The Grid Said No: How AI Compute Is Being Forced to Become a Power Company
FERC's bring-your-own-power order, Meta's 1.6GW Crusoe deal, and Switzerland's nuclear reversal mark the moment AI infrastructure stopped waiting for the grid and started building around it.
A regulator’s quiet capitulation
The most consequential sentence written about AI infrastructure this week did not come from Jensen Huang, Sam Altman, or any of the customary oracles. It came from the U.S. Federal Energy Regulatory Commission, which signaled it would order grid operators to expedite AI data center interconnection applications — provided those projects either bring their own generation or agree to curtail usage during peak demand (Tom’s Hardware). The grid operators have ninety days to comply.
Read past the bureaucratic syntax and what FERC is saying is this: we cannot build transmission and generation fast enough to serve you on the terms you want. So we will change the terms. If you want a hyperscale campus connected this decade, stop asking us to find you 500 megawatts of firm power. Bring it yourself. Or accept that your $40 billion of GPUs will sit idle for hours of the year when the air conditioners come on in Houston.
This is the moment the structural question stopped being theoretical. For three years, every patient observer of the AI buildout has known that the binding constraint was not silicon, not packaging, not capital, not even cooling water — it was electrons delivered to a substrate of geographic locations where the latency, taxes, fiber, and political climate cooperated. The constraint was always going to bite. Now it is biting, and we are watching the industry’s response in real time.
That response is not, as one might have expected, a polite queue at the utility’s door. It is a forced vertical integration of the AI industry into the power generation business — a transformation whose closest historical analog is the late-19th century moment when industrial customers stopped buying steam from central plants and started building their own. The hyperscalers, neoclouds, and AI labs are about to become, by necessity, utilities. The question is whether they understand what that means.
How we got here: the half-decade the grid was supposed to do nothing
To see why this week’s news constitutes an inflection rather than another data point, it helps to remember the assumption that governed U.S. electricity planning for most of the 2010s. After the financial crisis, U.S. electricity demand was flat to declining. Efficiency gains from LED lighting, building codes, and Energy Star appliances outpaced load growth from population and economic activity. Utility long-range plans projected something like zero percent compound annual growth in electricity sales through 2030. Grid planners optimized for retirement of coal, integration of intermittent renewables, and resilience — not for new bulk load.
This worldview produced an interconnection process designed for a steady-state grid. Want to plug in a 200 MW factory? Submit a study request. The utility will model your load against the regional transmission organization’s existing flows, identify required upgrades, allocate costs, and queue you behind everyone else who applied first. Median wait times in PJM and ERCOT, the two RTOs most exposed to data center growth, stretched to four or five years. For a generation project, the math was worse: queue times of seven years and interconnection costs that frequently exceeded the cost of the generation itself.
Onto this somnolent system, hyperscalers and AI specialists dropped a demand profile no one had planned for. Lawrence Berkeley National Laboratory’s December 2024 update to the U.S. data center energy report concluded that data center electricity consumption could reach between 6.7% and 12% of total U.S. electricity by 2028, up from roughly 4.4% in 2023 — a tripling in five years against a baseline that had been assumed flat. The IEA’s January 2024 Electricity 2024 report estimated global data center electricity demand could more than double from 460 TWh in 2022 to over 1,000 TWh by 2026, with AI as the dominant marginal driver.
These projections, which seemed aggressive when published, now look conservative. The reason is that they were modeled on data center load, while the actual constraint has turned out to be data center contracted capacity — the firm megawatts the utility must reserve regardless of utilization. A 1 GW training cluster might average 600 MW of actual draw, but it requires the grid to plan around its 1 GW nameplate, because when synchronous training kicks off, the load step is near-instantaneous. The grid is not just supplying more energy; it is being asked to absorb shocks that look like nothing it was designed for.
By mid-2025, the queue dynamics had become farcical. ERCOT’s large-load interconnection backlog exceeded 100 GW — more than the entire installed generation capacity of Texas. PJM’s queue showed years of waiting for projects whose business cases assumed 18-month deployment cycles. In Northern Virginia, Dominion Energy effectively paused new substation commitments in Loudoun County. The polite fiction that hyperscalers would simply wait their turn collapsed.
What replaced it is what we are watching this week.
The current state: vertical integration, in pieces
The clearest signal of the new regime is Meta’s reported 1.6 GW capacity agreement with Crusoe Energy, covering data centers in Childress, Texas and Warrenton, Missouri (Data Center Dynamics). Crusoe began life as a flared-gas-to-bitcoin company, monetizing stranded hydrocarbons by converting them to compute. It has reinvented itself as an integrated developer of behind-the-meter AI campuses: site control, gas turbines, generators, switchgear, and a leased GPU fabric, sold to a hyperscaler as a turnkey contract. Meta is not buying power from a utility. It is buying infrastructure-as-a-service from a company whose core competency is moving electrons from gas molecules to GPU sockets without crossing a regulated wires interface.
Crusoe is one node in an emerging archipelago. Aston Power raised $20 million this week from TDK Ventures and JLL Spark to scale a “private power platform for data centers” (DCD). Verse raised $54 million Series B for a platform to expedite data center connections, developed in partnership with Calibrant Energy (DCD). A new three-building campus filing in Northumberland, UK proposes natural gas-powered fuel cells for “bridging power” — power that exists between the moment the building is energized and the moment the utility finishes the grid upgrade, which might never arrive (DCD). Equinix is trialing hydrogen fuel cells as a diesel replacement at an Irish data center (DCD).
Underneath all of this sits an industrial cliché made suddenly load-bearing: the battery. As the sponsored DCD opinion put it this week, BESS — battery energy storage systems — is becoming “the bridge between AI data centers and the grid” (DCD). The bridge metaphor is precise. Batteries do not generate; they translate. They translate a 200 MW grid interconnection into a 400 MW peak by buffering. They translate intermittent solar into 24-hour data center power. Most importantly, they translate the AI cluster’s spiky, training-batch-synchronous load profile into something the grid sees as smoothly inductive — a load profile interconnect studies can actually evaluate without requiring transmission upgrades.
And on the demand side, FERC’s order completes the picture. If you cannot bring power, agree to be curtailable. The grid will treat you as a flexible resource: when peak load threatens reserve margins, you throttle. For an AI training workload, throttling is a tolerable cost; checkpoint, pause, resume. For inference at SLA, less so — but the load can be geographically routed to clusters where capacity is uncongested. The data center, long modeled as an inflexible baseload customer, is being redefined as a partially dispatchable resource.
Far away from FERC, the Swiss parliament voted this week to lift the country’s ban on new nuclear power plants, reversing a post-Fukushima moratorium (Bluewin). Switzerland is not building a hyperscaler campus. But the political economy that produced its reversal is precisely the political economy now driving Microsoft to recommission Three Mile Island, Amazon to acquire Talen’s Susquehanna campus, and the U.S. Department of Energy to backstop the first orders of small modular reactors. A baseload generation source whose chief political problem was that nobody needed it has suddenly become indispensable, because the load it serves was indispensible first.
This is the new architecture, assembled in real time and in fragments: behind-the-meter generation (gas turbines, fuel cells, reciprocating engines, eventually nuclear), buffered by BESS, contracted directly between an industrial customer and a developer, with the grid retained only as a backup interconnect or as a curtailable secondary supply. The wires utility has been demoted from primary supplier to insurance policy. And the AI company has been promoted, against its preferences, into the role of vertically integrated electricity buyer-of-record.
Three framings, steelmanned
There are at least three coherent ways to read this transition, and getting the structural call right means taking each seriously before choosing.
Framing one: bridging. The optimistic reading holds that behind-the-meter and bring-your-own-power are temporary expedients. The grid will catch up. Transmission projects underway in MISO, SPP, and ERCOT will deliver bulk capacity by 2028-2030. Reconductoring of existing rights-of-way with high-temperature low-sag conductors can double effective transmission capacity without new lines. The Inflation Reduction Act’s transmission financing tools and the DOE’s Coordinated Interagency Transmission Authorizations and Permits Program will compress queue times. By the early 2030s, the FERC order will be a historical curiosity — the regulatory equivalent of wartime rationing, irrelevant once supply normalized.
In this reading, the Crusoes and Aston Powers of the world are filling a five-to-eight-year gap. They will exist on the margin as backup, as edge inference, as crypto-residual hybrid sites. The hyperscalers’ core training campuses will revert to traditional utility relationships, sweetened with PPAs for matching renewable generation but procured at the wholesale electric rate they have always paid. The vertical integration will unwind because, on a competitive cost-of-electrons basis, integrated utilities with scale economies, regulated rate base recovery, and grid-balancing optionality will always beat a single-customer 1.6 GW gas-and-batteries installation.
This framing has the virtue of historical pattern-matching. Industrial customers have repeatedly built behind-the-meter generation during supply crises — the 1970s oil shocks, the 2000 California crisis — and have repeatedly reverted to grid power when supply caught up. There is no obvious reason to believe AI demand has broken this pattern.
Framing two: structural inversion. The middle reading is darker. It holds that the FERC order is not a stopgap but a permanent reorganization of how large industrial loads are served in a transmission-constrained world. The reason is that “the grid will catch up” assumes a planning, permitting, and construction regime that has demonstrably broken. New high-voltage transmission lines in the U.S. take 10-15 years from initial conception to energization. The land use politics around new rights-of-way are at this point unwinnable in many corridors. Even Texas, which has the most permissive grid build environment in the developed world, is producing transmission projects whose timelines no longer match the doubling pace of GPU deployment.
If you believe the grid cannot catch up on a timescale relevant to GPU depreciation cycles — a credible belief — then bring-your-own-power becomes structural. Hyperscalers and neoclouds will permanently maintain in-house power development teams, dedicated gas turbine fleets, BESS operations, and eventually nuclear partnerships. The capex line item that used to be “electricity expense” will migrate to “generation infrastructure.” Companies like Crusoe become not bridge solutions but the permanent supply chain. The data center industry effectively forks the grid: a regulated wires-and-residential-load grid on one side, and a deregulated industrial-load-with-self-generation grid on the other, with batteries and FERC-mandated curtailment as the interface.
This framing has the virtue of taking the grid’s actual construction velocity at face value. It is also consistent with what utilities themselves are doing. The major investor-owned utilities have shifted their capex toward distribution hardening and renewable interconnection, not toward the kind of massive bulk transmission that would let them serve hyperscaler load. They have read FERC’s tea leaves. The signal is unmistakable: they have been let off the hook for being the only solution.
Framing three: capex reckoning. The pessimistic reading is that the vertical integration is happening because the underlying economics of the AI buildout were already unsustainable, and forcing AI companies to internalize power costs will accelerate the moment of reckoning. The hyperscalers’ 2025-26 capex plans, now exceeding $400 billion annually in aggregate, were modeled on a particular set of unit economics: GPU cost, power cost, cooling cost, depreciation schedule, revenue per token. If you change one input — say, by forcing operators to pay the full marginal cost of bringing 1.6 GW of dedicated generation online instead of the average cost of utility power — you change the revenue-per-token threshold below which the entire build does not pencil.
In this reading, the FERC order will produce two effects within 24 months. First, the marginal AI training campus will be cancelled or quietly scoped down, as the full-loaded cost of behind-the-meter power exceeds what utility power was projected to cost. Second, the inference economics will get squeezed, because inference clusters cannot tolerate the same curtailment that training can, which means they bear a higher share of firm-power costs. The result is that the structural answer to AI’s power problem turns out to be: there is less AI than we expected, because the marginal customer cannot afford the marginal kilowatt-hour.
I find this framing easier to dismiss in aggregate than in particulars. In aggregate, frontier-lab demand and big-three hyperscaler demand will not be deterred by power cost increases of any magnitude that’s actually on offer — even doubling the all-in cost of electricity at the meter only moves a frontier training run from, say, $80 million to $120 million of energy in a $400 million budget. But in particulars, at the margin where neoclouds compete on $/GPU-hour for inference and fine-tuning workloads, a power cost increase of 30-50% is structurally fatal to many business plans. The capex reckoning, if it comes, will look like a consolidation of the neocloud sector, not an unwinding of the hyperscaler buildout.
First principles: why power, why now, why this shape
Strip away the news flow and the underlying drivers are physical and political.
The physical driver is that AI training and large-scale inference are the first computing workloads where the bottleneck has migrated from compute to power delivery rather than power cost. For all of computing’s history, the economic problem with electricity was its price. CPUs and even early GPUs were power-constrained only at the limits of thermal design. A data center could be sited wherever fiber, taxes, and weather aligned; power was an input variable, not a constraint.
That regime ended somewhere between the H100 generation and the B200 generation. A modern AI training rack draws 120-140 kW; full Blackwell pods exceed 1 MW per rack-row; a 100,000-GPU campus is a 150 MW continuous load with another 60-80 MW for cooling. These numbers do not exceed any utility’s theoretical capacity. They exceed utilities’ ability to deliver them to a specific physical location within the time horizon of a competitive AI training run. The fundamental physical fact is that 200 MW is now a load that requires its own substation, dedicated transmission interconnection, and frequently dedicated generation. It is no longer something you order from the utility like a pizza.
The political driver is that the people who live near the substation and the transmission line increasingly do not want the substation or the transmission line. The Amazon workers testifying against AI data center expansion this week — and being allegedly intimidated by their employer for doing so (Tom’s Hardware; CNBC via Hacker News) — are a leading indicator of broader political resistance. The Digi Power X lawsuit over noise from a New York crypto mining site is a symptom of the same dynamic (DCD). Behind-the-meter generation is not just an engineering response to grid queue times; it is also a political response. A site that doesn’t need transmission upgrades doesn’t need transmission permitting hearings.
These two drivers compose into a third, derived principle: the value of a unit of compute is becoming increasingly determined by the value of a unit of power-delivered-to-the-right-place. NVIDIA’s pricing power, AMD’s catch-up, Amazon’s reported willingness to sell Trainium to outside data centers (DCD) — all of this discussion about silicon is happening on top of an unstated assumption that the silicon will find a socket somewhere with power and cooling. As that assumption weakens, the silicon’s value becomes path-dependent on the power infrastructure surrounding it. A 1 GW deployment slot near a gas reservoir with permitting clarity is becoming more valuable than the GPUs that will fill it. This is the inversion that defines the new regime: in the AI buildout’s third act, the scarce resource is not chips. It is sited, energized, cooled megawatts.
Once you see this, the more exotic news items in this week’s feed stop looking like fringe oddities and start looking like rational adaptations to the constraint. Crusoe’s flared-gas model arbitrages a power source nobody else can monetize. Hyperscale Data’s plan to deploy humanoid robots inside its Michigan facility (DCD) is, in part, an answer to data center labor scarcity in remote behind-the-meter locations. California startup Orbital’s $5 million pre-seed to put Blackwells in orbit (DCD) is preposterous on cost grounds today but is a rational long-tail bet if you assume terrestrial power-delivery costs continue to inflate while solar panels in geosynchronous orbit do not need a substation interconnect study. Even the local-AI movement — the Tom’s Hardware writer running mini PCs to process millions of tokens daily on home wattage (Tom’s Hardware) — is best understood as edge-of-grid demand arbitrage against centralized AI power costs that have started rising faster than Moore’s Law can offset.
What it means: a partial taxonomy
If the structural inversion framing is roughly correct — that bring-your-own-power is permanent rather than transitional — several consequences follow that are worth naming explicitly.
First, the AI infrastructure capex line will start to look less like the cloud capex of 2015 and more like the upstream oil and gas capex of 1975. Long-cycle, capital-intensive, physically located, with significant project-finance components, joint ventures, and offtake agreements. The companies optimized for cloud-era capex deployment — engineering org structures, financial reporting, real estate functions — will need to absorb power-generation expertise that does not currently sit on their org charts. Expect a transfer of personnel from the regulated utility sector and from independent power producers into hyperscalers and neoclouds.
Second, the geography of AI compute will continue to shift toward locations where this kind of integration is politically and physically feasible. The Texas Panhandle. The North Dakota Bakken. Western Pennsylvania over Marcellus. The Permian. The Saskatchewan oil patch. The Norwegian fjords. The Gulf Cooperation Council where sovereign producers will be only too happy to bundle compute with their hydrocarbons. The legacy data center alleys — Northern Virginia, Santa Clara, Singapore — will continue to serve latency-sensitive inference but will lose their share of training compute as power constraints bind harder there than in resource-extraction regions. The Buzz HPC / Bell AI / Cohere $220 million Canadian cloud deal (DCD) is a small early data point in this geographic redistribution.
Third, the regulatory regime will continue to bifurcate. FERC’s order acknowledges, in regulatory politeness, that the agency has lost the ability to plan for AI load on the timescale the AI industry needs. State public utility commissions will increasingly find themselves regulating residential and small-commercial customers on one track while the bulk industrial AI load operates on a deregulated, contract-by-contract track. This is a serious public-policy problem because it means residential customers risk being stranded with the fixed costs of a grid built for a customer mix that no longer exists. The rate cases of 2027-29 will be where this fight gets adjudicated, and the politics will be ugly.
Fourth, the nuclear question becomes inevitable. Behind-the-meter gas is the bridge. Batteries are the buffer. But the only generation source that is both dispatchable, low-carbon, high-capacity-factor, and physically siteable next to a hyperscale campus is nuclear — small modular reactors first, then potentially conventional plants if the regulatory path can be shortened. The Swiss parliament vote, the Microsoft-Constellation TMI deal, the Amazon-Talen Susquehanna acquisition, the DOE’s SMR backstops, and now Jefferson Lab’s $49 million data center build (DCD) — these are nodes of a single emerging story in which national governments, AI companies, and regulators converge on nuclear as the only path that doesn’t require either accepting indefinite gas dependence or accepting that AI scaling stops at the boundary of intermittent renewables.
Fifth, the FERC curtailment provision deserves more attention than it is getting. If hyperscalers and neoclouds accept curtailability in exchange for queue priority, they implicitly accept a new operational constraint: training and inference scheduling must include grid-state as an input. This is technically tractable — checkpoint training during peak hours, run batch inference jobs at night, geographically distribute fine-tuning workloads to wherever the grid has slack — but it requires software infrastructure that nobody has built at scale yet. The first hyperscaler to operationalize grid-aware scheduling across its global fleet will have a meaningful cost advantage over peers who do not. This is a software problem of the same flavor as carbon-aware scheduling, but with binding economic teeth rather than ESG aesthetics.
What to watch
The leading indicators over the next two to four quarters are concrete and trackable.
Watch the interconnection queue clearance rates in ERCOT, PJM, and MISO after FERC’s ninety-day implementation window closes. If queues clear meaningfully — measured in projects energized, not just projects approved — the bridging framing has merit. If they don’t, the structural-inversion framing wins by default.
Watch hyperscaler 10-K disclosures in the next reporting cycle for “energy and infrastructure” capex breakouts. Microsoft, Amazon, and Meta have begun to separate power-related capex from data center capex in segment reporting; the rate at which the former grows relative to the latter will tell you how much vertical integration is actually happening on the balance sheet versus being announced for press releases.
Watch the neocloud earnings — CoreWeave, Crusoe (if it goes public), Nebius, and the Bell AI / Buzz HPC / Cohere cohort — for gross margin compression in inference services. If gross margins on per-GPU-hour pricing compress more than 200-300 basis points in the next two quarters, the capex-reckoning framing is starting to bite.
Watch FERC and state PUC dockets for the rate-design fights that will emerge over who pays for the grid that AI is leaving behind. The first major U.S. utility to file for a residential rate increase explicitly attributable to data center grid investments will be a tell. So will the first major utility to write down stranded grid assets attributable to load that never materialized because the customer went behind-the-meter instead.
Watch the SMR order books, because the nuclear path is the long-run resolution of the constraint. Real orders — not letters of intent — placed in the next 18 months will determine whether the late 2030s look like a return to abundant clean baseload power or like a continuation of the gas-and-batteries patchwork.
Finally, watch whether the AI labs and hyperscalers begin to publicly disclose load-flexibility commitments and grid-services revenue. The transition from baseload customer to flexible resource is conceptually huge, but the financial disclosure tooling does not yet exist. When it does, that’s the moment the new regime has become permanent enough to require accounting standards.
A closing observation
The framing that has dominated AI infrastructure commentary for two years — that the buildout is a story about chips — has always been incomplete. Chips are the most visible variable because they are designed and sold by a small number of legible companies. But chips have always been downstream of power, fab capacity, packaging, and political license. The story that the second half of the 2020s is going to tell is the story of those upstream constraints reasserting themselves, one by one, as the binding factor.
Power is the first to bind in a way that the industry cannot wave away with capex. FERC’s order this week is the regulator publicly acknowledging that the binding has happened and that the workaround will be vertical integration whether anyone wanted that or not. The hyperscalers will become power companies, the neoclouds will become co-located behind-the-meter developers, the batteries will hide the spike, the gas turbines will hum behind the fence, and somewhere in the late 2020s the first SMRs will come online next to the first data centers built to be permanently tethered to them.
This is not a story that ends with AI being limited by power. It is a story that ends with AI restructuring the electricity industry to fit the shape AI requires. The grid did not say no. The grid said: not on the terms you offered. So the terms changed. The grid said: bring your own power. So AI is becoming a power company. The next several years are about what happens to an industry, to a regulatory system, and to a built environment when the largest source of marginal industrial load in the modern era stops accepting the slow answer and starts buying its way past the queue.
The bill for that is going to be enormous, and the bill is going to be visible, and the bill is going to land somewhere — on shareholders, on residential ratepayers, on the climate, on local communities, on the marginal AI workload that no longer pencils. The structural question for the rest of the decade is not whether AI scales. It is who pays for the power that lets it scale, and what the political settlement looks like when the answer becomes legible.
We will know more by year-end. The ninety days FERC has given grid operators expire in mid-September. Whatever those grid operators file will be the first concrete shape of the new regime. Watch the dockets.