How long does a data centre grid connection take in the US?

It ranges from 36 months to 84 months depending on the market. Columbus, Ohio is the worst at 84 months; Silicon Valley, Sacramento and Portland require six years; Phoenix and Atlanta require five; and the fastest markets — Pittsburgh, Chicago, Houston and Dallas — still demand a minimum of three years. In 2007, a large industrial customer could achieve operation in under two years.

What does "small modular reactor" actually mean?

"Small" means less than 300 megawatts of electrical output, against the 1,000 to 1,600 megawatt reactors that define the current fleet. "Modular" means the reactor is manufactured in a factory, shipped to site in components, and assembled rather than built from scratch on location. A single 300 megawatt unit still powers roughly 250,000 homes, or a 250 megawatt data centre campus with margin to spare.

Which SMR vendor is furthest along?

GE Hitachi's BWRX-300 at Ontario Power Generation's Darlington site is the furthest advanced grid scale SMR construction project in the Western world. The Canadian Nuclear Safety Commission issued the licence to construct on April 4, 2025, construction began in May 2025, and commercial operation is targeted for end of 2030.

What would it cost to power a hyperscale data centre campus with an SMR?

For a 500 megawatt campus, on-site nuclear carries a 20 year NPV of $4.9 to $7.4B, against $3.5 to $5.2B for grid-connected power. The nuclear capital component alone is $2.5 to $4.0B net of ITC, plus $400 to $600M of O&M and $150 to $300M for the decommissioning fund. Against that it eliminates $750M+ of PJM capacity charges and $200 to $500M of interconnection queue cost.

How long does it take to license a nuclear data centre?

5 to 8 years from site selection decision to commercial power delivery. Pre-application engagement runs 18 to 36 months, the construction permit or combined license review targets 17 to 18 months under Executive Order 14300's mandate, and SMR construction runs 3 to 5 years.

What does a nuclear-powered data centre campus actually look like?

It is a single site pairing data halls with a dedicated on-site small modular reactor. DC Atlas's three reference designs span a 40 acre, 50 MW campus on four Oklo Aurora microreactors; a 250 acre, 500 MW campus on eight NuScale VOYGR modules; and a 600 acre, 950 MW complex on four BWRX-300 units.

A DC Atlas Frontier

Going Off-Grid: The Nuclear Frontier

The grid cannot power AI at the speed the industry needs it. We modelled what it actually costs to take a hyperscale campus off it, and what the reactor really buys you.

Built on DC Atlas's asset-level data across ~8,000 tracked facilities.

By DC AtlasPublished July 14, 2026Last updated July 14, 20266 modules · 87 min

Three ways to go off-grid, from 50 to 950 MW

One design language, three scales. The step from a 50 MW edge campus to a 950 MW hyperscale complex is a nineteenfold jump in IT load — and a very different capital, reactor and buyer profile at each stop.

Atlas Edge50 MWAtlas Sovereign500 MWAtlas Hyper950 MW

Minimum viable

50 MW

Atlas Edge

Reactor: 4× Oklo Aurora (15 MWe each → 60 MWe gross)
Capital cost: $583–848M
Campus footprint: 40 acres
Timeline to full power: 6–8 years
20-yr capacity-charge avoidance: ~$75M (50 MW basis)
Price-Anderson insurance: Reduced — site under 100 MWe ($4.5–74M)
Who builds it first: Enterprise AI, defence, or sovereign cloud

The Aurora is specified at 15 MWe because the campus is designed on that unit; Oklo's design has since scaled to a 75 MWe maximum, but the 15 MWe reactor is the intentional basis here.

Mid-scale

500 MW

Atlas Sovereign

Reactor: 8× NuScale VOYGR (616 MWe gross)
Capital cost: $4.2–5.7B (first-of-a-kind)
Campus footprint: 250 acres
Timeline to full power: Year 8 (phased from Year 5)
20-yr capacity-charge avoidance: ~$750M (500 MW basis)
Price-Anderson insurance: Full $500M — site over 100 MWe
Who builds it first: Sovereign wealth fund, Tier-1 colo, or hyperscaler

Maximum credible

950 MW

Atlas Hyper

Reactor: 4× BWRX-300 (300 MWe each → 1,200 MWe gross)
Capital cost: $16.6–19.5B
Campus footprint: 600 acres
Timeline to full power: Year 10 (Unit 1 in Year 6)
20-yr capacity-charge avoidance: ~$1.425B (950 MW basis)
Price-Anderson insurance: Full $500M — site over 100 MWe
Who builds it first: Hyperscaler (Meta / Amazon scale)

Capital ranges and the campus designs are DC Atlas modelling, not vendor quotes. The 20-year capacity-charge avoidance is an estimate: it scales the validated ~$750M-per-500 MW PJM figure by each campus's net IT load.

SourceDC Atlas analysis — the three campus designs, their capital ranges and their timelines are our own modelling, built on the module's reactor, siting and PJM capacity sources

The Frontier