In the past eighteen months, big tech has signed 13 nuclear deals committing over 9.8 GW of capacity, Microsoft locked in a $16 billion, 20-year agreement to restart Three Mile Island, Meta assembled a nuclear procurement book of up to 6.6 GW while contracting new gas plants in Louisiana, and the flagship Stargate data center in Shackelford County, Texas is being built entirely off-grid behind a 700 MW natural gas microgrid. The market keeps describing these moves as workarounds. They are not workarounds. They are the visible half of a permanent restructuring, and the invisible half, the reason it cannot reverse, is written in tariff filings that most technology executives have never read.
The grid stopped being a utility and became a counterparty
For a century, the American grid ran on one economic assumption: new load is good. Growth spread fixed costs across more kilowatt-hours, which lowered average rates, which meant every new factory made everyone's power cheaper. Load growth was welcomed, courted, and above all socialized: the infrastructure to serve it went into the shared rate base, because serving it benefited the pool.
AI load broke the assumption. A gigawatt campus does not spread fixed costs; it creates new ones, at marginal costs above average costs, concentrated in one county, on a timeline the grid cannot meet without emergency spending. The moment new load started raising rates instead of lowering them, a century of regulatory machinery reversed direction. The evidence is the price signal covered in full in Behind the Meter Goes Mainstream: PJM capacity prices ran from $28.92 per MW-day for 2024/25 to $269.92 the following year, then pinned the FERC-approved cap for three consecutive auctions while falling short of the reliability requirement, with annual capacity costs reaching roughly $16 billion. Somebody has to be assigned that bill, and the assignment process is the most underreported story in technology.
The cost will not be socialized. It will be assigned, and the addressee is the hyperscaler
Here is the part the market has not priced. The default assumption inside most capacity planning is that grid costs, however high, get spread across the rate base the way they always have, with data centers paying a large but proportionate share. The regulatory record of the last twelve months says otherwise, in plain language, in every major jurisdiction at once. The costs of the AI-era grid are being assigned, specifically and by name, to the loads that caused them, and the mechanism is the large-load tariff.
Read the record. In Virginia, the largest data center market on earth, the State Corporation Commission approved Dominion's GS-5 tariff: every customer at 25 MW or above is automatically enrolled starting January 1, 2027, required to sign contracts with minimum 14-year terms, pay for 85% of contracted transmission demand and 60% of contracted generation demand whether they use it or not, and post collateral of $1.5 million per megawatt. Oregon's POWER Act created a mandatory separate rate class for facilities at 20 MW and above, with 10-year minimum power purchase commitments and direct assignment of new transmission costs. Minnesota's HF 16 requires a "very large customer" rate class that carries all attributable costs of service. Ohio's approved AEP tariff puts an 85% minimum-take obligation on new data centers over 25 MW for up to 12 years, with exit fees equal to three years of minimum charges. Texas Senate Bill 6 goes further than money: it requires large loads to share the load-shed obligation, meaning that in a shortage, the data center is curtailed to protect residential customers. At least 18 states have introduced special rate-class legislation, Xcel is pursuing large-load tariffs in four states simultaneously under the explicit banner of ensuring data centers pay their own way, California and New York have opened formal proceedings on cost shifting, and the Edison Electric Institute, the utilities' own trade association, now maintains a public tracker of large-load tariffs approved or pending, describing their purpose as ensuring large customers pay their fair share.
Assemble what grid-served now actually means for a hyperscaler, because no single filing states it in one place: capacity purchased at the price cap, infrastructure costs directly assigned rather than socialized, 10-to-14-year take-or-pay commitments, collateral in the millions per megawatt, exit fees, and first position in the curtailment order when the grid runs short. That is not a power bill. That is a one-way counterparty exposure, drafted by the counterparty, enforceable by the state, with the hyperscaler as the designated residual payer of the entire energy transition. The political logic is airtight and permanent: residential ratepayers vote, affordability is the dominant utility-sector concern, and a customer worth three trillion dollars is the only party in any rate case that can absorb the bill without electoral consequence. This does not soften as the buildout matures. It compounds.
Grid-served is no longer a power bill. It is a one-way exposure to a counterparty that also writes the rules.
The rational response to being the residual payer is exit
Once the grid is understood as a counterparty offering one-way terms, the behavior of every sophisticated buyer stops looking like improvisation and starts looking like the only correct move on the board. The load that generates behind its own meter is the load no rate case can name, no capacity auction can bill, no minimum-take tariff can capture, and no load-shed order can curtail.
And the exit is already underway at scale, documented deal by deal. Oracle and OpenAI's Stargate site in Shackelford County runs on an on-site, behind-the-meter gas microgrid, built explicitly to bypass the strained grid and energize as soon as 2026. VoltaGrid, the microgrid provider behind it, has signed a 2.3 GW agreement with Oracle, energized xAI's Colossus campus in Memphis in months rather than years, committed over 1 GW across Vantage's North American campuses, and closed $5 billion of financing against a 4.3+ GW fully contracted deployment book. Crusoe secured multi-gigawatt turbine access through a joint venture with Engine No. 1 and Chevron. Bloom Energy's March 2026 industry survey found time-to-power running 1.5 to 2 years longer than buyers expected even recently, and CBRE reports the largest hubs, Northern Virginia and Dallas-Fort Worth, with declining under-construction capacity, citing power procurement itself as the obstacle. The interconnection queue analysis in The Queue Is the Product explains the physics of the delay; this article adds the ledger: even when the queue eventually delivers, what it delivers is the GS-5 terms above. Speed was the first reason to go behind the meter. The tariff record is the second, and the second one is permanent.
There is even a confession embedded in the new tariffs themselves: several, including Texas's framework, offer incentives such as reduced load guarantees and faster interconnection for customers that bring their own generation. The grid's own rulebook now pays large loads to self-supply. When the incumbent system starts subsidizing your exit, the argument about whether the exit is temporary has been settled by the incumbent.
Behind the meter is not a bridge. It is the silicon playbook, again
The standard objection is that on-site generation is a stopgap: endure it until the grid catches up, then return to normal service. The objection fails on the evidence, and it fails on precedent the hyperscalers themselves set.
The evidence first. Nobody signs a 20-year power purchase agreement to restart a nuclear reactor as a stopgap. Nobody assembles a 6.6 GW nuclear procurement book with delivery windows of 2032 to 2035, as Meta has across TerraPower, Oklo, Vistra, and Constellation, to cover a temporary queue problem. Amazon is investing over $20 billion converting the Susquehanna nuclear site into an AI campus, has taken a 1.92 GW PPA there, and put $700 million into X-energy's small modular reactors with a stated path toward 5 GW by 2039. Google ordered a 500 MW SMR fleet from Kairos with first units targeted for 2030. These are 20-and-30-year commitments to firm, self-directed generation, made by companies that model capital allocation more rigorously than any institutions on earth. The revealed preference is unanimous: power is being brought inside the strategic perimeter, permanently.
Now the precedent, because this movie has run before, inside these same companies. A decade ago, hyperscalers bought their processors from a market: merchant silicon, from vendors, on vendors' roadmaps. When compute became the strategic constraint, they did not wait for the market to improve. Google built the TPU, Amazon built Graviton and Trainium, Microsoft and Meta followed, and today custom silicon is a permanent, defining layer of hyperscaler strategy, run in-house at the design level and manufactured by specialists. Nobody calls the TPU a bridge until Intel catches up. Power has now crossed the same threshold silicon crossed: too strategic to rent from a market that cannot guarantee supply, price, or terms. Behind-the-meter gas is the first generation of that self-sufficiency, deployable in months; the nuclear book is the second generation, deliverable in decades; and the grid's remaining role, for the loads that keep a connection at all, is backup and negotiating leverage. The lifecycle mechanics of that stack, bridge fleets rolling into backstop and conversion roles as long-dated generation arrives, are mapped in Bridge to Permanent; the point here is simpler: every stage of the stack is self-directed, and none of it hands the terms back to the counterparty.
Unwilling is the operative word, and it has a resolution
One more fact completes the picture, and it is the most commercially important sentence in this article. Microsoft, the company with the largest single nuclear commitment in the set, has stated plainly that it has no intention of directly investing in its own plants or owning and operating generation assets. Hold both halves of the hyperscaler position side by side: power is now too strategic not to control, and generation is still a business they refuse to own and operate. Forced into generation; unwilling to become generators. Every nuclear PPA, every microgrid contract, every capacity reservation in this article is an attempt to resolve that contradiction, and the resolution has a shape: generation procured as a service, controlled by the buyer, owned by capital whose business is owning it, and operated by specialists whose business is running it. That is the same resolution the industry reached on chips, on buildings, and on every other layer of the stack, and the balance-sheet mechanics of it are the subject of From Capex to Opex. The companies being forced to become power companies will finish the transformation the way they finish every transformation: by defining the requirements, controlling the output, and paying someone else to hold the asset.
Step back and the whole arc resolves into three acts. Hyperscalers assumed they would plug into the grid the way every industry before them did; the grid answered with five-year queues and GS-5 terms. Nobody wants these companies to become utilities, least of all the companies themselves, and the one with the largest nuclear commitment has said so on the record. Then the capital markets compounded the problem: as From Capex to Opex documents, the spending curve is now moving stock prices across entire companies, which means even the buyers who could stomach the grid's terms can no longer simply buy their way out with balance sheet. The operational problem, source the equipment, deliver the campus, run the plant, and the financial problem, where all of it lands on the income statement, arrived at the same desk in the same quarter.
Markets resolve exactly this shape of problem the same way every time, and this industry has already watched it happen once. When hyperscalers could not build GPU capacity fast enough and would not carry all of it themselves, a specialist layer formed to source, finance, deploy, and operate compute at scale, and the industry gave it a name: neoclouds. Nobody sold the category into existence; the contradiction produced it. The same layer is now forming one level down the stack, underneath the rack, for the same reasons and on the same structural logic: specialists who source, deliver, and run the generation, and capital structures that determine how, and whether, any of it touches the customer's balance sheet. Power delivered as a service, on compute timelines.
It has a name too. Neopower.
The forward call stands, sharpened: by the end of the decade, the hyperscaler power stack looks exactly like the hyperscaler compute stack. Specified and controlled in-house. Sourced, financed, delivered, and operated by the neopower layer. And connected to a grid that has been demoted from sole supplier to backup generator, still mailing tariffs to loads that, increasingly, are no longer on the other end of the wire.