Tech companies keep announcing nuclear deals for their data centres. This bets almost none of it is switched on this decade.
80%likely
Well sourcedNext check
Watching
Why it is different
Not whether small reactors work, but whether any meaningful amount is actually generating power for data centres before 2030.
Why it matters
If the nuclear answer lands in the 2030s, the power problem has to be solved some other way first.
Next check · 31 Dec 2026
After the August 2026 hearing, does the regulator issue the construction permit for TVA’s Clinch River reactor by 31 Dec 2026?
How we scored it ·
from how similar cases went
If right, how big3/5How overlooked4/5Upside vs downside2/5Source trust4/5Independent sources4/5
The exact wording we score
At 31 Dec 2030, aggregate new US advanced-reactor capacity continuously supplying a commercial data-center load is below 1.0 GW. Conventional reactor restarts, test criticality, demonstration output and capacity contracted but not operating do not count. Resolves YES at 0.999 GW or less; NO at 1.000 GW or more.
PROOF
Evidence and provenance
The links below are the evidence recorded when this forecast was scored. Each note states source quality or bias; a citation is not an endorsement.
Recorded sources · 4 citations / 4 independent origins
2026-07-28 | sell-side: nuclear-for-AI narrative remains promotional; mainstream press: signed deals are commonly presented as near-term capacity; priced market: reactor-startup equities reflect substantial optimism; specialist: licensing and construction schedules cluster meaningful delivery after 2030 | E4
Revision history4 entries
Probabilities and interpretations are never silently overwritten. This is the append-only record of what changed and why.
The current claim now implements the log's requested repair: an exact 1.0 GW boundary, continuous commercial-load delivery, and explicit exclusions for restarts, test criticality, demonstrations and non-operating contracts. The title no longer makes the unscored “largely narrative” assertion.
(2nd pass) — corrgroup refined ai-capex → ai-capex-physical. Assumption tags added: this signal is short A2 (announced dates hold) while five medicine signals are long it — see assumptions/README.md. Cross-reference conflict logged: uranium-enrichment-swu-bottleneck names Centrus (LEU) as its only US-listed pure-play, whose growth case rests on Piketon HALEU for exactly the advanced-reactor programs this signal rates 80% likely to be 'largely narrative' through 2030. The book holds both positions and had never noted the tension.
v2 rescore. P held at 80, but the bull case's evidence base is partly falsified. The DOE Reactor Pilot Program met its EO-14301 4-Jul-2026 criticality goal: Antares Mark-0 (2026-06-04, INL), Valar Atomics Ward 250 (2026-06-18, Utah — reached 10 kWt and converted to electricity), Deployable Energy Unity 1-MWe (2026-07-01, INL); Kairos broke ground on Hermes 2 in Apr 2026; NRC issued TerraPower's Natrium construction permit. This falsifies the Bulletin of the Atomic Scientists line this file leaned on ('no SMR projects are currently under construction in the United States') — that source is downgraded 4→3 in the registry. Valar Atomics + Nvidia also announced a 30 MW nuclear-powered AI data center in Utah (~2026-07-01). Quantitatively this is ~3% of the 1 GW threshold and all units are kW-to-1-MWe demonstration scale, so the numeric claim holds — but 'largely narrative' is now too strong a framing, and the resolution language should be tightened to continuous MW delivered to a commercial data-center load, excluding demo/test criticality.
created from research sweep; Bulletin source spot-verified. Watch for restart-relabeling that could muddy the falsification test.
Method challenge
Adversarial review record
Admitted
Structured internal role review tied to this frozen claim. It is not independent human peer review. Independent specialist review not yet performed
technical-domain specialist
delivery and operations reviewer
policy or standards reviewer
forecasting-method reviewer
Strongest specialist challenge
The negative forecast can look easier than it is because microreactors and DOE-authorized pilot projects may reach criticality outside the standard NRC commercial-reactor pathway. Conversely, test heat or electricity must not be mislabeled as continuous commercial supply.
Outside view
First-of-a-kind nuclear projects have a poor record of meeting announced commercial dates. Fewer than four and a half years remain, and most named commercial projects are still in pre-application or early licensing, supporting a high probability below the 1 GW threshold.
Causal rival
Conventional reactor restarts, gas generation, or power-purchase contracts can satisfy data-center demand and make the narrative wrong without falsifying this advanced-reactor count. A fast fleet of small DOE pilots is the direct causal rival to the forecast itself.
Measurement risk
NRC status proves licensing and EIA status proves generation, but neither alone proves electricity continuously serves a commercial data-center load. Adjudication needs a named operating unit, capacity, commercial operation evidence, and a documented data-center offtake relationship.
Residual risk
The legal boundary between demonstration and commercial operation may evolve, especially for DOE sites. The exact 0.999 versus 1.000 GW rule is sound, but each qualifying unit will require a transparent evidence dossier rather than contract-announcement totals.
COLDTELL