▮▮Coloprice
← Guides and analysis

· power

SMRs and Nuclear Power for Data Centers: 2026 Status

Hyperscalers have committed close to 10 GW of nuclear capacity for data centers, but the first SMR-powered facility won't run before 2029-2030. Here's the real timeline.

SMRs and Nuclear Power for Data Centers: 2026 Status

Hyperscalers have committed close to 10 GW of nuclear capacity to data center power over the past two years, but no small modular reactor (SMR) is generating commercial power for a data center anywhere in the world as of August 2026. The earliest deployments — from Oklo and X-energy — target 2029-2030. The nuclear capacity that is closest to delivering, Microsoft’s restarted Three Mile Island Unit 1, is a conventional large reactor, not an SMR. Buyers sourcing capacity for 2026-2028 should treat SMRs as a roadmap signal, not a procurement option.

Key takeaways

  • Nearly 10 GW committed, zero GW delivered. Microsoft, Google, Amazon and Meta have signed 13+ nuclear agreements, but every SMR-specific deal remains pre-construction as of mid-2026, per SMR Intel.
  • Meta leads by volume: up to 6.6 GW across TerraPower, Oklo, Vistra and Constellation agreements targeting 2032-2035 delivery.
  • Microsoft has the nearest firm power: a ~$16B, 20-year PPA for the full 835 MW of the Crane Clean Energy Center (Three Mile Island Unit 1 restart), targeted for H2 2027 — but this is large-reactor restart capacity, not an SMR.
  • Realistic SMR timeline is 2029-2030 at the earliest, driven by sequential NRC licensing, first-of-a-kind construction, and commissioning — not by any single blocker.
  • Projected SMR cost ($60-80/MWh nth-of-a-kind) is unproven at scale. NuScale’s only late-stage US project saw estimated capex exceed $20,000/kW before cancellation in 2023.
  • Grid rules are still being written. FERC’s December 2025 order on co-located generation is still working through PJM compliance filings as of April 2026, so even sites with an adjacent reactor don’t yet have final interconnection economics.

For power benchmark data by market, see our power sector hub and colocation price index.

Why nuclear, and why now

Global data center power demand is projected to reach roughly 219 GW by 2030, and grid interconnection queues in top markets already run 3-7 years — the same bottleneck covered in our grid connection guide. Nuclear offers what solar and wind cannot: baseload, 24/7 output with capacity factors above 90%, matching the always-on draw of an AI training cluster. That combination — plus favorable political and tax treatment for advanced nuclear in the US since 2024 — is why every major hyperscaler now has at least one nuclear agreement on the books, per SMR Intel’s 2026 tracker and Data Center Dynamics.

The hyperscaler deal scoreboard

Buyer Capacity committed Reactor type / partner Target first power
Meta Up to 6.6 GW TerraPower Natrium, Oklo Aurora, Vistra, Constellation 2030-2035
Microsoft 835 MW (firm PPA) Crane Clean Energy Center (TMI Unit 1 restart, conventional) H2 2027
Microsoft Undisclosed (early-stage) Helion fusion PPA Post-2028, unproven technology
Amazon ~1.92 GW (existing plant) Talen Energy / Susquehanna PPA Already operating; expanded through 2042
Amazon Up to 5 GW (investment) X-energy Xe-100 SMR ($700M+ investment) End of 2030s
Google Up to 500 MW Kairos Power KP-FHR SMR fleet First unit ~2030, fleet by 2035

Figures compiled from SMR Intel, POWER Magazine and Utility Dive 2026 reporting; deal terms and MW figures are frequently revised as agreements are amended.

Two categories are being conflated in press coverage and buyer expectations alike: existing large-reactor PPAs (Microsoft’s Crane, Amazon’s Susquehanna) that deliver power on a near-term timeline because the plant already exists, and SMR development agreements (everything else in the table) that fund reactor design, licensing and first-unit construction with no operating precedent. Only the first category is real, contractable 2026-2028 capacity.

What an SMR actually is, and why the economics are unsettled

An SMR is a nuclear reactor under roughly 300 MWe, factory-fabricated in modules and assembled on site — the pitch is lower capital risk and faster build times than a traditional 1,000+ MWe reactor. Designs relevant to data centers today:

  • Oklo Aurora — liquid-metal fast reactor, expanded from 50 MW to 75 MW per unit to better match data center loads; targets initial deployment as early as 2029 and holds a 12 GW supply agreement with data center developer Switch, per Utility Dive.
  • X-energy Xe-100 — pebble-bed, gas-cooled, 80 MWe per unit in standard four-unit (320 MWe) plants; backs Amazon’s investment.
  • Kairos Power KP-FHR — fluoride-salt-cooled, backs Google’s 500 MW agreement with a first reactor targeted for 2030.
  • TerraPower Natrium — sodium-cooled fast reactor with integrated molten-salt storage, up to eight units (2.8 GW) planned for Meta, early 2030s delivery.
  • NuScale — the most mature by NRC design certification, but its flagship Idaho project was cancelled in 2023 after estimated capex rose past $20,000/kW; it has since pivoted to data center customers including Standard Power’s ~2 GW Ohio/Pennsylvania plan.

On cost, vendor and DOE projections put nth-of-a-kind LCOE at $60-80/MWh, falling to $45-65/MWh with the US production tax credit — in the same band as combined-cycle gas ($38-75/MWh depending on site logistics), per SMR Intel’s cost analysis. But “nth-of-a-kind” assumes serial factory production that has not yet happened anywhere; the one first-of-a-kind US data point (NuScale) came in roughly 3-4x over projection before cancellation. Treat every published SMR $/MWh figure as a vendor target, not a market price.

The realistic timeline

Phase Typical duration Status as of mid-2026
Site selection 6-18 months Mostly complete for announced deals (2024-2025)
NRC design certification / site licensing 18-24 months for pre-approved designs In progress; Part 53 technology-inclusive framework not available to applicants until late 2027
First-of-a-kind construction 3-4 years Not started for any data-center-dedicated SMR
Commissioning and ramp 6-12 months N/A
Earliest realistic first power 2029-2030 (Oklo, X-energy); 2030+ for most others

Compiled from Introl’s implementation timeline analysis and NRC process documentation. Every stage compounds: a design that clears NRC review in 2027 still faces a multi-year construction phase before it can serve load, which is why even the most advanced SMR programs land no earlier than 2029.

Regulatory wildcard: co-location and behind-the-meter rules

A parallel track matters as much as reactor engineering: how a data center connects to a co-located or adjacent nuclear plant. FERC’s December 2025 order found PJM’s tariff for “co-located load” arrangements — where a data center draws power behind the meter from an on-site generator — unjust and unreasonable, and directed PJM to write new rules covering rates, cost allocation to other ratepayers, and reliability obligations. PJM’s compliance filing (February 2026) was only partially accepted by FERC in April 2026, per Mintz’s analysis. Until this settles, even a data center sited next to an existing reactor faces open questions on transmission cost allocation and whether it can bypass grid queues entirely — the same queue dynamics covered in our power density guide.

What this means for a colocation or campus buyer in 2026

  1. Don’t budget SMR power into a lease or PPA signed before 2028. No announced deal delivers inside that window; use grid, gas, or existing large-reactor PPAs (Crane, Susquehanna) for near-term power modeling.
  2. Treat a landlord’s nuclear announcement as a balance-sheet signal, not a delivery date. A hyperscaler or developer with a Meta- or Amazon-scale nuclear commitment has demonstrated long-horizon capital access and site control — useful diligence input, not a power guarantee for your deployment.
  3. Watch the FERC/PJM co-location rules if you’re evaluating a site near an existing nuclear plant. The final tariff structure will determine whether adjacent-to-nuclear sites carry a real cost advantage over standard grid-connected capacity.
  4. For genuine 2029-2030 planning horizons, Oklo and X-energy are the furthest along on both regulatory progress and signed offtake volume — worth tracking specifically if your roadmap extends that far.
  5. Cross-check any vendor $/MWh pitch against the NuScale precedent. First-of-a-kind SMR economics have overshot projections before; model a wide cost band, not the headline nth-of-a-kind number.

Track live power and pricing benchmarks across markets in the colocation price index and browse individual facilities, including power sourcing details where disclosed, in the data center catalog.

Frequently asked questions

Are any data centers actually running on SMR power today?

No. As of August 2026, every announced small modular reactor (SMR) deal for data centers is pre-construction — the earliest operational units, from Oklo and X-energy, target 2029-2030. The one nuclear-data-center pairing already delivering power, Microsoft's Crane Clean Energy Center (the restarted Three Mile Island Unit 1), is a conventional large reactor, not an SMR, targeted for H2 2027.

How much nuclear capacity have hyperscalers committed to?

Microsoft, Google, Amazon and Meta have signed more than 13 separate agreements totaling close to 10 GW of nuclear capacity, per SMR Intel's 2026 tracker. Meta leads by volume with up to 6.6 GW across deals with TerraPower, Oklo, Vistra and Constellation; Amazon combines a ~1.92 GW Talen/Susquehanna PPA with an X-energy investment; Google holds a 500 MW Kairos Power agreement; Microsoft's Constellation deal covers the 835 MW Crane restart.

What does SMR power actually cost per MWh?

Vendor and DOE projections put nth-of-a-kind SMR LCOE at $60-80/MWh by 2030, falling to $45-65/MWh if the US Inflation Reduction Act's $15/MWh production tax credit applies — competitive with combined-cycle gas at roughly $38-75/MWh depending on site and fuel logistics. These are projections, not realized costs: NuScale's Idaho project, the only US SMR to reach late-stage development before cancellation, saw its estimated capital cost climb past $20,000/kW before the project was shelved in 2023.

Why can't SMRs solve the data center power crunch before 2030?

Three sequential bottlenecks stack up: NRC design certification and site licensing (18-24 months per the standard review, though the technology-inclusive Part 53 framework won't be available to applicants until late 2027), then 3-4 years of first-of-a-kind construction, then commissioning. Site selection for most current deals only closed in 2024-2025, which mechanically pushes first power into 2029 at the earliest even with a clean regulatory run.

What is 'behind-the-meter' or 'co-located' nuclear power?

It means a data center draws power directly from an on-site or adjacent generator rather than through the public grid, avoiding transmission queues and fees. FERC's December 2025 order found PJM's tariff for these arrangements 'unjust and unreasonable' and directed new rules; PJM's compliance filing was partially accepted and partially rejected by FERC in April 2026, so the rate structure for co-located nuclear load in the largest US grid region is still being finalized.

Should a colocation buyer factor SMR power into a 2026-2028 sourcing decision?

Not as a primary power source. No SMR will deliver commercial power to a data center inside that window under any announced deal. SMR commitments matter as a signal of a hyperscaler's or landlord's long-term capacity roadmap and balance-sheet backing, not as capacity you can contract against today — for 2026-2028 delivery, gas, grid, and existing large-reactor PPAs (like Crane) are the only firm options.

Which SMR designs are furthest along for data center use?

Oklo's Aurora (up to 75 MW per unit, expanded from an initial 50 MW design) targets initial deployment as early as 2029 and has a 12 GW supply agreement with data center developer Switch. X-energy's Xe-100 (80 MWe per unit, four-unit 320 MWe plants) backs Amazon's investment. Kairos Power's KP-FHR backs Google's 500 MW agreement, with a first reactor targeted for 2030. None has poured first concrete for a commercial data center campus as of mid-2026.

Sources

Primary sources cited in this article. Every figure links to where it comes from.

  1. SMR Intel: Every Nuclear-Powered Data Center Deal 2026
  2. SMR Intel: State of Small Modular Reactors 2026
  3. SMR Intel: SMR Cost Per kWh — Complete Economics Guide
  4. Utility Dive: Oklo reveals 75 MW reactor design, late 2027 commercial deployment target
  5. Utility Dive: Oklo 12 GW advanced reactor supply agreement with Switch
  6. POWER Magazine: Meta locks in up to 6.6 GW of nuclear power
  7. Mintz: FERC directs PJM to issue new rules for co-location of power plants
  8. Data Center Dynamics: Nuclear power and SMRs — the solution to data center energy woes?
  9. Introl: SMR Nuclear Power for AI Data Centers — Feasibility and Implementation Timeline
  10. Reason: Next-generation nuclear power can meet data center energy demand — if regulations allow it

Get Quotes

Tell us what you need — we match you with data centers in our catalog and return real quotes. Free for buyers.

We reply within one business day. No spam, no reselling your contacts.