Floating Nuclear Finally Has a Customer

The Atlantic Generating Station was going to be four 1,150 MWe Westinghouse pressurized water reactors on floating platforms, moored 2.8 miles off New Jersey behind a breakwater of 17,000 interlocking concrete dolosse. The platforms would be mass-produced at Blount Island near Jacksonville — a purpose-built reactor gigafactory rated at four units a year — towed to site, and connected to shore by 345 kV submarine cable.

It was designed. It was licensed. It was never built.

In late 1978 the utility cancelled all four orders, and the reason given was not safety, not regulation, and not cost overruns. A PSEG official put it plainly: "We simply will not need these units." Offshore Power Systems — the Westinghouse / Newport News joint venture behind it — closed, and sold the factory.

Forty-eight years later, Y Combinator has published a request for startups called Compute at Sea, and one of its companies, Atomarine, is building floating data center platforms explicitly designed to accept a nuclear vessel later.

The demand problem is solved. Everything else is back on the table.

What changed is the load

For most of the last decade the constraint on AI was chips. It isn't anymore. Power interconnection queues run four to seven years in many US regions, and industry estimates suggest roughly half of planned US data center projects may be delayed or cancelled in 2026 — not for capital, but for power availability, grid capacity and permitting that resolves in years rather than months.

Meanwhile a model generation is about six months. The infrastructure cycle and the model cycle have come completely unglued, and the resulting willingness-to-pay is unlike anything the nuclear industry has seen in fifty years: a counterparty with enormous capital, indifference to retail electricity prices, and an urgent need for firm, high-capacity-factor power that does not require a grid interconnection at all.

That last clause is the interesting one. A behind-the-meter offshore plant serving a colocated data center isn't competing with grid power — it's competing with not existing. That changes the economics that killed floating nuclear the first time.

Atomarine's actual bet: the berth, not the reactor

Atomarine's design is more conservative than the framing suggests, and the conservatism is the point.

They build floating data center platforms in shipyards, cooled by seawater, generating power on the platform, moored in coastal waters and cabled to shore. Nuclear is explicitly staged: as compact marine reactors come online, a nuclear vessel takes the same berth and the campus switches to steady, carbon-free power.

So the reactor is an upgrade path, not a precondition. Build the mooring, the platform, the cooling and the shore connection now against conventional generation; when marine SMRs clear licensing, a nuclear vessel takes the berth the campus was already designed around. That decouples the compute business from the reactor timeline — which, given that no US floating commercial reactor has been licensed since the 1970s, is a sensible way to avoid betting a company on the NRC's calendar.

The shipyard angle deserves attention on its own. Serial production in a controlled factory, with skilled trades and fixed-price shipbuilding discipline, addresses the failure mode that has defined Western nuclear construction for two decades: first-of-a-kind stick-built projects on greenfield sites. Offshore Power Systems understood this in 1970 — Blount Island was a four-units-a-year line. The idea was right and the market wasn't there.

The 2026 regulatory stack is genuinely new

For anyone who has watched this space, the meaningful development isn't another concept render — it's that a licensing pathway is actually forming.

NRC Part 53 took effect April 29, 2026. The optional risk-informed framework for advanced reactors matters here for a specific reason: its alternative siting criteria are, arguably, the most commercially significant provision for anyone deploying reactors to serve industrial loads or data centers.

The Marine Minerals Administration and the NRC have signed an MOU covering offshore nuclear projects on the Outer Continental Shelf, establishing how the agencies divide and sequence their reviews. Jurisdictional clarity offshore has been a standing obstacle; this is the first serious attempt to resolve it.

The NRC is developing a white paper on how existing frameworks apply to floating nuclear power plants and nuclear propulsion — building on microreactor licensing experience and on the agency's own record licensing the Offshore Power Systems design. The 1970s work was never wasted; it's precedent, and the NRC is going back to it.

Worth remembering, too, that the GAO's 1978 assessment was titled Before Licensing Floating Nuclear Power Plants, Many Answers Are Needed. Many of those answers still are.

The field, and who is betting on what

Atomarine is one entry in a suddenly crowded category, and the participants differ mainly in how much novel risk they've taken on:

Player	Approach	Novel risk
Atomarine	Shipyard-built compute platform, nuclear-ready berth	Marine reactor availability — but staged
Bluecore Energy	First barge and electric test reactor delivered to Port of Long Beach; ABS + USCG + NRC pathway	Full licensing, but hardware exists
Samsung Heavy / Sargent & Lundy	Reactor-agnostic offshore platform (MOU, July 2026)	Integration; hedges reactor choice entirely
Core Power	Feasibility study on BWXT mPower SMR for FNPPs (June 2026)	Reactor selection, no platform yet
Prodigy Clean Energy	Marine SMR platforms leased to coastal clients, conceptual design with NuScale	Business model
Seaborg	Compact molten fluoride salt reactor, 12-year refuelling, ABS feasibility statement	Gen-IV MSR maturity
ThorCon	Two MSRs on a shipyard-built hull, four-year refuelling	Same, plus jurisdiction
Maritime Fusion (also YC)	Fusion reactors for ships	Fusion

Two patterns stand out. Reactor-agnostic platforms are having a moment — Samsung Heavy / Sargent & Lundy explicitly, Atomarine implicitly via the berth. Everyone has noticed that the platform can be built and financed on a shipbuilding timeline while reactor licensing runs on its own clock, and decoupling them is the obvious hedge.

And the American Bureau of Shipping is quietly becoming the connective tissue. ABS issued Seaborg's feasibility statement, Bluecore is building to ABS requirements alongside USCG and NRC, and ABS has published its own Pathways to a Low Carbon Future: Floating Nuclear Power Data Center. Marine classification societies are used to certifying novel structures against performance requirements — a culture rather different from prescriptive land-based nuclear licensing, and possibly better suited to serial production.

What still isn't answered

For a technical audience, the open questions are the familiar ones, and they haven't moved much:

Emergency planning zones. The offshore EPZ argument is genuinely strong — a large water buffer and no resident population is close to the ideal case — but it has never been adjudicated for a commercial US floating plant. Part 53's siting criteria are where this gets decided.
Security. Armed response, exclusion zones and maritime interdiction against a moored nuclear facility are a different problem from a fenced land site, and the answers are neither cheap nor obvious.
Decommissioning and liability. Who holds Price-Anderson coverage for a hull, and what happens when a licensed vessel changes flag or owner mid-life?
The permitting assumption. The whole offshore thesis rests on maritime and environmental review being faster than grid interconnection. That is not obviously true, and nobody proposing it has demonstrated it.
Public consent. This is what actually killed the Atlantic Generating Station's political viability before demand finished it — nearby counties voted 2 to 1 against in referendums, and the New Jersey legislature refused to introduce the enabling bill. Being three miles offshore did not make the constituency disappear.

The inversion

The tidy way to put it is this. In 1978, floating nuclear had a design, a licence, a factory, and no customer. In 2026 it has a customer with effectively unlimited appetite, a serious shipbuilding industrial base, and no licensed plant.

Which of those is the easier gap to close is a real question, and the honest answer is that we're about to find out. Demand is the one thing an industry cannot manufacture for itself, and for the first time since the 1970s, nuclear isn't the one that has to.