This summer the Nuclear Regulatory Commission concluded that restarting Three Mile Island’s Unit 1 — the plant synonymous with nuclear disaster since 1979 — would have no significant environmental impact. Constellation Energy now expects the renamed Crane Clean Energy Center to generate power again in late 2027, under a 20-year, $16 billion agreement it signed with Microsoft to route the reactor’s output directly to Microsoft’s data centers, not the regional grid. Weeks later, PJM’s capacity auction — which prices electricity for 65 million Americans — hit its price cap for the third straight year, at $329 per megawatt-day, up nearly ninefold since 2024. Regulators attribute 63% of that jump to data centers: a $9.3 billion bill passed to ordinary ratepayers. Power has become the resource the AI race cannot route around.
THE CONTEXT
Until roughly 2023, the AI race was framed almost entirely as a chip story: who could buy enough Nvidia GPUs, and which countries Washington’s export-control regime would let access them. That regime still matters — the UAE was only elevated to the top “trusted” tier in July 2026, with chip access still restricted to a short list of approved entities. But training and running frontier models now demands electricity at a scale chips alone can’t explain: a single next-generation AI campus can draw 500 megawatts to over a gigawatt continuously, comparable to a mid-sized city. Since 2025, US grid operators have struggled to connect new large loads fast enough — interconnection queues in AI hotspots now run four to seven years, transformer orders take up to five, and only a fraction of announced capacity has broken ground. That gap is why Microsoft, Google, Amazon and Meta have each signed multi-billion-dollar, multi-decade deals to revive old reactors or pre-order ones that do not yet exist.
THE ARGUMENT
Stay ahead of the geopolitical week.
MD Briefing delivers expert analysis across five global fronts — the Indo-Pacific, energy, geoeconomics, European security, and the Middle East — every Monday morning. Free.
Most coverage of AI geopolitics still treats chip export licenses as the decisive lever — who Washington lets buy Nvidia’s best silicon, and who it cuts off. That framing is increasingly out of date. A chip is fungible: it can be bought, leased, routed through a subsidiary, or in the worst case smuggled. A gigawatt of firm, always-on power cannot. It requires transmission lines, transformers with multi-year lead times, and either an existing reactor fleet or a government willing to underwrite a new one. That difference is now the more decisive one.
Consider the Gulf. In July 2026, Washington eased chip export rules for the UAE, elevating it to its top country tier. Money is not the constraint either: Saudi Arabia’s HUMAIN alone has announced 6.6 gigawatts of AI capacity for 2034. Yet as of that same month, roughly 4,000 megawatts of announced Saudi and Emirati AI projects had zero documented energisations — industry shorthand for a project actually receiving grid power. Riyadh currently operates 467 megawatts of live data-centre capacity against that 6.6-gigawatt target: a fourteen-fold gap. Power-delivery timelines of four to seven years, driven by transformer procurement and utility agreements, now set the schedule that chip access alone cannot accelerate. Washington’s permission to buy chips is necessary. It is no longer sufficient.
The same constraint binds the country with the least reason to expect it: the United States itself. OpenAI’s $500 billion Stargate initiative showed, by one industry tracker’s count, no significant physical progress at its flagship Texas site as of April 2026, even as chip supply and capital were never in question. Nationally, roughly 12 gigawatts of AI data-centre capacity was promised for construction in 2026; only about 5 gigawatts actually broke ground. Interconnection queues in Northern Virginia, Phoenix and Dallas now run four to seven years. The best-capitalised, most chip-privileged companies on Earth are discovering that neither money nor an export licence can buy a faster grid connection.
The nuclear pivot is the industry’s answer, and it reproduces the export-control logic in a different institutional form. Restarting Three Mile Island, pre-ordering TerraPower’s Natrium reactors for Meta, or committing Amazon to X-energy’s small modular reactors are all bets that firm power, however slow to arrive, beats waiting on a grid that cannot expand fast enough. Crucially, exporting that solution abroad runs through the same kind of bilateral vetting as chip diffusion tiers: US civil nuclear cooperation requires a “123 Agreement” with each partner government, and 2026’s executive orders on nuclear set a target of twenty new ones by 2028. Access to American reactor technology and fuel, in other words, will be gated country by country — just as chip access already is.
The predictable objection is that this is more announcement than substance: most small modular reactor designs remain commercially unproven, and TerraPower’s and Oklo’s timelines stretch to 2032 and beyond. That is fair for the 2030s. It does not explain away the 2026-2030 window, which is the one that matters for the current capital cycle. Gas turbine order books are already sold out for years, renewables cannot supply the continuous baseload frontier clusters are contracted to need, and only a finite number of mothballed reactors exist to restart. Hyperscalers are locking in nuclear precisely because it is the only lever left that can plausibly deliver firm, gigawatt-scale power this decade — which is exactly why the deals are being signed years before a single additional electron flows.
THE SCENARIOS
Base case (roughly 55%): incremental progress. Crane and a handful of other US restarts come online close to schedule by 2027-28, and a first wave of small modular reactors reaches commercial operation by 2030. This assumes NRC licensing timelines hold and no flagship project suffers a Vogtle-style cost or schedule blowout. The US, and hyperscaler-run capacity in a small set of trusted partner states, consolidates a real lead. Gulf states continue announcing multi-gigawatt targets while operating a fraction of that capacity, hosting substantial but sub-frontier clusters through the decade.
Downside case: the bottleneck compounds rather than eases. A high-profile restart or small modular reactor project slips — first-of-a-kind nuclear projects have a long history of cost and schedule overruns — while transformer and turbine supply chains remain the binding constraint regardless of reactor type. Frontier capacity concentrates even further into the handful of jurisdictions with spare grid headroom, and the ratepayer backlash already visible in PJM’s $9.3 billion bill hardens into state-level restrictions on new data-centre interconnections, slowing even the jurisdictions currently winning. For Gulf and other capital-rich, power-poor states, this scenario means the gap between announced gigawatts and energised megawatts widens rather than closes before 2030.
Upside case: licensing genuinely accelerates. The Nuclear Regulatory Commission’s streamlined framework and Department of Energy co-location on federal land produce operating small modular reactors ahead of schedule, while a faster-than-expected run of new 123 Agreements extends trusted nuclear-and-AI status to partners beyond today’s shortlist. In this scenario the 123 Agreement, not the chip licence, becomes the document worth tracking — and frontier AI hosting capacity diffuses faster than current grid math suggests, reshaping which capitals matter to this story by the end of the decade.
THE TAKEAWAY
Chip export controls will keep generating headlines, but they are no longer the ceiling on who can build frontier AI — the grid is. A GPU can be bought, leased or smuggled; a gigawatt of firm nuclear power cannot. Whoever controls a credible path to one — a restartable reactor fleet, a small modular reactor order book, a 123 Agreement, or federal land for co-location — now controls the scarcer resource. Watch the Nuclear Regulatory Commission’s final environmental assessment on Crane Clean Energy Center, due this month, and its operating-licence decision expected in May 2027: the pace of that single approval will say more about who gets to host the next generation of AI than any export-control announcement out of Washington.
