A new Carnegie Endowment paper, published June 2 by John Pendleton and Mackenzie Schuessler, does the arithmetic the press-release cycle has been avoiding. Adding up every announced hyperscaler nuclear deal, including PPAs and direct SMR partnerships, the total comes to roughly 13 GW of potential capacity through the mid-2030s. If all of it materializes, the output is about 102 TWh per year. That figure is less than 20 percent of forecast US data-center demand through 2035.
The 13 GW splits in two halves. PPAs account for about 6.9 GW: Amazon’s 1.9 GW Talen agreement, Meta’s 1.121 GW Constellation credit deal plus 2.6 GW of life extensions, Microsoft’s 835 MW Three Mile Island restart, and Alphabet’s 50 MW Kairos/TVA contract. Direct partnerships account for the other 6.1 GW: Alphabet’s 1.8 GW Elementl arrangement (three 600 MW sites by 2035), Amazon’s roughly 5 GW X-energy framework with Energy Northwest, Meta’s 2.8 GW TerraPower commitment by 2035, and the 1.2 GW Meta/Oklo path by 2034.
The timing mismatch is the real story. The authors flag that conventional reactor builds run 10 to 15 years and no commercial-scale SMR is yet operating in the US. Meanwhile, Carnegie cites over 700 data centers under construction as of March 2026, with projected sector demand tripling or quadrupling by 2035. Quote in the paper: “Nuclear is a marathon. Signing up for the marathon does not move the sprint’s finish line.”
What this means for the supply-chain frame:
- Volume gap, not announcement gap. Hyperscalers are not under-contracting in headline terms. They are under-delivering in MWh terms, because the contracts run on 2028 to 2035 timelines while the load is showing up now. The bridge is gas, then storage, then anything else that can be permitted.
- SMR vendors are option value, not baseload. X-energy, TerraPower, Oklo, and Kairos collectively anchor about half the announced volume, but none have a commercial-scale US operating reference. Each project that slips moves more demand onto grid-scale storage and behind-the-meter generation in the meantime.
- Grid-scale storage is the residual claimant. Every gigawatt of nuclear that fails to arrive on time is a gigawatt that storage, gas peaking, or load curtailment has to cover. The Carnegie math reinforces the storage-demand thesis from the supply side, not just the demand side.
The pending FERC large-load interconnection order, expected before month-end on Docket RM26-4-000, is the next variable that will move how this gap gets covered. The Carnegie paper is the cleanest public sizing of that gap to date.