The hyperscaler nuclear deals of 2024–2025 are not generic “nuclear power” commitments. When Google, Amazon, and Meta each signed with a specific reactor developer, they picked a distinct technology — one with a different cooling method, fuel type, capacity, cost profile, and regulatory pathway. These are not interchangeable bets. This article is part of our comprehensive coverage of the broader nuclear buildout reshaping AI infrastructure.
Six reactor designs are now in active commercial development: Oklo‘s Aurora Powerhouse, TerraPower‘s Natrium, Kairos Power‘s fluoride-salt SMR, X-Energy‘s Xe-100, Radiant Nuclear‘s Kaleidos microreactor, and HGP Intelligent Energy’s Navy reactor repurposing proposal. They span a capacity range from 1 MW to 520 MW per unit. None is commercially generating power today. The earliest credible commercial milestone is 2030.
This article compares the leading designs on capacity, cost per MWh, timeline, and regulatory risk — so you can assess the deals accurately.
What is a small modular reactor and how does it differ from a conventional nuclear plant?
A small modular reactor (SMR) produces up to 300 MW of electricity per unit — roughly one-quarter of a conventional nuclear plant’s output.
Factory prefabrication, not on-site construction. Conventional plants are custom-engineered on-site over 10–20 years. SMR components are factory-manufactured as standardised modules and assembled on-site, targeting 24–36 month build times once certified.
Modular scaling. Commit to four modules, expand to twelve as demand grows — matching power delivery to data centre build-out instead of making a single large capital bet upfront.
Cooling technology diversity. Conventional plants almost universally use light-water cooling. The six designs here use sodium, fluoride salt (FLiBe), helium, heat pipes, and light water — and those choices are not interchangeable. They determine operating temperature, fuel type, and regulatory pathway.
One sub-category matters: reactors below 10 MW are technically microreactors. Radiant Nuclear’s Kaleidos (1 MW, semi-trailer transportable) raises different sizing questions for data centre use.
Key terms:
- LCOE (levelised cost of electricity): all-in cost per MWh over a plant’s lifetime. The apples-to-apples cost metric.
- PPA (power purchase agreement): the contract hyperscalers use to commit to SMR power before reactors are built. For pre-commercial SMRs, a PPA is a reservation — not an invoice.
- HALEU (high-assay low-enriched uranium): enriched to 5–20% U-235, versus 3–5% in conventional plants. Required by Oklo; domestic US supply is limited.
- TRISO fuel: uranium particles in multiple ceramic layers that physically resist meltdown even with total cooling loss. Used by X-Energy, Kairos, and Radiant Nuclear.
For context on the hyperscaler deals that reference these specific reactor types, the competitive dynamics behind each partnership matter as much as the technology.
How do the leading SMR designs compare on capacity, cost target, and timeline?
Here are the six designs in active commercial development.
Aurora Powerhouse — Oklo — 75 MW per unit — $80–130/MWh target — ~2030 (Meta deal, Pike County OH) — DOE RPP pilot (Aurora-INL, July 2026 target); NRC commercial licensing in progress — Backed by OpenAI / Sam Altman
Natrium — TerraPower — 345 MW plus up to 500 MW surge (molten-salt storage, 5+ hrs) — $50–60/MWh target (nth plant) — ~2032 (Meta deal, 2 reactors / 690 MW) — NRC COL process; Kemmerer WY groundbreaking June 2024 — Backed by Bill Gates / Nvidia
Fluoride-Salt SMR — Kairos Power — ~75 MW per unit (~500 MW across 6–7 units) — Cost TBC — First unit by 2030 (Google/TVA deal) — Pre-commercial; NRC COL process — Backed by Google
Xe-100 — X-Energy — 80 MW per module (320–960 MW per site) — Cost TBC — Cascade Advanced Energy Center; 2030 construction target, power “in the 2030s” — Pre-commercial; NRC COL process; $700M Series D (Nov 2025) — Backed by Amazon / Jane Street
Kaleidos — Radiant Nuclear — 1 MW per unit (transportable) — Cost TBC — DOME criticality at INL expected 2026; commercial TBC — Pre-commercial; $300M raised (Dec 2025) — Backed by Equinix (20 units pre-order)
Navy Reactor Repurposing — HGP Intelligent Energy — 450–520 MW (2 repurposed reactors) — $1–4M/MW conversion cost — TBC; DOE loan guarantee application pending — Proposal stage; no construction commenced
One design stands out.
TerraPower’s thermal storage advantage. The Natrium reactor produces 345 MW of heat continuously. When grid demand is low, excess heat goes into insulated tanks of molten salt. When demand peaks, stored heat boosts output to roughly 500 MW for five or more hours — without changing reactor output at all. It’s a built-in battery without the lithium chemistry constraints. No other design in this comparison can do that.
On cost targets. Only TerraPower has published a specific LCOE target ($50–60/MWh). Oklo’s $80–130/MWh range has been cited publicly. For Kairos, X-Energy, and Radiant, cost is genuinely TBC — these are physics-proving operations, not utility cost-modelling exercises.
What does SMR power actually cost — and how does that translate to AI infrastructure economics?
Start with where costs are today, not where developers hope they will be.
Current state: SMR LCOE is currently estimated at $89–102 per MWh for first-of-a-kind installations. Utility-scale solar runs $26–50/MWh; onshore wind $25–50/MWh; combined-cycle gas $40–75/MWh. SMRs are not cheap.
Target state: TerraPower’s target is $50–60/MWh for a mature nth-of-a-kind Natrium plant. Oklo targets $80–130/MWh. None of these apply to the first plants being built now.
At $89/MWh, a 100 MW AI training cluster running for 1,000 hours costs roughly $8.9 million in electricity. At TerraPower’s $50/MWh target, that drops to $5 million — a 44% reduction. Scale to a 10 GW campus at 8,760 hours per year and the difference is $342 million annually. That number is why hyperscalers are locking in capacity years before a single reactor is built.
A counterpoint worth noting. The Centre for Net Zero (Octopus Energy’s research group) found that renewables plus a small amount of gas costs 43% less than SMRs for a 120 MW facility. The premium is real. The case for nuclear rests on 24/7 carbon-free dispatchability — not raw cost competitiveness.
First-of-a-kind cost reality. NuScale‘s Idaho project escalated from $5.3B to $9.3B before cancellation. Vogtle Units 3 and 4 cost $35B — double the projection, seven years late. SMRs need to break this pattern before their targets deserve full credibility.
For a detailed comparison against alternatives, see how SMRs compare against renewable and gas alternatives.
How does NRC licensing affect SMR deployment timelines?
The Nuclear Regulatory Commission is the primary bottleneck for commercial SMR deployment in the US. Every design here requires an NRC Combined Licence (COL) for commercial operation. This is not a formality — it adds years, and outcomes for novel reactor designs are genuinely uncertain.
Oklo’s difficult path. After its Aurora licence was rejected — insufficient accident analysis — Oklo pivoted to two tracks. Track one: the DOE Reactor Pilot Programme (RPP), letting Aurora operate at Idaho National Laboratory under DOE jurisdiction with a first-criticality target of July 4, 2026. Physics test, not grid power. Track two: resubmitting the NRC commercial application.
TerraPower’s comparative success. Natrium broke ground at Kemmerer, Wyoming in June 2024 under a combined construction and operating licence — the first commercial advanced reactor construction in the US. TerraPower’s sodium-cooled design has decades of experimental data behind it, which appears to have eased the NRC’s review considerably.
Reforms on the books. The ADVANCE Act (2024) directs NRC to streamline licensing — fee reductions, new pathways for DOE-site demos, manufacturing licences for factory-certified modules. Part 53 (2024) is a new technology-neutral framework replacing the Part 50/52 light-water system. Both are directionally positive; Part 53 has not yet been used for a complete commercial application.
NRC licensing adds three to seven years beyond the construction period. Plan around it, not over it.
For a deeper analysis, see the NRC and UK DESNZ licensing context shaping these timelines.
What is the HGP Intelligent Energy Navy reactor repurposing proposal and why does it matter?
HGP Intelligent Energy is a Texas-based startup petitioning the DOE for a loan guarantee to repurpose two retired US Navy nuclear reactors — Westinghouse A4W (Nimitz-class carriers) and GE S8G (Los Angeles-class submarines) — into a 450–520 MW grid-connected plant at Oak Ridge National Laboratory.
The cost claim is the reason to pay attention.
HGP estimates conversion at $1–4 million per MW. New SMR construction: $3–6 million per MW. Conventional new nuclear: $7.675–12.5 million per MW. If accurate, this approach is 33–67% cheaper than the next cheapest option.
The logic holds on first principles: the reactor vessels, containment structures, and fuel systems already exist — you’re avoiding the largest capital cost components of a new build. ORNL’s involvement adds credibility. The capacity range puts it in the same tier as Kairos Power’s Google deal.
But the risks are substantial. Naval reactor designs have never been certified for civilian grid connection — no NRC or DOE precedent exists. The $1–4M/MW figure is HGP’s own estimate. Tom’s Hardware’s December 2025 story is the only detailed public coverage.
Track it, note ORNL’s involvement, treat the cost as a hypothesis. Don’t model it yet.
Which SMR designs are furthest along — and what are the realistic timelines?
No SMR is commercially generating electricity at grid scale in the US as of April 2026. China’s Linglong One (210 MW, Hainan Province, 2023) proved the technology works at commercial scale. The US regulatory environment is a different matter.
2026: physics proofs, not power. Oklo’s Aurora-INL targets first criticality at Idaho National Laboratory by July 4, 2026 under the DOE RPP — validating the design’s physics, not delivering grid power. Radiant Nuclear’s DOME criticality test at INL is also expected in 2026.
2030: first credible commercial milestone. Kairos Power’s first unit for Google targets operations at Oak Ridge, Tennessee by 2030 through a three-party structure: Kairos (technology), TVA (utility), Google (clean energy attributes). “First unit” is not full contracted capacity — the ~500 MW deal is phased through 2035.
2032: furthest-advanced construction. TerraPower’s first Natrium reactor for Meta targets commercial operation approximately 2032. Kemmerer, Wyoming broke ground in June 2024. Actual earth has been moved — that matters more than any announcement.
2030s (TBC): Amazon/X-Energy. Cascade Advanced Energy Center targets 2030 for construction commencement, power “in the 2030s.” X-Energy’s $700M Series D and 11+ GW orderbook are strong signals; construction has not started.
What a “deal” actually means. PPA announcements are reservations and option rights, not delivery dates. Meta’s January 2026 deals came from a December 2024 RFP for 1–4 GW by the early 2030s — securing priority access and de-risking developer financing. Not ordering next year’s power.
For capacity planners: 2030 is the optimistic end; 2032–2035 is more realistic for meaningful scale. TerraPower and X-Energy have the strongest combination of capitalisation, regulatory progress, and construction activity right now.
Frequently Asked Questions
What is TRISO fuel and why do multiple SMR designs use it?
TRISO (tristructural isotropic) fuel encases uranium particles in multiple ceramic layers inside graphite spheres roughly the size of a golf ball. The structure physically prevents meltdown — the fuel cannot reach failure temperatures even if all cooling is lost. X-Energy, Radiant Nuclear, and Kairos Power all use it. The passive safety profile simplifies regulatory approval compared to conventional uranium oxide fuel.
How does TerraPower’s molten-salt thermal storage system work?
The Natrium reactor produces 345 MW of heat continuously. Excess heat is stored in insulated molten-salt tanks during low-demand periods. When demand peaks, stored heat boosts output to approximately 500 MW for five or more hours — without changing reactor output. For AI data centres with variable workloads, a reactor that surges on demand is structurally more useful than one delivering flat baseload.
Why did Google choose Kairos Power, Amazon choose X-Energy, and Meta choose both Oklo and TerraPower?
Google/Kairos — the three-party TVA structure lets Google work through established utility infrastructure rather than taking direct development risk. Amazon/X-Energy — Amazon funded development, licensing, and construction directly, treating early-stage capital as the market gap to fill. Meta/Oklo+TerraPower — a diversified bet across two designs and two timelines, hedging a faster smaller-unit design against a more proven larger-unit design with storage. Geography and existing relationships likely played as large a role as technical selection.
How many SMR units would a large AI data centre actually need?
A 500 MW AI training campus would require approximately: 7 Oklo Aurora units (75 MW each), 1–2 TerraPower Natrium reactors (345 MW base), 7 Kairos Power modules, 7 X-Energy Xe-100 modules, or around 500 Radiant Nuclear Kaleidos units (1 MW each). That last number explains why Radiant targets edge deployments. At 140 kW per modern Nvidia rack, one Kaleidos unit powers two to seven racks — useful for forward-deployed inference, not training clusters.
Is HGP Intelligent Energy’s Navy reactor cost claim credible?
Plausible on first principles — the reactor vessels and containment already exist, so you avoid the biggest capital costs of a new build. But civilian grid certification for the A4W and S8G designs has no precedent. The $1–4M/MW figure is HGP’s own estimate with no independent validation, and Tom’s Hardware’s December 2025 story is the only detailed public coverage. Track it; don’t model it yet.
What is HALEU and why does it matter for Oklo’s timeline?
High-assay low-enriched uranium (HALEU) is enriched to 5–20% U-235, versus 3–5% in conventional plants. Oklo’s Aurora requires it. US domestic enrichment capacity is limited — Centrus Energy is the primary commercial source. Any supply delay directly delays Oklo’s commercial timeline. TRISO-based designs and TerraPower’s Natrium do not face this constraint.
What does “first criticality” mean and why does the Aurora-INL milestone matter?
First criticality is the moment a reactor achieves a self-sustaining chain reaction — the physics proof that the design works as modelled. Aurora-INL’s target of July 4, 2026, is a DOE Reactor Pilot Programme milestone at Idaho National Laboratory — not a commercial event and not an indication that Meta’s power is imminent. For investors, counterparties, and regulators, it is the most important near-term proof point for Oklo’s entire commercial programme.
What is the ADVANCE Act and does it actually accelerate SMR timelines?
The ADVANCE Act (2024) introduced 50% fee reductions for SMR applicants, new DOE-site demonstration pathways, manufacturing licences for factory-certified modules, and parallel processing of site permits and design certifications. Part 53 (2024) is the NRC’s new technology-neutral licensing framework replacing the Part 50/52 light-water system. Both are directionally positive; time savings are hard to quantify until the first advanced reactor completes the new process. See the regulatory environment governing each design’s timeline for the full picture.
What is a PPA and how do hyperscalers sign one with a pre-commercial reactor?
A power purchase agreement commits a buyer to purchase electricity at a specified price per MWh over a term — typically 10–20 years. For pre-commercial SMRs, the structure includes milestone-dependent obligations, option rights for additional capacity, and often direct development funding. Google’s three-party model (Google → TVA → Kairos) has TVA purchasing electricity while Google procures clean energy attributes. Amazon funded X-Energy’s development directly. Neither is paying for power today — both are securing priority access and helping ensure the plant gets built.
Why is Equinix buying Radiant Nuclear microreactors instead of a utility-scale SMR?
Equinix is a colocation provider, not a hyperscaler building AI campuses. Its use case is distributed resilience: capacity in markets where grid power is unreliable, expensive, or unavailable. The Kaleidos microreactor — 1 MW, semi-trailer transportable, TRISO fuel, refuelable up to four times over 20 years — is designed for exactly this. Equinix’s pre-order of 20 units (20 MW total) is a different thesis: edge resilience, not gigawatt-scale AI power.
Are SMRs carbon-free and does that matter for AI data centre sustainability commitments?
Nuclear produces near-zero direct carbon emissions, with lifecycle carbon intensity comparable to wind and solar. For hyperscalers with 24/7 carbon-free energy (CFE) commitments — Google’s 2030 target, Microsoft’s 2030 target — SMRs are attractive because they are dispatchable. A 345 MW Natrium reactor delivers 345 MW around the clock, closing the overnight and cloudy-day gaps that battery storage can only partially fill.
This is part of nuclear power’s AI renaissance — always-on, carbon-free generation that hyperscalers cannot achieve with renewables alone.
