orbit
0 m
−10 m
−30 m
−60 m
−200 m
abyss
DEPTH 0 m
Attention Deficit · Episode 4

Panthalassa

A $140M bet that the cheapest place to run AI is the open ocean. This story starts in orbit and ends on the seafloor — scroll to descend.

▼ BEGIN DESCENT
ORBIT · +550 km

First, look up: the datacenters leaving Earth

Before we get wet, meet the competition. The same power crunch pushing compute to sea is pushing it to space — and space is further along than most people think.

NOV 2025
Starcloud-1 launches with an NVIDIA H100 aboard — it ran Google's Gemini in orbit and trained a nano-GPT
$1.1B
Starcloud's valuation after a $170M Series A; the company claims "10x lower energy costs" from sun-synchronous orbit
×8
how much more productive a solar panel is in the right orbit, per Google's Project Suncatcher — two TPU satellites fly in early 2027, with 81-satellite arrays envisioned
1M
satellites in the constellation SpaceX is seeking approval for, targeting "100 kW of compute per tonne"

The catch, priced out loud: aerospace engineer Andrew McCalip calculated that 1 GW of orbital solar compute costs about $51.1 billion, versus $15.9 billion on the ground — "the margin stack and the mass tax eat you alive." Launch mass is the tax; radiation is the failure mode; nothing gets repaired, ever.

Which is exactly the opening for the other camp. Peter Thiel, announcing his bet: "Extra-terrestrial solutions are no longer science fiction. Panthalassa has opened the ocean frontier." Now we descend.

SURFACE · 0 m

Sea level: why compute is leaving land at all

AI's growth is no longer limited mainly by chips — the binding constraint has shifted to power and cooling. Microsoft's own chief executive put it plainly: AI chips are "sitting in inventory that I can't plug in."

Grid interconnects take years; cooling water is politically contested; land near fiber is scarce. So a strange race began: if the grid can't feed the datacenter, move the datacenter to the energy.

Orbit solves the energy problem with a mass tax. The ocean solves it with salt. Panthalassa's claim is that salt is the cheaper enemy: waves are the densest renewable on the planet, the water is the chiller, and a ship — not a rocket — is the repair truck. The rest of this page is that claim, examined.

A technician in safety gear walks across the white spherical head of a Panthalassa node floating in the ocean
A technician on the spherical head of a node at sea — the only part that ever breaks the surface. Photo: panthalassa.com
−10 m · THE COMPANY

Ten years underwater before you heard of it

2016
founded in Portland, Oregon as a public benefit corporation
$140M
Series B (May 2026), led by Peter Thiel personally
$210M
raised in total
120
employees building nodes near Portland

CEO Garth Sheldon-Coulson came from Bridgewater — a hedge-fund analyst, not a hardware founder. Co-founder Brian Moffat spent years on wave-energy systems at Spindrift. The Series B cap table reads like a who's-who: John Doerr, Marc Benioff's TIME Ventures, Max Levchin, Hanwha, Fortescue, Supermicro, Dylan Field — with Founders Fund returning. The same names from our defense-tech episode, now at the root of the compute frontier.

A huge white sphere under construction inside a factory bay labeled BAY 4
Bay 4 at the Oregon plant. The bet: build nodes like boats, not buildings. Photo: panthalassa.com
−30 m · THE MACHINE

Tap the parts — how one node works

Interactive — hover or tap the numbered parts
1 2 3 4 5 not to scale · sphere ≈ 30 ft · hull ≈ 200 ft
Hover a number. Each part answers one question: where does the power come from, where does the compute live, how do bits get out.

The design has almost no moving parts you'd recognize. A 30-foot buoyant sphere rides the swell; a 200-foot tube hangs beneath it. As waves lift and drop the sphere, seawater is forced through the tube, spinning internal turbines through a closed hydraulic loop. In 2025 Puget Sound testing, one node — Ocean-2 — generated up to 50 kW in decent wave conditions.

The compute rides below the waterline: hermetically sealed chip containers, chilled by the ocean itself — the cooling bill that dominates land datacenters simply doesn't exist. Queries and answers travel by Starlink. The node even steers itself using its hull shape — no engine, no fuel, no crew.

Two engineers standing inside the enormous tubular hull of a node, dwarfed by the turbine assembly
Inside the hull: the turbine assembly, with two engineers for scale. Photo: panthalassa.com
−60 m · THE MATH

Drag the slider — how many nodes is a datacenter?

Ocean-2's tested 50 kW sounds small because it is: a single NVIDIA GB200 NVL72 rack draws roughly 120 kW. The whole bet is fleet economics — thousands of cheap, unmanned, mass-produced nodes. So: how big a fleet buys how much compute?

Fleet calculator · assumes 50 kW per node (Ocean-2 tested output)

1 node100 nodes20,000 nodes
5.0 MW
fleet power
~5,000
H100-class GPUs (~700 W + overhead)
~41
GB200 NVL72 racks (120 kW ea.)
0.5%
of one 1 GW AI campus

Read it both ways. At 100 nodes, the fleet is a rounding error next to a hyperscale campus. At 20,000 nodes — the "thousands of nodes" the company describes for the open Pacific — it's a gigawatt, with no grid interconnect queue, no land, no water permits. The whole question is whether building and maintaining 20,000 sea robots is cheaper than one interconnect fight.

−200 m · THE RACE

Ocean vs. orbit, head to head

Ocean — Panthalassa
energy: waves · cooling: seawater

Serviceable by ship. Salt, storms, biofouling. Starlink latency limits it to batch inference. Energy density: the northern Pacific's winter waves are among the most energy-dense renewable resources on Earth. Demonstrated: 50 kW/node, sea trials since 2021.

Orbit — Starcloud, Suncatcher
energy: solar ×8 · cooling: radiators

No weather, near-continuous sun. But: launch mass costs, radiation-hardening, no repairs, thermal radiators instead of free water. Aerospace engineer Andrew McCalip priced 1 GW of orbital compute at $51.1B vs $15.9B on the ground — "the margin stack and the mass tax eat you alive." Demonstrated: one H100, running Gemini, since Nov 2025.

The open question nobody has answered

The orbital version has a public cost teardown. The ocean version doesn't — nobody has published the honest per-kW economics of a maintained wave fleet. Until someone does, "cheaper than land" is a hypothesis, not a number.

−1,000 m · THE RECORD

A decade of sea trials, in order

2016
Founded in Oregon
Years of quiet wave-energy engineering — no AI angle yet.
2021
Ocean-1, Strait of Juan de Fuca
First at-sea prototype of the node concept.
2024–25
Ocean-2 · Puget Sound
Up to 50 kW in decent wave conditions; the sealed-compute concept validated at sea.
MAY 2026
$140M Series B
Thiel leads; Thiel: "Panthalassa has opened the ocean frontier."
2026
Ocean-3 pilot series
Deploying to the northern Pacific this year.
2027
First commercial systems
The target. Thousands of nodes is the stated ambition.
A welder works on the massive curved steel hull of a node, sparks flying
Hull welding at the Oregon facility. Photo: panthalassa.com
ABYSS · WHAT WE DON'T KNOW

The honest unknowns

Maintenance at sea

No human within hundreds of miles. Salt, storms, biofouling, and physical security of unattended hardware — the ocean is the harshest colocation facility on Earth.

Bandwidth per node

Every token in and out rides Starlink. Fine for batch jobs; a real ceiling on interactive workloads.

Why almost nobody is talking about it

Hacker News has four Panthalassa posts ever — none above 3 points. The $140M round got 1 point and zero comments. A Thiel-led, Founders-Fund-backed frontier-compute company with a decade of sea trials is sitting in a near-total coverage vacuum. You are hearing about it here first.