5-7 yrs
Waiting for the grid
New data centers can wait five to seven years for a grid connection. AI companies need capacity this year.
Now
Phase 0: building the first pod. Dock trial starts December 2026
Seabed compute
AI compute, powered by the Gulf Stream.
We're building sealed pods of AI chips that sit on the seabed off Miami and run on the ocean current flowing past, day and night. No grid connection. No land. No fresh water.
No grid. No land. No water.
40 years
of continuous NOAA measurements show a remarkably stable current
~1.5 m/s
average flow at 50 m off Fort Lauderdale, every hour of the year
1/8
the failures: sealed servers on the seabed versus servers on land
24/7
current power, day and night, forecastable a season ahead
The problem
AI is running out of places to plug in.
Demand for AI computing is growing exponentially. The bottleneck is no longer chips or software. It is electricity, land and permission to build.
1,000+
Communities pushing back
More than a thousand US communities moved to curb AI infrastructure in 2026, over power bills, land and water.
47
Projects stalled
At least 47 large data center projects were delayed, withdrawn or blocked in 2026 alone.
How it works
Power comes down from the current. Results go up to the satellite.
- 01
The current
The Gulf Stream flows north a few miles off South Florida at about 1.5 m/s, day and night.
- 02
Underwater generators
Proven tethered generators fly figure-eights through the current and send power down their tethers to the seabed.
- 03
The seabed pod
AI chips sit in a sealed steel vessel filled with dry nitrogen. Seawater flowing past the wall carries their heat away.
- 04
Satellite to shore
A data-only fiber rises to a small buoy with Starlink. AI results are tiny, so they travel cheaply; electricity never has to reach shore.
Why the seabed
Chips last longer where nobody can touch them.
One-eighth the failures. Microsoft sealed 855 servers in a nitrogen-filled pod and left it on the seabed off Scotland for two years. Only 6 failed, against 8 of 135 identical servers kept on land: one-eighth the failure rate.
No corrosion
Dry nitrogen instead of oxygen and humidity.
No people
Nobody bumps, unplugs or mis-handles the equipment.
Still and steady
Cold, constant water and no vibration. The pod never moves.
Free cooling
No chillers or cooling towers. The ocean does the work.
Why the Gulf Stream
Power you can forecast a season ahead.
NOAA has measured the Florida Current with a submarine cable since 1982 and found it remarkably stable. Its year-to-year swings follow El Niño with about a three-month lead, so we can forecast each node's output a season ahead. No wind or solar farm can offer that.
- One of the fastest ocean currents on Earth, and it never stops
- A few miles off one of the largest metro areas in the US
- Home to a federal ocean-current test site off Fort Lauderdale
Progress
Building in the open.
- In progressBench and dock trialDec 2026 - Feb 2027
- NextPilot pod on the seabedMar - Jun 2027
- PlannedOn ocean-current powerJul - Sep 2027
- PlannedDR-100 demo node2028
- PlannedFirst DR-3500 node2030
Latest · Nov 2026
Architecture locked: chips in a sealed seabed pod, power from separate ocean-current generators. Pilot vessel designed from standard steel pipe and flanges.
The nodes
From a four-GPU pilot to a megawatt on the seabed.
DR-1 · Pilot
2027
- Chips
- 4 GPUs
- Vessel
- 0.6 m × 1.5 m
- Rated depth
- ~100 m
- Goal
- Prove the pod
DR-100 · Demo
2028
- Chips
- ~30 GPUs
- Generator
- 100 kW class
- Site
- Off Fort Lauderdale
- Goal
- Paying customers
DR-3500 · Standard node
2030
- Chips
- 288 latest-gen
- Power
- 1.28 MW average
- Generators
- 2 × 1.75 MW class
- Design life
- 20 years
Clean by design
Carbon-free compute, every hour of the day.
Solar stops at night and wind drops. The Gulf Stream runs around the clock, which will make our nodes one of the few ways to run AI on clean power every hour.
No gas turbines, no diesel backup, no fresh water for cooling and no land. We will publish a full lifecycle carbon figure and open monitoring data on marine life from our first deployment, rather than claim zero emissions.
Journal
Notes from the build.
Put the computer where the electricity is
AI's bottleneck is power, not chips. Here is why the answer might be a steel cylinder sitting in the Gulf Stream.
Read the post →Team
A lean team, building in Miami.
Adam Todd
Founder
Former LIFFE floor trader and early Betfair API developer who has built trading, marketing and AI products. Leads strategy, partnerships, funding and Deeprun's software.
Alex Abalionov
Founding Engineer, Fabrication and Marine Operations
25 years building and repairing complex mechanical systems, and at home on boats. Builds the pods and runs everything at sea.
We are recruiting advisors in subsea engineering and power electronics. Get in touch.
Questions
What people ask us.
Is this an underwater data center?
Close, but not quite. Microsoft's Natick project proved that sealed servers last longer on the seabed, but it was plugged into the shore grid. Our pods make their own power from the Gulf Stream and need no grid at all. We call it seabed compute.
Why not just build data centers on land?
Because the grid can't keep up. A node at sea needs no grid connection, no land, no fresh water and no local zoning fight, so capacity can come online years sooner.
How do you fix chips you can't reach?
Each pod carries spare chips and is built to be lifted and swapped. We plan to bring a pod up only for its chip refresh, about every five years, lowering a fresh pod in its place so the node keeps earning.
What about hurricanes?
The pod and the generators' anchors sit on the seabed, well below storm waves. Only the small surface buoy is exposed, and it is designed to break away and be recovered.
Is satellite fast enough for AI?
For inference, yes. AI agents run long tasks that tolerate the fraction of a second a satellite link adds, and the results are tiny. At fleet scale, nodes can share a fiber cable to shore.
What about marine life?
We survey every site to avoid coral and sensitive habitat, and we will publish monitoring data from our first deployment onward, including the small amount of heat our pods add to the current.
Who buys the computing?
AI agent and app companies, AI labs, GPU clouds and marketplaces. We will sell capacity by the hour on contracts and by the token on model marketplaces, starting with our first commercial node.
Deeprun in one sentence
Say it the way we say it.
Writing about us, introducing us, or explaining us to a colleague? Use these.
The sentence
AI compute, powered by the Gulf Stream.
Out loud
We're putting AI chips on the seabed off Miami and running them on the Gulf Stream.
The category
Seabed compute.
Three facts
- No grid. No land. No water.
- One-eighth the failures: sealed servers on the seabed failed at one-eighth the rate of servers on land.
- Power you can forecast a season ahead: 40 years of NOAA data on the Florida Current.
The words we use
- Seabed compute
- AI computing that runs on the seabed, powered by the ocean current around it. Our category.
- Pod
- A sealed, nitrogen-filled steel vessel of AI chips resting on the seabed.
- Node
- One or more pods plus the generators that power them. DR-1, DR-100, DR-3500.
- Current power
- Electricity made from the Gulf Stream's steady flow. Not tidal, not wave.
Contact
Seabed compute skips the grid queue.
We're building our first pod now, and our first commercial node is planned for 2028. Join the early-access list to get first call on capacity and help shape what we build, or talk to us about partnering.