AI capex met the grid: when the megawatt curve breaks

The question behind the headline #

The four largest hyperscalers (Microsoft, Alphabet, Meta, Amazon) guided combined capex above 360 billion dollars for 2025 and roughly 420 to 470 billion dollars for 2026. Add NVIDIA's own infrastructure spend, Oracle's OCI build out, and the rapidly capitalized neocloud tier (CoreWeave, Crusoe, Nebius, Lambda, Nscale, Applied Digital), and the AI capex envelope for 2025 to 2026 clears 500 billion dollars. Most of that money buys silicon, integration, and buildings.

None of it generates an electron. The chips have to plug into something, and the something is the slowest moving asset in the AI stack. The average new generator interconnection request in the United States now waits more than 4 years for energization, according to the 2024 Lawrence Berkeley National Lab queue report. The PJM 2025 to 2026 capacity auction cleared at 269.92 dollars per megawatt-day, an order of magnitude above the prior decade's average. The Northern Virginia data center cluster (Loudoun and Prince William counties) has booked Dominion Energy load growth requests that exceed the utility's entire 2018 peak.

The question is not whether AI workloads can pay for power. They can, easily. The question is whether the megawatts arrive on the calendar that the capex assumed. This brief works through where the constraint binds first, what the candidate solutions actually deliver, and what the gap means for AI investment ROI.

What the queue data actually says #

Lawrence Berkeley National Lab's annual queue report is the cleanest single source. The 2024 release covered all seven US ISO/RTO regions plus the non-ISO utilities. Active capacity in the queue at the end of 2023 was 2,597 gigawatts of generation and 1,028 gigawatts of storage. For context, total US installed generating capacity is roughly 1,300 gigawatts. The queue is twice the existing fleet.

Headline queue size is misleading on its own because the historical completion rate is low. Across 2000 to 2018, only 19 percent of generation projects that entered an ISO queue ultimately reached commercial operation. The 2024 report shows that completion rate has fallen further for the 2014 to 2018 cohort. The pipeline is fat. The throughput is not.

The other axis is time. Median time from interconnection request to commercial operation rose from 2.1 years for projects completed in 2008 to 5.0 years for projects completed in 2023. The PJM and MISO queues run longer than the median. CAISO and ERCOT run shorter.

Where the binding constraint actually bites #

Queue averages hide the location problem. AI training and inference clusters are not distributed evenly across the country. Six markets carry the bulk of the announced 2025 to 2028 hyperscale capacity: Northern Virginia, Phoenix and Maricopa County, the Atlanta corridor, central Washington (Quincy and Moses Lake), Dallas Fort Worth, and ERCOT West Texas. Each one faces a different binding constraint.

Northern Virginia: transmission, not generation. Loudoun County alone hosts above 4,500 megawatts of operating data center load and another 7,000 to 9,000 megawatts in active development through 2028. Dominion's most recent integrated resource plan filed with the Virginia State Corporation Commission shows peak load growing from 22.5 gigawatts in 2024 to roughly 39 gigawatts by 2039. The constraint is the high voltage backbone. Two greenfield 500 kilovolt projects are needed to land the new generation in the data center alley. Both have permitting timelines that push first energization to 2028 at the earliest.

Phoenix and Maricopa: heat, water, and Salt River Project tariff posture. APS and SRP have both moved to large load tariffs that require minimum 4 to 5 year lead times and capacity reservation deposits. The 2024 to 2025 SRP rate case introduced a separate Extra High Load Factor schedule with explicit data center provisions.

Atlanta corridor: Georgia Power is the constraint. The utility's 2025 IRP update added 8,200 megawatts of new resource need by 2031, almost entirely driven by data center and EV manufacturing load. The PSC approved an accelerated procurement path, but new build natural gas generation realistically energizes 2027 to 2029.

Quincy Washington and central Washington: hydroelectric is sold out. Grant County PUD, Chelan County PUD, and Douglas County PUD have all moved to closed or constrained allocation models for new large loads. The cheap hydro power that drew Microsoft, Yahoo, Sabey, and Vantage to Quincy starting in 2007 is no longer available at the margin.

ERCOT West Texas: generation is plentiful. Transmission is the constraint. The Permian Basin Reliability Plan working through PUCT and ERCOT contemplates roughly 13 to 15 billion dollars of new 765 kilovolt transmission to move West Texas wind, solar, and gas to the load centers. First in service is targeted for 2030 at the earliest.

Dallas Fort Worth: distribution capacity within Oncor's footprint is now the visible constraint, with multi-year wait lists for new substations above 200 megawatts.

Announced data center capacity by region #

The announced pipeline is not the same as the energized pipeline. The table below is the announced 2025 to 2028 nameplate, drawn from utility integrated resource plans, regional planning organization filings, and county economic development announcements.

The contrast between announcement and delivery is the heart of the gap. Northern Virginia and Phoenix between them have above 27 gigawatts of announced load through 2028. Realistic energized load over the same window, based on signed Engineering Procurement Construction (EPC) contracts and confirmed transmission scope, is closer to 11 to 13 gigawatts.

PPA prices: the second order signal #

Power purchase agreement prices are the cleanest market signal of how tight the constraint has become. LevelTen Energy's quarterly PPA Price Index covers North America wind, solar, and storage. The trend since 2022 is monotonic. Prices are up across every region, every technology, every contract length.

The bilateral data center PPA market sits well above the LevelTen index for two reasons. First, hyperscalers are taking the entire output of new projects under additionality requirements that limit the qualifying universe. Second, hourly carbon free energy commitments require firmed shaped product, not vanilla solar or wind, which clears at a meaningful premium.

Behind the meter: the 18 month bridge #

Behind the meter generation, on site at the data center campus and not requiring transmission interconnection, is the single most popular bridge solution in the 2025 to 2027 window. The economics are straightforward. A reciprocating engine or aeroderivative gas turbine plant of 100 to 400 megawatts can be designed, permitted, and energized in 14 to 22 months, against 4 to 7 years for grid connected generation in the same footprint.

The catch is air permitting. Title V major source thresholds and PSD review push permitting timelines past 24 months in non attainment areas. The clean alternatives (fuel cells, on site solar plus storage) carry energy density limits that cap them at single digit percent of the load on a typical hyperscale campus. Bloom Energy's solid oxide fuel cell deployments at AWS Oregon and Equinix Silicon Valley sites are the most visible examples, but each is sub 50 megawatts.

Stargate and the 1 gigawatt scale projects coming online in Abilene Texas and central Wisconsin in 2026 lean heavily on behind the meter natural gas with optionality for hydrogen blending. The Abilene site uses 18 GE LM6000 aeroderivative units in a configuration that can island the campus from ERCOT for extended periods. This is not a sustainability play. It is a calendar play.

The economic logic: a hyperscaler paying 5 to 7 cents per kilowatt-hour all in for behind the meter gas, against 9 to 14 cents for landed grid power including transmission and distribution charges in PJM or Dominion territory, captures both the cost differential and the schedule advantage. The reputational and ESG cost is the offset. Most hyperscalers are choosing to pay the reputational cost in 2025 to 2027 and bridge to cleaner firmed solutions later.

Small modular reactors: real roadmap, late delivery #

The SMR conversation went from speculative to capitalized over 2023 to 2025. Amazon, Google, Microsoft, Oracle, and Meta have each signed letters of intent or anchor offtake agreements with SMR developers. The economics on paper are competitive. The schedule is the open question.

TerraPower's Natrium project at Kemmerer Wyoming targets first power in 2030, with the 2024 fuel supply delays now in the rear view. X-energy's Xe-100 design with Amazon and Energy Northwest in central Washington targets 2032 to 2033. NuScale's domestic deployments effectively reset after the UAMPS cancellation in 2023, with current activity focused on overseas deployments and a new US site selection process. Oklo's microreactor design at the Idaho National Laboratory is targeting 2027 to 2028, but at single unit scale of 15 to 50 megawatts.

Realistic SMR contribution to the 2025 to 2030 megawatt gap is therefore zero to 500 megawatts. Real contribution begins 2030 to 2032 and accelerates 2033 onward. The honest framing is that SMRs solve the 2030s gap, not the 2025 to 2028 gap that AI capex is racing into. PPA pricing for early SMR offtake reflects this. Microsoft's Three Mile Island Unit 1 restart agreement with Constellation, announced in September 2024, is for 835 megawatts of restart capacity (not new SMR) at a 20 year contract length, and even this restart targets 2028 first power.

Energization timeline scenarios for hyperscale data centers #

Combining the queue data, the regional binding constraints, and the PPA market gives a structured view of how long an announced megawatt actually takes to energize, by procurement path and region.

The fastest path in any region is behind the meter gas with no transmission upgrade requirement. The slowest path is greenfield generation that requires both new generator interconnection and new transmission backbone. Most real projects sit somewhere in between.

What the megawatt constraint means for AI capex ROI #

The capex side of the AI economics equation has been studied to death. Chip costs, depreciation schedules, utilization curves, cloud margins. The opex side is now the variable that moves the answer. Three specific places the megawatt constraint reshapes the ROI math.

First, time to revenue. A GB200 NVL72 rack purchased and shipped in early 2026 that cannot energize until late 2027 because the Northern Virginia substation is delayed loses 18 to 22 months of useful life against a 36 to 48 month depreciation schedule. That is not a small number. On a 4 million dollar rack with 60 percent utilization assumptions, the schedule slip alone destroys 35 to 45 percent of the cumulative gross margin.

Second, location optionality value. The hyperscaler that can route training workloads to whichever campus has marginal megawatts available captures option value the single site operator cannot. This favors the four largest cloud platforms over the neocloud tier and over enterprise on premises. Workload portability across regions, including across countries (the European, Middle Eastern, and Latin American buildouts now matter for capacity reasons rather than latency reasons), becomes a first order strategic capability.

Third, the cost of power moves from a nuisance line item to a binding economic input. A frontier training run that consumes 50 megawatts continuously for 90 days uses roughly 108,000 megawatt-hours. At 4 cents per kilowatt-hour (cheap industrial power circa 2022), that is 4.3 million dollars. At 9 cents per kilowatt-hour (Northern Virginia behind the meter gas plus transmission and distribution circa 2026), it is 9.7 million dollars. At 14 cents per kilowatt-hour (PJM peak with capacity charge pass through circa 2026), it is 15.1 million dollars. Power is now 8 to 18 percent of the all in cost of a frontier training run, against 2 to 4 percent in 2022.

PJM as the canary #

PJM Interconnection covers 65 million customers across 13 states and DC, including the Northern Virginia data center alley. The PJM 2025 to 2026 Base Residual Auction cleared at 269.92 dollars per megawatt-day, against an average clearing price of 28.92 dollars per megawatt-day across the prior 10 auctions. The 2026 to 2027 auction settled in mid 2025 at 329.17 dollars per megawatt-day. Capacity prices are now an order of magnitude above the prior decade norm.

The cause is structural. Coal and older gas retirements are accelerating ahead of new generation interconnection. The capacity emergency reserve target was breached on the demand side by data center load growth that PJM's own forecasts repeatedly underestimated. Federal Energy Regulatory Commission filings through 2024 and 2025 show PJM revising its 15 year load growth forecast upward four times.

The downstream consequence for AI economics: PJM territory data center power costs include capacity charge pass through that has roughly tripled since 2023. A 100 megawatt data center in Loudoun County now pays 8 to 12 million dollars more per year in capacity related charges than it did three years ago. That cost flows directly into hyperscaler unit economics on AI workloads located there.

Where the math gets soft #

Three places the analysis above can move materially.

Demand response and flexibility. Data center load has historically been treated as inflexible baseload. AI training is not. A 60 minute pause for a transmission contingency event is operationally trivial for a pretraining run, and economically trivial against the value of grid flexibility services. If the industry signs onto curtailable tariffs at scale, both the queue and the capacity prices look different. Early indicators from ERCOT's 4 Coincident Peak program and PJM's Capacity Performance product suggest 5 to 15 percent of hyperscale load could be flexed without material workload impact.

Permitting reform. The Energy Permitting Reform Act of 2024 narrowed but did not pass. State level permitting reforms in Texas, Pennsylvania, and Virginia in 2025 have shaved schedule on specific projects. Federal action that materially compresses NEPA review on transmission could pull 12 to 18 months out of the worst case scenarios above.

Utility willingness to take on data center cost recovery risk. The Ohio AEP rate case in 2024, the Virginia SCC's 2025 large load tariff proceeding, and the Indiana IURC's 2025 ruling on minimum demand contracts are reshaping who bears the stranded cost risk if announced data center load fails to materialize. The trend is toward shifting more risk to the data center customer through long minimum demand contracts and exit fees. This is the right direction, but it raises the cost of optionality and may slow announcement velocity.

What enterprises and investors should plan for #

Five working assumptions hold for the next 36 months of AI capex planning.

One. The binding constraint on US AI capacity through 2028 is megawatts on a calendar, not chips, not capital, not policy. Investment theses that assume the chip pipeline drives the deployment curve are incomplete.

Two. The four hyperscalers will continue to outbid the neocloud tier for the scarce megawatts. This is a structural advantage that compounds as power gets tighter. Neocloud growth depends on either secondary markets that hyperscalers exit, or geographies the hyperscalers cannot reach quickly.

Three. Behind the meter gas is the dominant 2025 to 2028 bridge solution. Treat ESG narratives that exclude this honestly. The math forces it.

Four. SMRs solve the 2030s problem, not the late 2020s problem. PPA portfolios should be built on a wind, solar, storage, restarted nuclear, and new gas backbone for the next 60 months, with SMR optionality layered on for 2030 onward.

Five. Regional differentiation in PPA prices and energization timelines now exceeds 50 percent across the major data center markets. Where you build matters more than at any point in the post deregulation era. Northern Virginia and Phoenix are saturating. ERCOT West Texas, Columbus Ohio, and select Southeast markets are the relative bargains for the 2026 to 2028 window. Quincy and the cheap hydro era is over.

The enterprises that build the megawatt curve into their capex models the way they currently build in chip generation curves will compete on a different basis from those that do not. The gap is large enough to be the difference between a deployment that earns its capital cost and one that does not.

How we work this problem #

Two of our anchor platforms are built directly against this question. Athena (see /platforms/athena) models compute economics end to end, from chip cost curves through utilization to landed dollar per effective FLOP across procurement paths and regions. Promethean (see /platforms/promethean) models energy infrastructure, PPA structures, queue dynamics, behind the meter economics, and SMR roadmaps at the regional and ISO level. The two platforms are designed to plug into each other so that a capex decision can be tested against both the chip side and the megawatt side of the equation in a single workflow.

Engagements typically combine a workload specific compute model, a regional megawatt feasibility study across two to four candidate sites, and a procurement path tree that maps each site to PPA, behind the meter, and grid connected scenarios with dated energization windows. Reach out at /engage if a decision is coming up.

Sources #

• Lawrence Berkeley National Lab Queued Up 2024 interconnect queue annual report
• IEA World Energy Outlook data center power demand projections
• EIA hourly electric grid monitor
• FERC interconnection studies and electric power markets
• PJM, ERCOT, MISO, CAISO interconnect queue data
• Dominion Energy Virginia 2024 Integrated Resource Plan
• Georgia Power 2025 IRP update filing
• LevelTen Energy PPA Price Index
• ERCOT Permian Basin Reliability Plan
• PJM 2025 to 2026 Base Residual Auction results