Hook: The loudest AI-energy headlines promise “nuclear-powered compute.” The quieter, more useful detail is sequencing. At Savannah River Site, the U.S. Department of Energy’s NNSA selected Amentum to negotiate a phased lease for a 1 GW AI data center backed by roughly 2 GW of on-site generation—starting on natural gas, with a contractual pathway to advanced nuclear later (World Nuclear News; Amentum, July 20, 2026).
Key takeaway: This is not “SMR flips a switch tomorrow.” It is a gas bridge with a nuclear option, sized so the campus does not dump a gigawatt onto ratepayers while reactors finish licensing, fuel, and financing.
What actually happened
NNSA chose Amentum as the counterparty to negotiate—not yet to build under a final lease award. Official and company statements are careful on that point: selection for negotiations does not equal a signed lease, and any deal still faces permitting, safety, security, and other federal approvals (Nuclear Engineering International; Amentum release).
The scale numbers are the engineering hook:
- ~1 GW of AI data-center load.
- ~2 GW of dedicated on-site generation capacity.
- Generation plan: natural gas bridging to nuclear / advanced nuclear (NNSA framing via WNN / Amentum).
Why 2 GW for 1 GW of IT? Public sources do not lock a single answer. Plausible engineering reasons include headroom for cooling and balance-of-plant, phased IT build-out, export to the grid, redundancy, or some mix. Treat the exact split as 待核 until the lease and interconnection studies are public.
SRS itself is the site advantage: a large federal nuclear reservation with existing nuclear culture, security posture, and industrial land—one of several DOE sites previously flagged for AI-plus-power co-location (WNN summarizing the April 2025 federal-site shortlist and later RFP path).
Engineering mechanism: time-to-power vs time-to-nuclear
AI halls want high capacity factor. They do not wait politely for first-of-a-kind reactor schedules. Three clocks collide:
- 1) IT schedule
Hyperscale campuses sell tokens and contracts on calendar years, not decade-long nuclear FOAK timelines. - 2) Grid interconnection
Elsewhere, high-voltage connections can stretch years. A Centrica / Ceres report highlighted UK HV waits of roughly 5–15 years versus SOFC installs potentially under 12 months—different geography, same constraint: time-to-power (Energy Live News / Energy Jobline, Sept 17, 2026). - 3) Nuclear delivery
Advanced reactors still need design certification / licensing maturity, fuel-cycle readiness, and project finance that survives schedule risk. Analysis of the SRS structure correctly frames gas as the bridge and NRC predictability as a delivery variable (SMR Intel, Sept 2026).
On-site gas generation answers clocks (1) and (2) with known EPC playbooks. The nuclear pathway answers the long-term carbon and fuel-price story—and gives lenders a story only if the off-take is real. A near-continuous 1 GW AI load is, in project-finance language, an unusually clean anchor tenant: high utilization, long infrastructure horizon, behind-the-meter or dedicated-supply economics that avoid merchant power markets (SMR Intel framing).
Ratepayer-protection logic (policy + power engineering):
- Pair new AI load with dedicated on-site MW so the campus does not simply queue ahead of utility upgrades.
- Oversize generation (~2 GW vs ~1 GW IT) leaves room for losses, cooling, contingency, and possible grid support—exact allocation still 待核.
The misconception worth killing
Misconception: “If the press kit says nuclear AI campus, day-one electrons are nuclear.”
Reality check: Day-one electrons follow the technology that can be permitted, financed, and synchronized first. At SRS that is explicitly a gas bridge. Nuclear is the destination—if licensing, fuel, and financing close inside the campus ramp window.
Limits / 待核:
- Lease negotiation ≠ construction NTP. Schedule slides are normal until NEPA, security, and interconnection clear.
- “Advanced nuclear” is not yet a named SMR design + firm COD in the public selection notices we cited.
- If nuclear licensing slips, the project can remain a large gas plant with a nuclear option—commercially rational, policy-disappointing.
- Cross-region SOFC “<12 month install” figures are Centrica/Ceres report claims for a different market—use as time-to-power context, not as SRS design basis.
Takeaway judgement
If you design AI power architecture in 2026, read Savannah River as a template, not a miracle:
Separate three questions: What energizes the racks in year 1? What keeps ratepayers whole? What must be true for nuclear to actually arrive?
- Year-1 power will usually be gas, grid islands you already own, or fast-deploy generation (fuel cells, turbines)—not FOAK reactors.
- Dedicated on-site MW is becoming a political and engineering requirement, not a nice-to-have slide.
- SMR / advanced nuclear only “wins” when licensing clocks and fuel supply beat the gas-phase extension risk.
That is less sexy than a nuclear AI billboard. It is how gigawatt campuses actually get built.
Related: @TheEngineeringCore-v · @Vkinng · vkinngworld.blogspot.com
Sources: Amentum IR (2026-07-20); World Nuclear News (2026-07-21); Nuclear Engineering International; SMR Intel (2026-09); Energy Live News / Centrica–Ceres time-to-power coverage (2026-09-17). Figures marked 待核 where public notices leave allocation or COD unspecified.
—— The Engineering Core
