A single Nvidia H100 GPU has run customer AI workloads from a floating data center in San Francisco Bay, drawing power from a 20kW solar array. NetworkOcean, the startup behind the trial, claims this is a first for the industry.

The company has not disclosed which customer used the GPU. Footage released by NetworkOcean shows 12 floating solar panels linked to a container aboard a ship, with the GPU server housed inside.

Founded in 2024 by Sam Mendel and Eric Kim, NetworkOcean began as an underwater data center venture. It has since pivoted to a floating design that runs on solar energy at sea.

What does the CEO say about deployment speed?

CEO Sam Mendel frames the project around power deployment speed rather than compute density. Writing on LinkedIn, he said: "Our future is bottlenecked by the speed at which we can deploy power."

He continued: "We're building floating solar arrays powering data centers at sea. This will soon be the fastest way to deploy new compute, skipping years of land, grid connections, and traditional infrastructure delays."

Mendel stated the company moved from "design to deployed in the water" in 53 days. He added: "This will soon become the quickest way to deploy new power, powering the most efficient and sustainable data centers on the planet."

Why did the startup abandon underwater data centers?

NetworkOcean is a spin-out from the Y Combinator incubator. Its original plan placed servers underwater, but that approach hit regulatory trouble. Reports claimed the firm lacked the correct permits for a trial deployment in the Bay Area.

The floating solar concept sidesteps the permitting problems that stalled the underwater design while keeping the core idea of marine-based compute.

How large can the system scale?

The company's roadmap, published on its website, outlines an aggressive expansion:

• A 1MW solar array deployed at sea next year
• Up to 1GW of power for compute by 2030

The jump from the current 20kW trial array to 1MW represents a 50-fold increase in generation capacity. Reaching 1GW by 2030 would require scaling up another three orders of magnitude from that milestone.

The current trial involves one GPU. A 1GW power budget would support thousands of accelerators, assuming consumption in the range of today's high-end AI hardware.

What does this mean for the compute-at-sea sector?

NetworkOcean is one of several ventures pursuing compute at sea, a sector that has attracted attention as land-based data centers face power and grid connection constraints. Mendel's argument centers on time: grid interconnection queues and land acquisition can add years to traditional data center projects, while a floating array can theoretically deploy in weeks.

The 53-day design-to-deployment timeline is the company's key evidence for that claim. Whether the model scales to megawatt and gigawatt levels — and whether it can secure the permits that derailed the underwater predecessor — remains untested.