The AI labs have run out of internet. What they have not run out of is hospitals.

Getting to those hospitals is not an engineering problem, which is why this company is more interesting than it looks.

What they built

Osseus calls itself the frontier data platform for health and biological AI. It provides medical and biological datasets, either off the shelf or built to order, along with reinforcement learning environments, and sells them to the frontier labs. It works with hospitals to get the material.

The company's own framing is worth repeating: in an era of cheap intelligence, the most important use of AI will be improving human health, globally.

Anish Mariathasan previously co-founded a compliance AI startup and did machine learning research at a Harvard lab. Ayush Patel interned twice at Apple, where he built an agentic quality assurance system, and holds a master's in computing.

Selling data to labs is now a large industry. Scale AI built the category, with Surge AI, Labelbox and Turing alongside it. In health specifically, Truveta and Datavant assemble records across hospital systems, Tempus built a business on cancer data, and Flatiron Health sold to Roche for exactly this kind of asset. UK Biobank and All of Us are the public collections everyone benchmarks against. Most of them serve drug companies and researchers. Selling directly to the frontier labs, in the shape a lab wants, is newer.

The point: the hard part is legal, and that changes who should build it

AI has created a new golden place in the supply chain, and it is data collection. That much is widely understood. What is less understood is where the difficulty actually sits.

Anyone can move files. Extraction is the easy half, and it has been easy for twenty years.

The hard half is making the data legally usable: anonymous enough that a hospital and its regulator will release it, and still detailed enough to be worth something record by record. Those two requirements pull against each other, and holding both at once is the product.

This is our reading rather than the company's own framing, and it fits what they are. The founders are a compliance AI founder and an engineer who built quality assurance systems. That is a legal and procedural pairing, not a data engineering one.

It is worth being exact about what the work is, because the word people reach for is wrong. This is not getting around the rules. For patient records the rules are serious, and a company that treated them as an obstacle would lose every hospital it has. The work is the opposite: consent that actually covers the use, removal of identifiers that survives an attempt to reverse it, access that lets a model learn without the records leaving the building, and a trail showing exactly what happened. Done properly, the answer becomes yes. That is engineering the permission rather than evading it.

Who this field is open to

Here is the part worth acting on, and it is opinion rather than anything the company has said.

This is a welcoming field for beginners, and for founders whose strength is creativity and persuasion rather than code.

The job is mostly talking to the people who hold the goldmine and convincing them you have a safe way to open it. That happens before any data moves, and it is what decides whether the company exists.

A great many technical founders simply do not have that skill, and would not enjoy the eighteen months of hospital committees, ethics boards, data protection officers and general counsel that the work actually consists of. Someone who is good with people and inventive about arrangements has a real advantage here, and it is one of the few AI businesses where that is true.

So the exercise is open to almost anyone. Ask where else a goldmine of records sits behind an institution that will never simply publish it, and what arrangement would let it be used safely. Court files, insurance claims, bank transactions, factory sensor histories, farm yields, shipping logs, school outcomes, utility meters, pharmacy dispensing, clinical trial archives. Every one of them is valuable, every one is locked, and in every case the lock is legal rather than technical.

We will see a lot more of these companies, and there is no reason not to start one.

Why it is bigger than it sounds

The short description is a company that sells medical data to AI labs. What it is really selling is permission.

  • Extraction was never the constraint. Moving files is easy, and the reason this data is not already in a model is that nobody has been allowed to use it.
  • Anonymous and individualised at the same time is the actual product. Those two pull against each other, and a company that holds both has something no amount of engineering alone produces.
  • It is open to non-technical founders. The deciding work is persuading institutions, which is a skill most AI startups never need and most technical founders do not have.

What to watch

The measure is how many hospital agreements are signed, and how long each one took. That number is the company. A falling number of months is the clearest sign that the arrangement itself has become repeatable, which is when this turns from a series of deals into a business.

The second is whether a frontier lab renews. A first purchase is curiosity, and a second means the data changed what the model could do.

The third is whether any hospital comes back for more, because an institution that reopens the door has decided the arrangement was safe, and that verdict is worth more than any contract.