts more than the carbon ledger itself. Enhanced geothermal is firm and low-carbon but cannot yet be developed everywhere at comparable scale or speed; three gigawatts by 2033 is transformative for an industry and marginal against a 100-gigawatt gap [12] [20]. Nuclear, including the 45-gigawatt SMR pipeline and Fermi’s proposed AP1000s, offers firm low-carbon power on development timelines measured toward the 2030s [19] [39]. Renewables and batteries remain the volume leaders — the IEA expects them to meet half of global datacenter demand growth — but intermittency and transmission constraints cap their share of firm, sited load [18]. The debate, properly framed, is therefore not clean-versus-dirty but a five-dimensional trade among time, cost, reliability, geography, and carbon — with each project striking a different bargain and each state, per Section 4, writing different rules for what bargains are permissible. Whether behind-the-meter gas serves as the bridge its builders promise or the lock-in its critics fear will be decided less by intentions than by what gets financed next to it: the storage, geothermal, and nuclear capacity that could eventually displace the turbines, or merely more turbines.
5.5 The Five-Layer Consequence: Geology Can Set the Price of Intelligence
Bring the argument back to economics, because the deepest implication of Subsurface Compute is a proposition about prices.
Suppose one AI campus obtains predictable electricity at 5.85 cents per kilowatt-hour from owned gas beneath its feet, per the Alpha Compute evaluation [8], while a competitor pays the equivalent of roughly 15 cents under a scarcity-priced dedicated PPA [32], and a third floats on a congested wholesale market with capacity-price exposure. Electricity is among the largest recurring costs of inference and a first-order cost of training; a persistent two-to-three-fold spread in delivered power cost, compounded across millions of GPU-hours, propagates directly upward through the stack:
Cheaper Energy → Cheaper Compute → Cheaper Training and Inference → Cheaper Models → Cheaper AI Applications and Agents
The transmission is imperfect — chip efficiency, utilization, model architecture, and software all intervene — but the direction is unambiguous, and every layer above Layer 1 competes on margins thin enough for it to matter. Token prices are already an arena of ferocious competition; the producer whose marginal electron is cheapest and firmest can hold price points its rivals cannot, or reinvest the spread in the next training run. Over time, sustained energy advantage becomes model advantage becomes market advantage.
Eventually, therefore, a geological advantage becomes a software-economic advantage. This may be the paper’s single most important proposition, and it deserves italics:
The cost of artificial intelligence may increasingly be determined by resources that existed millions of years before the first transistor was invented.
The Marcellus was deposited roughly 390 million years ago; the heat beneath Utah’s Cape Station derives from the deep radiogenic and primordial energy of the crust. That these endowments should surface in the unit economics of a chatbot subscription in 2030 is the kind of joke history tells — and the kind of structural fact strategists ignore at their expense.
Section 6: What Have We Learned? Seven Pillars
The argument of this paper can be compressed into seven pillars — five structural, two cautionary.
Pillar 1 — Geology Is Becoming Digital Infrastructure
The first lesson of Subsurface Compute is that the boundary between physical and digital infrastructure is collapsing. Gas formations, geothermal reservoirs, pipelines, pore space, and energy-producing land can become upstream components of an AI system just as surely as GPUs and fiber networks — priced, financed, litigated, and acquired as such. When a Nasdaq-listed compute company files a term sheet for 1,800 mineral acres [8], and a national laboratory builds an AI digital twin of a geothermal reservoir on GPU infrastructure [13], the categories have merged in both directions. The cloud has a geology.
Pillar 2 — The Electricity Bottleneck Is Moving Corporate Strategy Upstream
When electricity can be obtained easily, technology companies purchase it. When electricity becomes difficult to obtain, they secure generation — Kilby, Hyperion, Matador. When generation becomes difficult to secure — turbines sold out, queues at 474 gigawatts [20] [23] — they move farther upstream still, toward fuel, pipelines, storage, frameworks, and resource ownership. Scarcity redraws the boundary of the firm, exactly as the theory of the firm predicts; AI has merely supplied the most dramatic input scarcity in modern industrial history. Each additional year of constraint pushes the frontier of control one link deeper into the Fuel-to-Compute Chain.
Pillar 3 — Energy-Rich Land Can Become Compute-Rich Land
The next major American datacenter hubs will not be determined exclusively by historical internet geography. Pennsylvania’s Marcellus tier, the Permian-to-Processor corridor of West Texas, Louisiana’s gas-and-grid cluster, and the geothermal fields of the interior West sketch an alternative map in which the question where can intelligence be manufactured most reliably? is answered, in part, from underground. Fiber made Ashburn; fuel is making Reeves County and Richland Parish. The valuation consequences — for rural land, for midstream assets, for utilities like UGI suddenly worth a 21 percent control premium [14] — are only beginning to be priced.
Pillar 4 — The Political License May Become as Important as the Mineral Right
Owning gas rights does not guarantee permission to build an AI campus. Developers must also obtain environmental permits, local approval, water, turbines, financing — and, as 2026 has made unmistakable, political legitimacy. Pennsylvania now conditions state review on local consent [3]; Texas conditions grid access on audit compliance [23]; Virginia conditions service on 14-year commitments and megawatt-scale collateral [55]. The election cycle demonstrates that voters have become participants in AI infrastructure development, through the most ordinary instruments of democracy: gubernatorial ballots, county land-use boards, and rate cases. Therefore the valuable asset is not simply the Mineral Right. It is:
Mineral Right + Power Right + Permit Right + Community Acceptance + Compute Demand
Absent any one term, the product of the expression approaches zero — a lesson at least one Texas developer, exiting after failing the audit standard, has already learned in public [52].
Pillar 5 — Layer One Can Determine the Economics of Layers Two Through Five
The Five-Layer AI Economy begins with Energy for a reason. Without adequate Layer 1 capacity, a Blackwell-class or future accelerator produces no useful computation; without compute, models cannot train economically; without models, applications and autonomous agents cannot scale. And because energy-cost advantages propagate upward with little attenuation, Layer 1 does not merely enable the upper layers — it can price them. The 5.85-cent electron and the 15-cent electron [8] [32] eventually confront each other in a token-price war neither chip nor model architecture can fully arbitrate. The intelligence economy is ultimately constrained — and increasingly priced — by the physical economy beneath it.
Pillar 6 — Verification, Not Announcement, Is the Scarce Commodity
A cautionary pillar. The record of 2026 is saturated with proposed gigawatts: a 474-gigawatt Texas queue against a 91-gigawatt peak [23], more than 100 Pennsylvania proposals against five fully permitted projects [1], phantom plants acknowledged even within BloombergNEF’s own tracking [25], and utility resource plans sized to demand that may arrive late, smaller, or never [45]. Pennsylvania’s anti-speculation provisions, Texas’s audits, and Virginia’s minimum-payment collateral are all, at bottom, verification technologies — institutional attempts to distinguish real load from optioned load before steel and ratepayer money are committed. Analysts and regulators alike should treat announced capacity the way geologists treat unproven reserves: as a resource category, not a fact. The 1970s nuclear overbuild, which left Washington State bondholders holding more than $2 billion in defaults [45], is the standing historical warning.
Pillar 7 — The Carbon Ledger Travels with the Molecule
The final pillar keeps the contradiction of Section 5.4 permanently in view. Subsurface Compute, in its 2026 configuration, is carbon-intensive at the margin: on-site gas capable of lifting U.S. power-sector emissions by a fifth is being planned in the name of an industry publicly committed to decarbonization [25] [36]. The same upstream integration that secures fuel also concentrates accountability — a company that owns the molecule owns its combustion products in a way a tariff customer never did. Pore-space acquisition [8], geothermal frameworks [12], nuclear pipelines [19], and matched renewable procurement [41] are the visible hedges; whether they mature faster than the turbine fleet locks in is the central open question of AI’s environmental economics. The pillar’s rule of thumb: every claim about Subsurface Compute’s speed should be footnoted with its ledger.
Conclusion: Intelligence Has a Basement
Artificial intelligence is often described as though it exists above the physical world. We speak of cloud computing, neural networks, frontier models, autonomous agents, and eventually superintelligence. The vocabulary points upward — toward abstraction, software, and cognition.
But the industrial system now being constructed to produce that intelligence is moving in the opposite direction.
It is moving downward.
Down through the server rack. Down through the electrical busway. Down through the substation. Down through the turbine. Down through the pipeline. And, increasingly, down into the geological formations beneath the land itself.
That is why I chose the title Subsurface Compute.
The Alpha Compute transaction in Pennsylvania makes the phrase more than a metaphor. When an AI infrastructure developer proposes acquiring datacenter land together with approximately 1,800 mineral acres of Marcellus gas rights and the pore space below them, the physical boundary of the AI economy changes: the resource beneath the datacenter becomes part of the datacenter’s strategic architecture [7] [8]. Chevron and Microsoft’s Project Kilby expands the idea to multi-gigawatt, twenty-year scale [10]. UGI and Prime Data Centers weld pipeline infrastructure and underground storage directly to an AI campus [9]. Meta’s Louisiana development binds one hyperscale project to an entire portfolio of new generating resources and a state’s transmission map [41] [42]. Google and Fervo demonstrate that the same logic can reach beneath the ground for heat instead of hydrocarbons — and that AI itself can be turned around to model the reservoirs that power it [12] [13]. KKR’s bid for UGI shows the capital markets pricing all of it in real time [14].
At the same time, politicians have discovered that the geology-to-compute relationship cannot be separated from the electorate. Pennsylvania’s August 18 policy shift, Texas’s audits and cost-responsibility directives, Virginia’s GS-5 architecture, and the campaign advertising now running in two of the nation’s largest states demonstrate that the race to construct AI infrastructure is producing a genuine negotiation among hyperscalers, utilities, energy companies, investors, communities, regulators, and voters [3] [6] [23] [54]. The 2026 midterms will not settle that negotiation; they will formalize it.
The most important implication extends beyond natural gas, beyond any single fuel, and beyond any single election.
Subsurface Compute is ultimately a theory of upstream integration. As artificial intelligence demands ever larger quantities of continuous energy, competition pushes technology companies progressively backward through the physical supply chain — from electricity contracts to generation, from generation to fuel infrastructure, from fuel infrastructure to resource ownership, and from conventional procurement toward new combinations of geothermal, nuclear, gas, storage, and other firm-power technologies. Each step backward is rational for the firm taking it; collectively, the steps are reorganizing the corporate structure of two of the world’s largest industries and redrawing the economic geography of a continent.
That transformation also changes the Five-Layer AI Economy itself. Layer One is no longer merely underneath the other four layers conceptually. It is increasingly underneath them literally. The shale beneath Pennsylvania, the gas infrastructure of West Texas, the engineered heat beneath Utah, the uprated reactors of Louisiana, and whatever firm resources are developed next will influence where the coming generation of AI factories is built, which corporations bring compute online fastest, which states capture the investment, which communities consent to host it, what it emits — and, ultimately, what it costs to manufacture intelligence.
For decades, the technology economy was organized around the proposition that information could escape geography. The AI economy is reminding us that energy cannot. And if artificial intelligence becomes one of the largest industrial consumers of electricity in modern history — the trajectory every serious forecast now describes [17] [18] — then the competitive search for computation will increasingly become a search for the physical resources that can sustain it.
That is the deeper meaning of Subsurface Compute:
Before there is an AI agent, there is a model; before the model, there is compute; before compute, there is electricity; and increasingly, before the electricity, there is the earth beneath our feet.
Footnotes / Endnotes:
[1] Commonwealth of Pennsylvania, Office of the Governor, “Governor Shapiro Signs Executive Order on Data Center Development in PA,” August 18, 2026. https://www.pa.gov/governor/newsroom/2026-press-releases/governor-shapiro-signs-executive-order-on-data-center-developmen
[2] WHP / NBC affiliate reporting, “PA governor signs executive order on AI data centers: ‘nation’s strictest guardrails’,” August 18, 2026. https://nbcmontana.com/news/nation-world/pa-governor-signs-executive-order-on-ai-data-centers-giving-local-communities-more-power-josh-shapiro-governors-responsible-infrastructure-development-standards-pennsylvania-pa
[3] Katie Bernard et al., The Philadelphia Inquirer, “Gov. Josh Shapiro signs executive order restricting data center development in Pennsylvania,” August 18, 2026. https://www.inquirer.com/politics/pennsylvania/josh-shapiro-data-center-order-20260818.html
[4] WESA (Pittsburgh NPR), “Pa. Gov. Shapiro signs executive order to ‘rein in’ data center development,” August 18, 2026. https://www.wesanews.org/politics-government/2026-08-18/shapiro-data-center-executive-order
[5] CNHI / Mesabi Tribune wire reporting, “Shapiro signs data center order barring fast-track permitting, NDAs,” August 18, 2026. https://www.mesabitribune.com/around_the_web/news/shapiro-signs-data-center-order-barring-fast-track-permitting-ndas/article_adb26ef9-3018-5f19-b59e-0dd107bf6729.html
[6] Marc Levy, Associated Press, “Governors’ races are being increasingly buffeted by the toxic politics of data centers,” August 18, 2026. https://www.news4jax.com/news/politics/2026/08/18/governors-races-are-being-increasingly-buffeted-by-the-toxic-politics-of-data-centers/
[7] Laila Kearney, Reuters, “Exclusive: Alpha Compute to buy Pennsylvania land, gas rights for $55 million data center campus,” August 11, 2026. https://finance.yahoo.com/technology/ai/articles/exclusive-alpha-compute-buy-pennsylvania-170027769.html
[8] Alpha Compute Corp., Form 6-K and announcement, “Alpha Compute Signs Binding Term Sheet for Planned 200 MW Natural Gas-Powered Data Center Campus,” August 11, 2026 (via StockTitan). https://www.stocktitan.net/sec-filings/ALP/6-k-alpha-compute-corp-current-report-foreign-issuer-2d502ec76b84.html
[9] Marcellus Drilling News / UGI Corporation announcement, “UGI Selling Property, Partnering with Prime to Build PA Data Center,” May 7, 2026. https://marcellusdrilling.com/2026/05/ugi-selling-property-partnering-with-prime-to-build-pa-data-center/
[10] Chevron Corporation, “Chevron Signs 20-Year Power Agreement with Microsoft for West Texas Data Center,” June 22, 2026. https://www.chevron.com/newsroom/2026/q2/chevron-signs-20-year-power-agreement-with-microsoft-for-west-texas-data-center
[11] Yahoo Finance / TechCrunch reporting, “Meta expands Louisiana Hyperion data center to 5 gigawatts,” July 13, 2026. https://finance.yahoo.com/technology/articles/meta-expands-louisiana-hyperion-data-121522809.html
[12] Congressional Research Service, “Enhanced Geothermal Systems (EGS) Commercialization,” In Focus IF13278, updated July 2026. https://www.congress.gov/crs-product/IF13278
[13] Pacific Northwest National Laboratory, “PNNL Teams Up with Fervo Energy and NVIDIA to Accelerate Geothermal Energy Development,” June 22, 2026. https://www.pnnl.gov/news-media/pnnl-teams-fervo-energy-and-nvidia-accelerate-geothermal-energy-development
[14] Katha Kalia, Reuters, “KKR makes $9 billion takeover bid for energy distributor UGI, WSJ reports,” August 18, 2026. https://www.aol.com/articles/kkr-makes-9-billion-takeover-153328000.html
[15] International Energy Agency, “AI is set to drive surging electricity demand from data centres while offering the potential to transform how the energy sector works” (Fatih Birol remarks on the Energy and AI report), April 2025. https://www.iea.org/news/ai-is-set-to-drive-surging-electricity-demand-from-data-centres-while-offering-the-potential-to-transform-how-the-energy-sector-works
[16] Leda Zimmerman, MIT News / MIT Energy Initiative, “Confronting the AI/energy conundrum” (MITEI Spring Symposium featuring Prof. William H. Green), July 2025. https://www.ll.mit.edu/news/confronting-aienergy-conundrum
[17] U.S. Department of Energy, “DOE Releases New Report Evaluating Increase in Electricity Demand from Data Centers” (Lawrence Berkeley National Laboratory, 2024 Report on U.S. Data Center Energy Use), December 20, 2024. https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase-electricity-demand-data-centers
[18] International Energy Agency, Energy and AI — Executive Summary, 2025. https://www.iea.org/reports/energy-and-ai/executive-summary
[19] Capacity Media, “AI data centres could triple electricity consumption by 2030, IEA warns” (IEA Energy and AI 2026 update; Birol remarks), April 17, 2026. https://capacityglobal.com/news/iea-ai-data-centres-energy-grid-concerns/
[20] Utility Dive, “AI data center growth could force US utilities to rethink generation plans, BofA says,” July 17, 2026. https://www.utilitydive.com/news/ai-data-center-growth-utilities-generation-plans/825541/
[21] UGI Corporation / Yahoo Finance, “UGI Energy Services and Prime Data Centers Forge Strategic Partnership to Power Data Center Campus,” May 6, 2026. https://finance.yahoo.com/sectors/energy/articles/ugi-energy-services-prime-data-201500707.html
[22] Daily Energy Insider, “UGI, Prime Data Centers partner on major Pennsylvania gas infrastructure project,” June 18, 2026. https://dailyenergyinsider.com/news/52211-ugi-prime-data-centers-partner-on-major-pennsylvania-gas-infrastructure-project/
[23] Sonal Patel, POWER Magazine, “Abbott Orders Full Audit of Texas Data Center Interconnection Queue, Threatens to Deny Grid Access,” August 2026. https://www.powermag.com/abbott-orders-full-audit-of-texas-data-center-interconnection-queue-threatens-to-deny-grid-access/
[24] Utility Dive (opinion/analysis citing BloombergNEF), “Behind-the-meter data center gas plants will raise US energy bills,” June 8, 2026. https://www.utilitydive.com/news/data-centers-raise-energy-bills-not-for-reason-you-think/822205/
[25] Bloomberg News via Fortune, “Data center gas plants to boost U.S. power emissions by 20%,” August 18, 2026. https://fortune.com/2026/08/18/data-center-gas-plants-to-boost-u-s-power-emissions-by-20/
[26] Brian Sozzi, Yahoo Finance (citing Goldman Sachs), “Meta, Microsoft, Amazon, and Alphabet are about to spend a shocking amount of money to dominate the AI era,” June 3, 2026. https://finance.yahoo.com/sectors/technology/article/meta-microsoft-amazon-and-alphabet-are-about-to-spend-a-shocking-amount-of-money-to-dominate-the-ai-era-115359575.html
[27] Uncover Alpha, “Amazon, Google, Microsoft, Meta Q2 earnings: The AI CapEx ROIC is bad thesis is DEAD,” August 2026. https://www.uncoveralpha.com/p/amazon-google-microsoft-meta-q2-earnings
[28] Yahoo Finance, “Amazon, Meta, and Microsoft stocks surge as AI hyperscalers post strong earnings results,” August 2026. https://finance.yahoo.com/technology/article/amazon-meta-and-microsoft-stocks-surge-as-ai-hyperscalers-post-strong-earnings-results-163729332.html
[29] Office of the Texas Governor, “Governor Abbott Directs PUC And ERCOT To Shield Texans From Data Center Infrastructure Costs,” June 10, 2026. https://gov.texas.gov/news/post/governor-abbott-directs-puc-and-ercot-to-shield-texans-from-data-center-infrastructure-costs
[30] BIC Magazine, “Project Kilby: Chevron and Microsoft Partner on 2.67 GW West Texas Power Plant,” June 24, 2026. https://www.bicmagazine.com/industry/powergen/chevron-signs-power-agreement-microsoft/
[31] Quartz, “Chevron signs 20-year natural gas deal with Microsoft for AI data center,” June 2026. https://qz.com/chevron-microsoft-natural-gas-data-center-west-texas-062226
[32] Global Data Center Hub, “Chevron and Microsoft Sign $9B West Texas Power-and-Compute Deal” (TD Securities estimates), July 2026. https://www.globaldatacenterhub.com/p/chevron-and-microsoft-sign-9b-west
[33] Jordan Blum, Fortune, “Microsoft and Chevron enter exclusivity deal on powering West Texas AI data center complex,” April 1, 2026. https://fortune.com/2026/04/01/microsoft-chevron-exclusivity-powering-west-texas-data-center-complex
[34] Pulse2, “KKR Reportedly Makes $9 Billion Takeover Bid For UGI At $42.50 Per Share,” August 18, 2026. https://pulse2.com/kkr-reportedly-makes-9-billion-takeover-bid-for-ugi-at-42-50-per-share/
[35] Fervo Energy, “Fervo Energy Reports First Quarter 2026 Results” (IPO, Cape Station, Google GFA), June 22, 2026 (via StockTitan). https://www.stocktitan.net/news/FRVO/fervo-energy-reports-first-quarter-2026-855vy8lpe8rk.html
[36] Spotlight PA / Environmental Integrity Project, “New power plants for data centers would worsen pollution,” July 2, 2026. https://www.spotlightpa.org/news/2026/07/pennsylvania-data-centers-emissions-gas-plants-climate-environment/
[37] EnergyNow / Bloomberg, “Microsoft and Chevron Sign 20-Year Power Deal For Texas Data Center” (BNEF state rankings; Gustavson remarks), June 2026. https://energynow.com/2026/06/microsoft-and-chevron-sign-20-year-power-deal-for-texas-data-center/
[38] Data Center Dynamics, “Fermi taps Hillcore to construct 2.6GW gas plant for up to 11GW Project Matador in Amarillo, Texas,” August 2026. https://www.datacenterdynamics.com/en/news/fermi-taps-hillcore-to-construct-26gw-gas-plant-for-up-to-11gw-project-matador-in-amarillo-texas/
[39] Fermi America / PR Newswire, “Fermi America and the State of Texas Announce Preliminary Approval for First 6 GW … on Project Matador’s 11 GW Private HyperGrid Campus,” November 4, 2025. https://www.prnewswire.com/news-releases/fermi-america-and-the-state-of-texas-announce-preliminary-approval-for-first-6-gw-of-one-of-the-worlds-largest-clean-natural-gas-facilities-on-project-matadors-11-gw-private-hypergrid-campus-302603582.html
[40] Fermi America, SEC filings excerpt via MarketBeat, “Project Matador — Advanced Energy and Intelligence Campus at Texas Tech University,” 2026. https://www.marketbeat.com/stocks/NASDAQ/FRMI/earnings
[41] ConstructConnect, “Meta Expands Louisiana Data Center to 5GW, Lifts Richland Parish Investment Above $50 Billion,” July 14, 2026. https://news.constructconnect.com/meta-expands-louisiana-data-center-to-5gw-lifts-richland-parish-investment-above-50-billion
[42] Engineering News-Record, “$27B Meta Data Center Pushes Louisiana Toward Massive Power Expansion,” April 2, 2026. https://www.enr.com/articles/62766-27b-meta-data-center-pushes-louisiana-toward-massive-power-expansion
[43] Daily Energy Insider, “Meta’s $50B AI expansion gives Entergy a massive power play in Louisiana,” July 15, 2026. https://dailyenergyinsider.com/news/53021-metas-50b-ai-expansion-gives-entergy-a-massive-power-play-in-louisiana/
[44] Straight Arrow News, “Meta fights order to disclose job, power figures for $50B Louisiana data center,” July 2026. https://san.com/cc/meta-fights-order-to-disclose-job-power-figures-for-50b-louisiana-data-center/
[45] RMI (Rocky Mountain Institute), “Planning for Uncertain Data Center Demand,” August 2026. https://rmi.org/resources/planning-for-uncertain-data-center-demand/
[46] Google, “Google and Fervo launch first-of-its-kind geothermal project,” corporate blog. https://blog.google/company-news/outreach-and-initiatives/sustainability/google-fervo-geothermal-energy-partnership/
[47] Seequent, “Case Study: Geothermal Energy 3D Models Powering Google’s Data Centers,” 2025. https://www.seequent.com/the-geothermal-innovation-powering-googles-data-centres/
[48] Latitude Media, “How Fervo plans to spend its $1.9-billion IPO,” May 13, 2026. https://www.latitudemedia.com/news/how-fervo-plans-to-spend-its-1-9-billion-ipo/
[49] Fervo Energy, “Fervo Energy and PNNL Leverage AI and NVIDIA Accelerated Computing for New Digital Twin Platform Designed to Advance Geothermal Development,” June 22, 2026. https://fervoenergy.com/fervo-energy-and-pnnl-leverage-ai-and-nvidia-accelerated-computing-for-new-digital-twin-platform-designed-to-advance-geothermal-development/
[50] Texas Policy Research, “Abbott’s Texas Data Center Directive,” June 10, 2026. https://www.texaspolicyresearch.com/abbotts-texas-data-center-directive/
[51] Matt Vincent, Data Center Frontier, “Texas Tightens Oversight of Data Center Development,” August 10, 2026. https://www.datacenterfrontier.com/hyperscale/article/55396202/texas-tightens-oversight-of-data-center-development
[52] The Center Square, “More data centers, companies, announce they will comply with Abbott directive,” August 2026. https://www.thecentersquare.com/texas/article_5fec315d-c1dd-4fdb-85be-1b92981e877a.html
[53] POWER Magazine, “More Data Center Operators Commit to Abbott’s Texas Standards as Power Companies Endorse ERCOT Batch Framework,” August 2026. https://www.powermag.com/more-data-center-operators-commit-to-abbotts-texas-standards-as-power-companies-endorse-ercot-batch-framework/
[54] Virginia State Corporation Commission, “SCC Issues Order on Dominion Energy Virginia Biennial Review 2025” (GS-5 rate class), November 25, 2025. https://www.scc.virginia.gov/about-the-scc/newsreleases/release/scc-issues-order-on-dev-biennial-review-2025/scc-rules-in-dev-biennial-review-case.html
[55] Charles Paullin, Inside Climate News, “Virginia Regulators Approve New Dominion Rates, Assign More Costs to Data Centers,” January 7, 2026. https://insideclimatenews.org/news/07012026/virginia-regulators-approve-new-dominion-rates/
[56] Dara Abasiita, Forbes, “Virginia Now Makes Data Centers Post $1.5 Million A Megawatt,” June 9, 2026. https://www.forbes.com/sites/daraabasiita/2026/06/09/virginia-now-makes-data-centers-post-15-million-a-megawatt/
[57] Data Center Dynamics, “Virginia regulators order Dominion Energy to directly assign transmission costs to data centers,” August 2026. https://www.datacenterdynamics.com/en/news/virginia-regulators-order-dominion-energy-to-directly-assign-transmission-costs-to-data-centers/
[58] Columbia University Sabin Center for Climate Change Law, “President Trump Orders Expedited Permitting for Data Centers” (Executive Order 14318), 2025. https://climate.law.columbia.edu/content/president-trump-orders-expedited-permitting-data-centers
[59] Jessica Colarossi, Boston University College of Engineering, “Is AI Slowing Climate Progress? It’s Complicated” (Prof. Ayşe Coşkun), August 6, 2025. https://www.bu.edu/eng/2025/08/06/is-ai-slowing-climate-progress-its-complicated
[60] Statista, “Big Tech’s AI Spending to Reach $760 Billion in 2026,” August 2026. https://www.statista.com/chart/35046/capital-expenditure-of-meta-alphabet-amazon-and-microsoft/
