AI data centres need 945 terawatt-hours of electricity by 2030 — every second of every day, not just when the sun shines. Big Tech has collectively committed to over 10 gigawatts of new nuclear capacity. Amazon bought a data centre campus directly powered by the Susquehanna nuclear plant. Google signed the first corporate SMR power purchase agreement, with Kairos Power. Microsoft is backing Bill Gates's TerraPower. China's Linglong One is on track to become the world's first land-based commercial small modular reactor in 2026. The nuclear renaissance was supposed to be about decarbonisation. It has become about something more urgent: keeping the AI revolution powered, all the time, without interruption.
Not investment advice. Data sourced from Introl SMR Nuclear Data Centers Report 2025, smrintel.com State of Small Modular Reactors 2026, Energy Solutions Intelligence SMR Report June 2026, IEA Nuclear Capacity Projections, WWT Big Tech Nuclear Bet Analysis, company disclosures and NRC filings. All figures current as of June 2026.
For two decades, the nuclear industry's primary argument was climate change — nuclear as the carbon-free baseload that could decarbonise grids without the intermittency problems of solar and wind. That argument was true and largely unpersuasive to a public scarred by Chernobyl and Fukushima, and to investors who watched nuclear construction projects run years over schedule and billions over budget. The argument that has finally moved markets, governments, and Big Tech balance sheets simultaneously is different: artificial intelligence needs electricity that never stops, and nothing else can reliably provide it at the scale required.
An AI training cluster running continuously for months cannot tolerate the intermittency of solar generation that drops to zero at night or wind that varies with weather. Battery storage at the scale required to bridge multi-day renewable gaps for gigawatt-scale data centres remains prohibitively expensive. AI data centres are projected to require 945 terawatt-hours of electricity by 2030 — a figure that, if it were a single country, would rank among the top ten electricity consumers on earth. The hyperscalers — Microsoft, Amazon, Google, Meta — have done the arithmetic and concluded that nuclear is the only technology that can deliver gigawatts of always-on, carbon-free power at the pace their AI roadmaps demand.
The result: a nuclear investment wave driven not by government decarbonisation mandates but by the balance sheets of the four most valuable technology companies on earth. This is a fundamentally different and more durable demand driver than the policy-dependent nuclear cycles of the past — corporate capital allocated against contracted, multi-decade power purchase agreements does not evaporate with the next election cycle.
TerraPower's Natrium design is the most technologically ambitious and closely watched advanced reactor in the Western pipeline. The 345 MWe sodium-cooled fast reactor pairs with molten salt energy storage — enabling output to surge to 500 MW during peak demand, a flexibility profile precisely suited to the variable loads of AI training workloads that spike during intensive compute phases. The Kemmerer, Wyoming plant secured its NRC construction permit — a major regulatory milestone.
The honest setback: HALEU fuel supply scarcity has pushed the project's timeline from an original 2028 target toward 2030 or beyond — the clearest illustration that even the best-resourced, most prominent advanced reactor project in America is constrained by a fuel supply chain problem that money alone cannot immediately solve. TerraPower's struggle is the industry's struggle in miniature.
X-energy has secured the broadest commercial and government backing of any advanced reactor developer. In the US, it is partnered with Dow Inc. for a demonstration project at a Gulf Coast chemical facility — targeting industrial heat applications beyond pure electricity generation. A landmark Joint Development Agreement with Centrica, announced September 2025, aims to deploy up to 12 Xe-100 reactors in the United Kingdom — the most significant single international SMR commitment disclosed to date.
X-energy has also forged an alliance with Amazon, Korea Hydro & Nuclear Power, and Doosan Enerbility specifically to explore powering AI infrastructure — demonstrating the cross-border, cross-sector coalition-building that characterises the most credible advanced reactor developers. The pebble bed design's meltdown-proof fuel characteristics make it a strong candidate for industrial site co-location where safety case simplicity matters commercially.
NuScale is positioned as the bridge between the legacy nuclear industry and the new factory-built, scalable paradigm — the most mature SMR technology and the go-to partner for risk-averse utilities. Its 77-megawatt modules combine in configurations of 4, 6, or 12 units to create plants ranging from 308 to 924 MW. In September 2025, NuScale announced a partnership with ENTRA1 Energy and the Tennessee Valley Authority for a massive 6-gigawatt SMR deployment programme — the largest single SMR commitment by capacity disclosed in the industry to date.
NuScale also supports the RoPower project in Romania, backed by international financing — demonstrating the technology's applicability beyond the US market. NuScale's design certification with the NRC, completed years ahead of its newer Gen IV competitors, gives it a regulatory head start that translates into earlier commercial deployment even if its underlying technology is less novel.
Kairos secured the distinction of the first-ever corporate SMR power purchase agreement, with Google. The Hermes 2 project in Oak Ridge, Tennessee uses fluoride salt-cooled high-temperature reactor technology — a Generation IV design offering enhanced safety characteristics through the inherent properties of molten salt coolant, which does not pressurise and cannot boil away in an accident scenario the way water-based coolants can.
The initial 50 MW commitment to the TVA grid by 2030, scaling toward 500 MW for Google's data centre operations, represents the template for how a hyperscaler de-risks a first-of-a-kind reactor commitment — start with a modest grid-connected demonstration, then scale the relationship as the technology proves itself operationally. Kairos's molten salt approach is among the most technologically novel in the advanced reactor field, carrying both the highest potential safety and efficiency upside and the highest first-of-a-kind execution risk.
China's Linglong One is on track to become the world's first land-based commercial SMR, with commercial operation expected in the first half of 2026 — beating every Western competitor to actual commercial deployment by years. This is the nuclear equivalent of the sodium-ion battery story (Letter 91): China moving from technology follower to deployment leader through state-directed industrial policy, abundant capital, and a regulatory environment that can move faster than Western licensing frameworks.
China and Russia's advancement in advanced reactor deployment is the geopolitical subplot running beneath the Western hyperscaler nuclear story. If China demonstrates that SMRs can be commercially deployed years ahead of the US and Europe's first-of-a-kind projects, the technology leadership narrative around nuclear's AI-era renaissance shifts meaningfully — and the West's HALEU fuel supply dependency on historically Russian enrichment becomes even more strategically uncomfortable.
The single biggest risk to the entire advanced nuclear industry
Almost every advanced reactor design that matters for the AI nuclear renaissance — TerraPower's Natrium, X-energy's Xe-100, Oklo's Aurora, Kairos's Hermes, Radiant's Kaleidos — requires High-Assay Low-Enriched Uranium (HALEU), enriched to between 5% and 20% U-235, well above the roughly 5% used in conventional light-water reactors but below weapons-grade material. The supply gap between current HALEU production and the volume that reactor developers will need is, by the industry's own assessment, the single biggest risk to the entire advanced nuclear sector.
The geopolitical dimension is stark: Russia's Rosatom/TENEX was historically the only commercial-scale HALEU supplier in the world. The United States banned Russian enriched uranium imports in 2024 — a necessary strategic move that simultaneously removed the industry's primary fuel source overnight. Centrus Energy is currently the only US HALEU producer, running a full-scale 120-centrifuge cascade at Piketon, Ohio, targeting 6 metric tonnes per year of capacity within 42 months of its January 2026 milestone. The NRC authorised enrichment up to 10% in September 2025, with first LEU+ deliveries to fuel fabricators expected in 2026.
The UK is planning a HALEU facility at Capenhurst for the early 2030s. A separate US enrichment project has an NRC license application planned for H1 2026, targeting production by 2031 at a total cost of approximately $5 billion — significant new Western enrichment capacity, but years away from operational scale. This is the same fuel supply chain anxiety that runs through hydrogen (Letter 92), sodium-ion battery materials (Letter 91), and critical minerals (Letter 97) — the AI-driven energy transition is creating fuel and material bottlenecks faster than the industrial base can expand to meet them. Until HALEU supply scales meaningfully, every reactor developer's commercial timeline is hostage to a fuel availability constraint that capital alone cannot solve quickly.
The economic case for SMRs rests on an assumption with strong historical precedent but no guarantee: that factory-based serial production will dramatically reduce per-unit costs compared to today's bespoke, first-of-a-kind projects, the same way standardised manufacturing made solar panels and lithium-ion batteries dramatically cheaper over the past fifteen years. Current first-of-a-kind SMR projects face capital costs of $3,000–6,000 per kilowatt. Manufacturers project these will fall below conventional large nuclear's $7,675–12,500/kW through series production — the entire investment thesis for SMRs depends on this cost reduction materialising as more units are built using standardised factory processes rather than bespoke construction.
The levelized cost of electricity from SMRs currently ranges from $89–102 per megawatt-hour — meaningfully higher than wind and solar at $26–50/MWh, but the comparison is misleading without accounting for capacity factor. SMRs offer capacity factors exceeding 95% — meaning they generate power essentially around the clock — compared to solar's roughly 25% and onshore wind's roughly 35% capacity factors. For a hyperscaler that needs firm, continuous, dispatchable power for a data centre running 24 hours a day, the relevant comparison is not the levelized cost of intermittent renewable generation but the all-in cost of renewable generation plus the battery storage and grid infrastructure needed to make it equally reliable — a comparison in which nuclear's economics look considerably more competitive.
Every SMR project powering a data centre today is a first-of-a-kind deployment of a reactor design that has never delivered commercial power before. Kairos's fluoride salt reactor, X-energy's helium-cooled pebble bed, Oklo's fast microreactor — none of these technologies has an operational track record at commercial scale. Demonstration plants and prototypes, heavily funded by the US Department of Energy, must prove the designs work as expected before commercial confidence is justified. TerraPower's slip from a 2028 target toward 2030+ due to HALEU scarcity is not an aberration — it is the predictable pattern for genuinely novel nuclear technology, and investors should expect more such delays across the sector before, not after, commercial maturity arrives.
The IEA's own projection — 10 to 25 gigawatts of installed SMR capacity globally by 2035, representing just 1–3% of global nuclear capacity — describes a steady evolution, not the overnight revolution that headlines suggest. The hyperscaler deals are genuinely significant as demand signals and de-risking mechanisms for the industry, but the physical reality of building, licensing, fuelling, and commissioning novel reactor designs at scale operates on a timeline measured in years and decades, not the quarterly cycles that technology markets are accustomed to. The nuclear renaissance is real. It is also slower than the AI infrastructure narrative implies — and the gap between AI's electricity demand growth and nuclear's physical deployment timeline is itself a structural risk that other power sources (natural gas peaker plants, existing nuclear plant restarts, grid-scale battery storage) will need to fill in the interim.
The fuel supply chain dependency on historically Russian enrichment capability is the AI nuclear renaissance's most underappreciated vulnerability. The US ban on Russian enriched uranium imports was strategically necessary but has left the entire advanced reactor industry dependent on a US domestic HALEU supply chain that is, as of mid-2026, represented by essentially one commercial-scale producer ramping toward modest initial capacity. Every major advanced reactor company's commercial timeline is hostage to this single point of failure in a way that mirrors the semiconductor industry's Taiwan dependency (Letter 97) — a strategically critical input concentrated in too few hands, with the difference that the nuclear fuel bottleneck is even less diversified today than chip fabrication was before the CHIPS Act response began.
Long-horizon thinking on capital, technology, and the forces shaping the next decade of wealth creation. Written from first principles. Not consensus. Not noise.