Data Center On-Site Nuclear & SMR

This is the first coverage this KB has of on-site nuclear or small modular reactor (SMR) power for AI data centers — before this cycle, the electrification KB had zero prior sources on the topic despite it being a live, fast-moving hyperscaler capital-deployment story since 2024-2025. A continuously-updated industry-intelligence tracker puts committed hyperscaler nuclear/SMR capacity at 9.8 GW across 13 deals and 7 buyers as of the May 2026 aggregate cut (1 project under construction, 12 in planning; underlying dataset last verified 2026-07-06).

The tracker's central finding is a clean split between two nuclear pathways with very different speed-to-power profiles: restarting an existing, previously-licensed reactor is structurally faster than building a new SMR fleet. Microsoft's Crane Clean Energy Center — the restart of Three Mile Island Unit 1, contracted with Constellation Energy for 835 MW under a 20-year, $16B power purchase agreement — is the fastest deal in the tracker, targeting full commercial power by 2027 after a FERC transmission-waiver approval cleared June 1, 2026. By contrast, Meta's up-to-6.6 GW multi-developer portfolio (spanning TerraPower's Natrium fleet, Oklo's Aurora SFR Power Campus, and PPAs with Vistra and Constellation) is the largest single commitment in the tracker but has the longest horizon, 2032-2035 — new-build SMR construction, even at hyperscaler capital scale, does not compress to a 2020s timeline.

This matters directly for how nuclear/SMR should be read against the near-term power-constraint story in AI Data Center Grid Interconnection: nuclear is not, and will not soon be, an answer to the interconnection-queue and time-to-power problem through 2028. Bloom Energy's concurrent industry survey (same discovery cycle) shows fuel cells (47%) and reciprocating engines/turbines (38%/33%) leading current onsite-technology evaluation, with nuclear barely registering as a near-term option — consistent with the deal tracker showing only one restart-class deal (Microsoft/TMI) landing before 2028, and every new-build SMR fleet (Meta/TerraPower, Meta/Oklo, Amazon/X-energy, Google/Kairos) clustering in the 2030s.

Deal Table (as of May 2026 aggregate cut; tracker updates weekly beyond this cut)

Source: Nuclear/SMR Data Center Deal Tracker, SMR.INTEL.

Key Claims

• 9.8 GW of nuclear/SMR capacity committed across 13 deals and 7 hyperscaler buyers, as of the May 2026 aggregate cut. Evidence: moderate — single industry-intelligence tracker (analysis-grade, not peer-reviewed or regulatory-filed); treat as a snapshot, not a current total (Deal Tracker)
• Restart-of-existing-reactor deals deliver power fastest: Microsoft/Constellation's Three Mile Island Unit 1 restart targets 2027, the fastest date in the tracker, following a FERC transmission-waiver approval (cleared June 1, 2026). Evidence: moderate (Deal Tracker)
• New-build SMR fleets cluster in 2030-2035, regardless of buyer capital scale — Meta's TerraPower Natrium fleet, Oklo Aurora Power Campus, Amazon/X-energy Cascade facility, and Google/Kairos Power's 500 MW fleet all land in this window. Evidence: moderate (Deal Tracker)
• Meta's up-to-6.6 GW multi-developer portfolio is the largest single commitment but has the longest timeline (2032-2035) — illustrating the tracker's core trade-off: existing-reactor restarts deliver power fastest, new-build SMRs offer scale but not speed. Evidence: moderate (Deal Tracker)
• Nuclear/SMR is not a near-term (pre-2028) answer to the AI-datacenter power constraint — only one deal (Microsoft/TMI) lands before 2028; this is corroborated by Bloom Energy's concurrent survey, where nuclear barely registers against fuel cells (47%), reciprocating engines (38%), and mobile turbines (33%) in near-term onsite-technology evaluation. Evidence: moderate, cross-source corroboration between an analysis-grade deal tracker and a survey-grade industry report (Deal Tracker; Bloom Energy 2026 Report)

Benchmarks & Data

• Aggregate committed capacity: 9.8 GW across 13 deals, 7 buyers (as of May 2026 cut) (Deal Tracker)
• Fastest deal: Microsoft/Constellation Crane Clean Energy Center (TMI-1 restart), 835 MW, $16B 20-year PPA, target 2027 (Deal Tracker)
• Largest single-buyer commitment: Meta, up to 6.6 GW across 4 developers/structures, target 2032-2035 (Deal Tracker)
• Amazon/X-energy Cascade Advanced Energy Facility: 960 MW (12 Xe-100 HTGR units), $700M investment (Deal Tracker)
• Projects under construction vs. planning: 1 vs. 12 (Deal Tracker)
• Tracker last verified: 2026-07-06; aggregate totals as of May 2026 (Deal Tracker)

Open Questions

• HALEU (high-assay low-enriched uranium) fuel-supply-chain risk is not quantified in this source. Earlier (2025) industry writing flags Russian-supply linkage for HALEU enrichment; this cycle's tracker does not address it. This is a named knowledge gap (see frontier.md).
• Many capacity, cost, and timeline fields in the deal table are marked "Undisclosed" (most of Meta's individual sub-deals, Amazon's Talen and Dominion deals) — true committed capacity and cost may be understated.
• The tracker is an industry-intelligence blog, not a peer-reviewed or regulatory-filed source — deal terms should be cross-checked against primary company disclosures (10-Ks, PPA filings) before use in any published report.
• Will any 2026-2027 restart or new-build slip past its target date? The Microsoft/TMI restart (2027) is the first real near-term test of nuclear's credibility as an AI-power source on the timeline hyperscalers are claiming.
• The tracker's "as of May 2026" aggregate will already be stale relative to its live (weekly-updated) state by the time this KB topic is next refreshed.

Related Concepts

• AI Data Center Grid Interconnection — The near-term (through ~2028) power constraint that nuclear/SMR does not solve; behind-the-meter gas and fuel cells are the technologies actually moving at AI-industry speed in this window, per Bloom Energy's survey. Nuclear/SMR is the 2030s-and-later generation-side answer to the same structural problem.

Backlinks

Pages that reference this concept:

• AI Data Center Grid Interconnection — cross-links here as the generation-side, multi-year-lagged counterpart to interconnection/curtailment mechanisms

Changelog

• 2026-07-23 — New concept, first KB coverage of on-site nuclear/SMR for data centers. Compiled from 1 source (SMR.INTEL deal tracker, analysis-grade, snapshot as of May 2026, last verified 2026-07-06). Full 13-deal table reproduced; headline distinction is restart-of-existing-reactor (fast, 2027-2028) vs. new-build SMR fleet (scale, 2030-2035). Cross-linked to the new AI Data Center Grid Interconnection concept as the near-term counterpart this concept does not address.