Convergence Watch: The Self-Driving Stack Found a Job. It’s Watching Salmon.

Buried inside US patent 12,077,263, granted to TidalX AI on September 3, 2024, is a phrase that should make every laid-off Waymo perception engineer sit up: an “anti-fish-startling model evaluation engine.” The patent claims a method for training a neural network on logs of past robot actions, the conditions under which they happened, and whether the nearby fish flinched. Once trained, the model rewrites the robot’s plan to keep the salmon calm.

If you swap the word “fish” for “pedestrian,” you have a self-driving car.

That swap is happening at scale. Ten years ago, the people writing this kind of code worked on autonomous vehicles, robotic warehouse pickers, and drone delivery. Today, the same stack of calibrated cameras, real-time object detection, behavior models, and actuator-level safety policies is being shipped into floating cages in Norwegian fjords, Scottish lochs, and Patagonian inlets. The biggest winner so far is the company that built it inside Alphabet.

The moonshot that made it out

In July 2024, Alphabet’s X graduated a project named Tidal as an independent company under the name TidalX AI, with funding from Perry Creek Capital and two Norwegian funds connected to the family that controls SalMar, one of the world’s three largest salmon producers. SeafoodSource covered the launch; TechCrunch called it the moonshot factory’s latest spinout. The pitch was specific: an underwater camera and software platform that monitors fish growth, detects sea lice, and runs feeders autonomously.

The company’s own numbers, taken from its public materials in May 2026, are the giveaway: more than 700 pens instrumented worldwide, more than 1.5 petabytes of underwater video processed, more than 50 million individual fish tracked through their grow-out cycle. That’s a perception dataset comparable to mid-2010s autonomous-driving fleets, except every “scene” is forty meters underwater and the things being tracked don’t follow lane lines.

The patent record matches the marketing. As of May 9, 2026, X Development and TidalX have together been granted at least 51 US utility patents whose titles explicitly involve aquaculture, fish, sea lice, underwater sensing, marine pens, or ectoparasites. Aquabyte, the venture-backed San Francisco competitor that PR Newswire covered when it raised a $10 million Series A in 2019, has three. Innovasea, a longtime aquaculture-equipment supplier, has seven. The single-firm gap is enormous.

Read the claims

The titles tell you what kind of company TidalX has been quietly becoming.

US 12,131,568, granted October 29, 2024, claims a method for estimating fish weight by selecting high-quality “tracks” of fish swimming past a stereo camera and computing a representative weight. This is the same multi-object tracking problem self-driving cars solve every fifteen milliseconds for cars and pedestrians, just retrained on bodies that are silver, hydrodynamic, and prone to schooling.

US 12,078,533, granted September 3, 2024, repurposes the underwater camera itself as a light meter: it points the lens upward, measures ambient light passing through the surface of the water, and uses the reading to drive other actions in the pen. It’s a classic sensor-fusion trick from autonomous robotics, applied to a context where every additional cable into a salt-water cage is an expensive failure point.

US 12,137,674, granted November 12, 2024, is even more clever. It estimates the water current inside an enclosure by measuring how feed pellets drift between video frames. Optical flow on pellets, in other words. The same technique that lets a Tesla guess how fast it’s moving from a forward-facing camera lets a salmon farm guess whether the cage is in a riptide.

And US 12,185,701, granted January 7, 2025, claims a self-calibrating system for using ultrasonic pulses to knock parasites off fish, with calibration targets and feedback loops borrowed wholesale from industrial robot vision. Sea lice cause the salmon industry more than a billion dollars in annual losses, by Norwegian government estimates that SeafoodSource has tracked. A vision system that points a transducer at the parasite, fires, and verifies the hit is the agricultural version of laser weeding, and a far cheaper alternative to the chemical and thermal treatments that currently dominate.

The literature got there earlier

The academic side has been pointing at this convergence for years. Across OpenAlex’s 357 million indexed scholarly works, papers tagged with both “aquaculture” and one of “deep learning,” “computer vision,” “machine learning,” or “convolutional neural network” went from 8 in 2014 to 445 in 2025, a fifty-five-fold rise in eleven years.

The most-cited single paper in that body of work, by a team out of China’s Dalian Ocean University and published in Computers and Electronics in Agriculture in 2021, modifies YOLO-V4 to detect uneaten feed pellets in underwater video in real time. Hitting that target matters: feed is the largest single cost line for a salmon farm, and uneaten pellets are simultaneously a profit leak and an environmental liability that Norwegian regulators increasingly fine. The paper has been cited 251 times. A 2022 follow-up from a separate lab, “YOLO-Fish,” generalized the approach to species detection in dirty water; a 2022 Reviews in Aquaculture survey concluded that deep-learning detectors had become the default tool for visual aquaculture tasks. The methods being grafted onto fish are the same methods being yanked off the autonomous-driving stack: YOLO, Faster R-CNN, ResNet, DeepSORT.

Why this swap, why now

The conditions are textbook adjacent-possible. The aquaculture industry has a labor shortage, a parasite problem that is genuinely catastrophic to its economics, and a regulator pushing it toward closed-containment systems that work only with continuous monitoring. The autonomous-driving industry, after a brutal 2022 to 2024, has a glut of perception engineers, a glut of edge-deployable computer-vision models, and a hardware supply chain that knows how to make ruggedized stereo cameras at automotive volumes. Salmon farmers have every reason to buy what AV teams already know how to build.

The patent geography is the most telling artifact. X Development LLC, the entity that filed most of the foundational TidalX inventions, is the same address that filed Project Wing’s drone-delivery patents and the early Waymo robotic-arm work. Patent 11,633,848, granted to X Development in April 2023, is titled “Independent pan of coaxial robotic arm and perception housing.” Read the claims and the specification reads almost identically to a manipulator on a delivery robot, except its job is to point an underwater camera at a particular fish.

The competitive map says the same thing. Aquabyte’s 2025 grant US 12,418,721 claims a learned policy for positioning the underwater camera based on where the fish actually are: reinforcement learning on a sensor mount, the same idea that drives active perception in mobile robots. Merck’s animal-health subsidiary Intervet has filed parasite-detection patents using the same multi-spectral imaging tricks first developed for medical computer vision.

For an R&D director, the question is not whether autonomous-vehicle perception will reach fish farming. It already has. The question is whether the data moat of 1.5 petabytes of underwater video and 50 million annotated fish will let one Alphabet spinout compress a $19 billion farmed-salmon industry, by recent industry-analyst estimates, into a software business with one or two dominant platforms. Norwegian and Chilean regulators, the only authorities whose paperwork ultimately decides which sensor stack a pen is allowed to use, are watching. So is every consumer of farmed salmon, which is to say most of us.

The robot is learning not to startle the fish. The fish, in return, are training the robot.

Method note

Patent counts are from a snapshot of the USPTO bulk grant feed covering 9.3 million issued utility patents through May 5, 2026. Aquaculture-related grants were identified by full-text search and assignee match; assignee counts collapse common variant spellings. Scholarly trend counts are from a local index of 357 million OpenAlex works, restricted to records whose full-text fields match both an aquaculture term and a deep-learning or computer-vision term. Sea-lice cost figures are from the Norwegian Veterinary Institute and the Marine Resource Economics literature, as summarized by SeafoodSource. Industry-revenue figures are 2024 farmed-salmon estimates aggregated from analyst reports cited by FERN and Statista. TidalX deployment statistics are taken from the company’s public materials and IEEE Spectrum’s recent profile; readers should treat self-reported pen and dataset counts as the company’s own numbers.