In partnership with
In London this week, more than 500 nuclear and data centre leaders meet to ask which reactor can feed AI. In New York, the General Assembly opens its climate and energy summit on Wednesday. Both rooms face the same fact. Every route to firm power is booked at once.
The evidence sits in three datasets. GE Vernova holds 100 gigawatts (GW) of gas turbines under contract. About 10 GW of slots remain across 2029 and 2030. Hyperscalers have signed nuclear agreements worth up to 13 GW, most of it not due before the 2030s. And 2,061 GW of generation and storage waits in the US interconnection queue, with a median wait over five years.
The bottleneck has moved from the chip to the megawatt. This month, all three supply routes report full.
The opening is three queues, not one. Gas, nuclear and grid report full in the same month
Each constraint has been reported on its own. The point is that they now bind together. A campus that cannot get a grid date turns to gas. Gas needs a turbine slot, and the slots are nearly gone through 2030. Nuclear is the long answer, and the first new units arrive in the 2030s. There is no unconstrained route left to switch to. That is new. The industry meets this week with every door partly closed.
Gas: 100 GW under contract and about 10 GW of slots left before 2031
Start with the fastest route. GE Vernova signed 21 GW of new turbine agreements in one quarter. Its contracted total rose from 83 to 100 GW, and slot reservations from 43 to 56 GW. About 20% of that volume is tied directly to data centre load. Customers pulled hard into 2030, so roughly 10 GW of capacity remains across 2029 and 2030 combined. Prices on new orders run 10 to 20 points higher per kilowatt than late 2025. Wood Mackenzie expects turbines to reach £451 ($600, €528) per kilowatt by end 2027, near triple the 2019 level.
Proposals outrun the forges. 1,047 GW planned, two thirds with no named maker
The demand side of gas has run ahead of the supply side. Global gas capacity in development rose 31% in 2025, to 1,047 GW. The United States nearly tripled its pipeline to 252 GW, a third of it for on-site data centre power. Building all of it would cost over £313bn ($416bn, €366bn). Yet two thirds of projects in development name no turbine manufacturer. Three makers hold over 75% of the named market, and their backlogs run through 2030. Capital is available. The machine is not.
Gas slots, nuclear units and grid dates are all years away. Which do you secure first?
Nuclear: up to 13 GW signed, and the new build arrives only in the 2030s
The long route is spoken for too. The Carnegie Endowment counts hyperscaler nuclear agreements at up to 13 GW. About half are power purchase agreements (PPAs) on existing plants. The rest are new-build partnerships. The new-build half could add about 6.1 GW by the mid-2030s, if developers deliver. Financial terms of the largest deals were not disclosed. The capacity available this decade comes from plants already built, restarted or extended. The new units are a decade out, and the best sites are already under contract.
Grid: 2,061 GW waits five years, and PJM pays its cap of £251 per MW-day
The shared route is the slowest. Lawrence Berkeley National Laboratory (LBNL) counts about 8,200 active projects in the US queue. Together they hold 1,312 GW of generation and 749 GW of storage. The median project built in 2025 spent over five years waiting. Only 13% of capacity requested between 2000 and 2020 ever reached operation. The price of the shortage is visible. PJM runs the grid for 67 million Americans. It cleared its 2027/28 capacity auction at its cap of £251 ($333, €293) per megawatt-day. Data centres drove nearly 5,100 MW of a 5,250 MW load increase.
Demand does not wait. The IEA sees 415 TWh become 945 TWh by 2030
The pull behind all three queues keeps growing. The International Energy Agency (IEA) puts data centre consumption at about 415 terawatt-hours (TWh) in 2024, 1.5% of global electricity. Its base case doubles that to about 945 TWh by 2030, growing 15% a year. AI-focused servers grow fastest, at 30% a year. The United States, China and Europe account for most of the increase. Supply is constrained on every route, and demand is compounding on all of them at once.
Britain sorts 100 GW of demand into blocks that reach 2035
The pattern is not American. Britain's National Energy System Operator (NESO) rebuilt its connection queue in 2025. It cut a 700 GW generation pipeline to 238 GW retained. More than 100 GW of large demand, mostly data centres, was sorted into blocks stretching to 2035. Only about 13 GW of firm demand can connect before 2030. Lead times run five to ten years. A £40bn ($53bn, €47bn) clean-power plan underwrites the fix. The government has offered 500 MW per AI Growth Zone to bypass the queue. The offer shows how scarce the queue is.
The Gulf and Asia answer with their own supply. Stargate UAE, Korea's 18.4 GW, China's 40 GW
Regions with sovereign capital build around the queues. Abu Dhabi's Stargate campus draws nuclear, solar and gas built alongside the halls. Its first 200 MW goes live this month. South Korea's AIDC Alliance targets 18.4 GW by 2035, though its own grid operator carries heavy debt and delayed lines. China reaches about 40 GW of data centres this year and moves hubs west to find power. We note China as an engineering observation, not a model. Each route still meets a constraint. None escapes all three.
The investment implication: sequence the routes, because none clears alone
The capital lesson is about order, not choice. No single route delivers a campus on time, so the schedule has to stack them. Existing and restarted nuclear delivers first, from 2027, but the volume is small and largely contracted. Gas delivers next, but only for buyers holding one of the few 2029 or 2030 slots. New nuclear and grid expansion arrive in the 2030s. A credible plan names which route covers which year. Underwriting that prices one route and ignores the other two misreads the asset. The General Assembly's energy summit on 23 September will hear the same arithmetic from governments.
The bottom line
The chip is no longer the ceiling. The megawatt is, and every route to it is booked. Gas is sold through 2030, nuclear is contracted into the 2030s, and the grid waits five years. Demand doubles by 2030 regardless. This week's London summit asks which reactor feeds AI. The honest answer is a sequence, not a source. Intelligence can map which route clears first for a given site. Someone still has to build all three.
Next week: Sovereign AI is an energy policy. The standards decide who can build fast.
For product teams moving at AI speed.
AI makes it easier to ship anything, even bad ideas. The hard part is knowing which ideas are worth building.
Jira Product Discovery brings your ideas, customer insights, and priorities into one place, so your team can decide what to ship and move forward with confidence.
Capture ideas, prioritize with evidence, and build living roadmaps your team can rally around—all while staying connected to delivery in Jira, so everyone can see what’s being built and why.
Better product decisions in the AI era.