The first piece in this series argued the AI buildout is a financing story: $700-730 billion in 2026 hyperscaler capex, up 70-80% from last year, and money is not the constraint. This one is about the constraint that actually is.
The industry is spending $700 billion a year to build data centers faster than anyone can power them.
The demand is not theoretical.
Lawrence Berkeley National Lab's August 2026 update puts US data-center electricity at 521-843 TWh by 2030, with a reference case of 649 TWh, or 11.8% of all US electricity. Goldman Sachs sees US data-center power demand going from 31 GW in 2025 to 66 GW in 2027, from 4.1% to 8.5% of peak summer demand. Moody's puts the bill at roughly $110 billion for 45 GW of new generation by 2030. Globally, the IEA has data-center consumption rising from ~415 TWh in 2024 to ~945 TWh by 2030.
Treat these reference cases as ceilings, not central estimates.
Big Tech is buying electrons. Some are real.
The deal flow of the past year splits cleanly into two buckets: contracted electrons with dates, and capacity options dressed as announcements.
In the first bucket: Microsoft's 20-year PPA with Constellation to restart Three Mile Island Unit 1, 835 MW, targeted for 2027. Amazon's $650 million for a campus co-located at the Susquehanna nuclear plant, 960 MW initially rising to 1.92 GW through 2042, wired directly behind the meter. Oracle's October deal for 125-250 MW from Wisconsin's Point Beach plant.
In the second bucket: Google's deal with Kairos Power for up to 500 MW, first 50 MW unit around 2030. Amazon's $700 million into X-energy for up to 12 SMRs. Meta's three agreements totaling up to 6.6 GW across TerraPower, Oklo, Vistra, and Constellation, with first power "as early as" 2030.
Add it up and Big Tech has contracted more than 10 GW of new nuclear capacity in the past year. But notice the structure: the restarts and co-locations are power. The SMR deals are development pipelines. Options, not electrons.
The SMR backlog race, ranked by paper quality.
Three public companies are competing to be the SMR story, and their backlogs tell you everything about how seriously to take each one.
Oklo has the biggest headline: 18.1 GW at end of 2025 (about 14 GW in 2026 reporting), though roughly 12 GW of that comes from a single customer under a largely nonbinding deal. But it also has two things neither rival can claim. First, the only Big Tech check actually written to an SMR developer: Meta's January deal for up to 1.2 GW at a Pike County, Ohio campus, where Meta prepays for power and funds early development. Google and Amazon signed future-power agreements; Meta put money down. Second, the fastest plausible path to first power: Oklo's Aurora unit targets 2028 on the DOE pilot track at Idaho National Lab, which can move quicker than the full NRC licensing queue everyone else is stuck in. Oklo also sells electricity rather than reactors, so its incentives point at operating plants, not winning design contests. It reported its first-ever revenue in Q2 2026: $1.2 million, from services, not power. Early, but real.
NuScale has the best paper of the three and the worst precedent. It's the only SMR with a full NRC design certification (January 2023), and its uprated 77 MWe module got design approval in May 2025. Its backlog is 6-7 GW, nearly all from one eastern-US utility customer, largely nonbinding. Its most advanced project, in Romania, now expects first commercial operation in 2033.
GE Vernova is the one actually pouring concrete. Its BWRX-300 is under construction at Darlington in Ontario since May 2025, with Unit 1 targeted for 2029. The NRC issued the first US construction permit for the design to TVA's Clinch River site in September 2026. Nuclear is under 3% of GE Vernova's sales; the SMR is an option, not the thesis.
The honest ranking: GE Vernova has the only project under construction. NuScale has the only certified design and no firm build. Oklo has the largest headline on the weakest paper, but also the only Big Tech check and the fastest plausible route to first power. And none of the three has delivered a commercial electron from a new-build SMR.
The skeptic's ledger.
Small modular reactors have already failed on cost once. NuScale's flagship Utah project with UAMPS was cancelled in November 2023 after the price went from $4.2 billion to $9.3 billion and the target price from $58 to $89/MWh, up 53%. Even DOE's $1.36 billion cost-share couldn't save it. Current SMR cost estimates run $100-180/MWh, and a European study puts average SMR capital cost at EUR 7,031/kW, about 41% above large reactors.
The bridge fuel is bottlenecked too. Gas turbine costs are up 195% since 2019. Industry output is 60-70 GW a year against 110+ GW of orders. GE Vernova's gas equipment backlog sits at 116 GW, with delivery slots offered into 2031. Operators are paying $25 million just to reserve a 2030 slot, and non-refundable reservation fees are now standard.
The fuel doesn't exist at scale. Centrus, the only US producer of the high-assay uranium advanced reactors need, has made less than 2 metric tons since 2019. Large volumes don't arrive before 2029.
US residential electricity rates are up 42% since 2019. PJM's December capacity auction cleared at the price cap and still came up short, with prices up 9x since 2023. Oracle's Wisconsin deal is at the center of a proposed $176 million rate hike fight.
When do the electrons actually arrive?
Three Mile Island restarts in 2027. Oklo targets 2028 for its first Aurora unit, company guidance, with the NRC license still pending. Darlington's first unit is targeted for 2029. TerraPower's Natrium, Kairos's first 50 MW unit, and the Meta-Oklo campus all point to 2030 "at best." Romania is 2033.
Read that list again. The only new nuclear electrons contracted for this decade come from restarting reactors built in the 1970s. Everything else is 2029-2033, at best, assuming nothing slips. Nuclear has never not slipped.
What fills the gap is natural gas (more than 30 of Moody's 45 GW), grid uprates, behind-the-meter projects (about 30% of new capacity is going on-site), aeroderivative turbines, and fuel cells. Fossil fuels still supply more than 60% of AI power today.
One more risk, on the demand side.
This week, four rival lab CEOs agreed for the first time that the frontier should slow down, and the infrastructure trade sold off the same day. It may be a press cycle. But file it: if the labs ever mean it, every 2030 demand forecast above is the bull case.
So where does it break?
Not a prediction. A watchlist. Whether Three Mile Island actually restarts in 2027 on schedule. Whether Oklo's NRC license application survives contact with the regulator that denied its 2020 application. Whether Darlington's 2029 target holds. Whether gas turbine delivery slots stretch past 2031. Whether the slowdown talk becomes more than a press cycle.
The first piece said the question was never whether AI needs the capacity, but who holds the risk when the music slows. The power story sharpens it: the risk isn't just financial. It's physical. You can finance a data center in a quarter. You can't build a grid connection in one, and you can't pour a reactor in two.
Money was never the constraint. The electrons were. And they're on a 2029-2033 schedule against a $700-billion-a-year construction program.
Originally published on AI & Markets: https://aiandmarkets.substack.com/p/the-ai-buildout-has-a-power-problem
Top comments (1)
Power is the constraint that shows up after the capex number. If you cannot energize a site, a lower inference cost does not help the workload you already built the GPUs for. Which constraint do you think binds first in 2026: the substation, or the model you can run on the power you already have?
iin1006h09