Key Takeaways

  • Underwater acoustic telemetry caps out at roughly 10 kilobytes per second over five kilometers, making remote cloud processing physically impossible beneath the surface.
  • Ulysses splits maritime robotics into two vehicles: Leviathan, an autonomous surface craft with a 2,000-nautical-mile range and Starlink, and Mako, a modular underwater drone.
  • To run computer vision and localization models without a live data umbilical, the Mako vehicle carries an onboard NVIDIA 5090 GPU and an Intel CPU inside its pressure vessel.
  • Saltwater corrosion is managed by anodizing and coating metal surfaces, avoiding the bespoke, cost-heavy metallurgy common in legacy defense contracts.
  • O'Brien targets commercial applications like offshore wind inspection and seagrass restoration to drive production volume and undercut defense prime contractors on cost.

The 10 Kilobyte Bottleneck Forces Compute to the Edge

Most robotics companies take high-bandwidth communication for granted. A wheeled delivery robot or a self-driving car can compress video, ping a cell tower, and stream telemetry back to an operations hub in real time.

That mental model dies five meters underwater. Radio waves do not travel through saltwater. Sound waves do, but acoustics give you almost nothing to work with. As O'Brien explained to John Collison and Chris Sacca, “You can do like 10 kilobytes per second. Over like five kilometers.”

At 10 kilobytes per second, you cannot stream low-resolution video. You cannot offload sensor fusion to a server rack on land. You cannot keep a human in the loop to bail out a confused autopilot. If a robotic sub cannot make up its own mind immediately, it gets lost or hits a reef.

Ulysses solved this by stuffing desktop-grade silicon into the vehicle itself. “We run a full NVIDIA 5090 onboard with like an Intel CPU,” O'Brien said. “And that's very important to our autonomy stack.” The sub processes its own sonar sweeps, interprets optical sensors, and runs local perception models without asking permission from shore. When bandwidth drops to zero, the edge machine has to be as capable as the server room.

Leviathan and Mako Divide the Ocean into Two Problems

Trying to build a single vehicle that travels thousands of ocean miles and dives deep into target zones creates an engineering trap. Long surface transit requires fuel capacity and satellite dishes. Deep diving demands heavy pressure housings and tight hydrodynamics. Mixing both sets of constraints into one chassis guarantees a vehicle that does neither job well and costs millions of dollars to assemble.

O'Brien split the physics into a parent-child architecture. “From first principles, if you want to build the ocean robot company, you need a few things. You need full XYZ access on the ocean as a domain,” he noted. “So we developed this autonomous surface vehicle called Leviathan, 2,000 nautical mile range. And that goes and drops the kids off to school, so to speak, to go do what they want.”

Leviathan stays on the surface, sucking down satellite telemetry through Starlink and running long diesel or hybrid legs across open water. Once it reaches the survey zone, it deploys Mako, the sub-surface drone. Mako dives, completes its mission using modular sensor packages, and returns to the surface craft.

The hardware materials follow the same pragmatic approach. Instead of building hulls out of exotic titanium alloys that inflate build budgets, Ulysses uses standard metals protected by surface treatments. “Corrosion and pressure are huge external forces,” O'Brien pointed out. “But corrosion, you can get around it by coating, doing anodizing the surface of the metal, which gives it some resistance.” By solving corrosion with standard surface chemistry and keeping compute self-contained, Ulysses can mass-produce submersibles at commercial prices rather than defense-contractor markups.

What to Do With This

Audit your system architecture for hidden bandwidth dependencies. If your field units rely on constant connectivity to execute core logic, force your engineering team to run tests under simulated 10-kilobyte caps this week. Shift every safety-critical decision and perception loop directly to local edge compute so an uplink failure never halts an operation.