The ocean has no observability stack

A survey vessel leaves port carrying about as much sensing as a self-driving car, multiplied. Multibeam, side-scan, sub-bottom profiler, an inertial navigation system, a CTD, cameras. Over a week it records terabytes about a part of the world almost nobody has seen.

Then it comes back, and someone copies the files onto a drive.

On land this got solved a long time ago. You instrument a system while it runs, keep the signals somewhere you can query, and make any moment a link. An engineer gets paged, opens a URL, and is looking at the exact request that failed ninety seconds ago. Nothing like that exists underwater.

Nine ways to write down a ping

Start with the smallest task: open the file. Qimera, one of the standard processing packages, lists nine raw sonar input formats — DB, ALL, S7K, HSX, JSF, GSF, XSE, R2SC, KMALL.^[1] Nine ways of recording the same event: a sound went out, a sound came back, here is when.

Now add the vehicle log. Add the CTD. Add the cameras. One dive, five or six files, different clocks, a different tool for each. That is why so much marine analysis happens on screenshots — a screenshot is the only artefact that reliably opens on someone else’s machine.

Why nobody fixed it

Three reasons, none of them stupid.

Software was an accessory. Sonar manufacturers sell instruments; the software existed to help sell the next one. Under that incentive a file format isn’t a standard to agree on, it is part of your product.

The market looked too small. Survey and inspection work is specialised and project-based, and for thirty years there wasn’t the volume to justify a platform. So the industry got very good desktop applications, sold per seat and installed per machine — which architecturally can never be a link.

And the standards work happened at the wrong end of the pipe. IHO S-44 sets survey orders, coverage and uncertainty with real rigour.^[2] It says how good the measurement has to be. It says nothing about how the data is stored, exchanged or reviewed. The messy engineering middle, where the hours actually go, was left alone.

What changed

The volume showed up. Offshore wind hit 83 GW installed by the end of 2024, with annual builds forecast to reach 34 GW a year by 2030.^[3] There are over 1.5 million kilometres of subsea cable across more than six hundred systems, and the AI build-out is adding more.^[4] All of it needs survey before, inspection during, and investigation when something breaks.

Autonomy got cheap, which flips the bottleneck. When one vessel ran one line, reviewing each file by hand was proportionate. When a fleet produces more in a week than a team can open in a month, the constraint stops being how much you can record.

Robotics solved the format problem next door. MCAP became the default recording format for ROS 2 in May 2023.^[5] The important part isn’t compression — it’s that the file carries the schemas needed to read it, so any tool can open it without the original stack. That is exactly what the marine formats don’t do, and subsea autonomy is increasingly built on the same robotics stack as everything on land.

And the browser got good enough. Rendering a sonar waterfall and synchronised video used to need a workstation and an install. It doesn’t now.

Why we think we win

Seabed 2030 puts 27.3% of the ocean floor mapped to modern standards.^[6] That gets read as how much is left. Read it the other way: a quarter has already been measured, and much of that measurement is unreadable to everyone except whoever recorded it. Whoever makes it openable becomes where it lives.

You don’t win a market like this with an enterprise platform and a two-year procurement cycle. You win it by killing the smallest and most universal annoyance in the job — I have a file and I can’t show it to anyone — and giving that away, with no account needed to open a shared link. Every link is the product demonstrating itself to whoever needed it.

The workspace follows, because the moment a team can see a mission they want to mark it up and compare it against the last one. And underneath sits the part that compounds: a neutral layer that normalises anything you feed it, streams while the vehicle is still down, and answers a question across a whole archive.

Neutral is the strategy, not a slogan. A layer serving manufacturers, autonomy teams, contractors and hydrographic offices at once only works if none of them has to lose. Being the thing everyone can safely standardise on is the prize.

The ocean isn’t unknowable. It’s unobserved. Unknowable is physics. Unobserved is a file format, a licence key, and the habit of waiting for the vehicle to come back.