A chemist, a lawyer, and an engineer walk into a bar. It sounds like the setup to a joke. For me it was a summer evening, in my own backyard, under a canopy, with a drink in hand.
The chemist is a friend of many years, a medicinal chemist by training. He has lived the whole arc of a drug: basic bench chemistry, the clinic, and lately process development, the quiet craft of turning a molecule into something you can actually manufacture. The engineer is his wife. She works at one of the big security companies, mid-career and deep in it, on the team that keeps data secure in the cloud. The lawyer is me.
We were not solving anything that night. We were talking about the thing everyone is talking about: what happens when artificial intelligence meets life sciences. And somewhere in the conversation I realized the three of us were the whole story. The science, the machine, and the law. Take any one away and the future we were describing does not get built.
Let me walk you through it.
Start with the machine, because that is where the noise is.
The AI we grew up with was narrow. Think of the old Google search box. One model, one job. Rank the pages, flag the spam, recommend the video. It did not understand anything. It matched patterns. Today's models are different in kind. Train a system on a huge slice of human knowledge to predict the next word, over and over, and something surprising happens. It can write, reason, and code across almost any subject. That knowledge lives in the weights, the billions of numbers set during training. The weights are the asset.
Underneath it all sits a stack. Chips and data centers at the bottom. Foundation models above them. Tools and applications on top. Most companies that call themselves "AI companies" are that top layer, a clever front end built on someone else's model. Easy to build. Easy to replace, unless it owns something the model does not. Because the value lives one layer down, in the model and the data, and those belong to someone else.
A company that runs on someone else's model, with data that blends into everyone else's, owns almost nothing durable.
It is renting its advantage. The winners will own theirs.
That is the whole game, and it comes down to what you own.
Which is where the lawyer earns his seat at the table.
Strip AI of the hype and it runs on old, durable law. Where did the training data come from, and did anyone have the right to use it? That is provenance, and provenance is contracts. Who owns what the model produces? Property and intellectual property. What can cross a border, and what gets you sued when it does? Privacy, licensing, jurisdiction. When companies that compete pool what they know, where is the line? Competition law. When people build these systems inside companies, what do they owe, and to whom? The plain old duties of loyalty and care.
None of this is new. The subject is new. The law is not. My foundation has always been transactional law, and transactional law is exactly the right tool for a world where value moves from one bucket to another and someone has to make sure it arrives clean and stays owned. The technology will change every six months. The law underneath it has held for a century. Build on the part that lasts.
Here is what most people miss. In life sciences, data is an asset only if the rights around it hold up when the company changes hands. Most AI-biotech data is licensed access, not owned. Licenses carry the terms that bite at exactly that moment: consent to assign, change of control, field limits, and the patient consent underneath all of it. A buyer's counsel reads those before the science.
Whoever holds the pooled-data rights, and a clean method for proving where the data came from, owns the real thing.
Everything else is packaging.
So what do you actually build?
The answer is a federation: separate companies sharing one data-and-model layer, with the legal right to pool and cross-use what each of them generates. The structure is the product.
But be honest about the hard part. Two good datasets do not automatically make one better dataset. If they come from different labs, instruments, or patients, pooling them can simply teach a model which company produced the sample, not the biology. The question is never "does everyone have data." It is "does the combined data compose into one loop that compounds." Almost no one asks that before they combine. A lawyer who does earns his fee.
Then put the three people in the same room. Literally. Palantir did not sell its software over the wall. It sent its own engineers inside the customer to build next to the people who owned the problem. Do that in life sciences: engineers at the bench, next to the biologists, building against real experiments and real data. Then do it again in the conference room, because that is where these ideas actually mature, and where they get monetized. The scientist, the engineer, and the lawyer, at the bench and at the table.
Point all of it at one place first. My choice is metabolic disease.
It is the largest and still-widening market in medicine. And the gaps that matter most, who responds, who keeps the weight off, whose liver actually improves, all turn on long-run human data that no single company legally controls. That is precisely the data a federation is built to assemble and own. The gap in the science and the moat in the law turn out to be the same point, seen from two sides.
A chemist, a lawyer, and an engineer walk into a bar.
The joke works because we assume they have nothing to say to one another. The truth is the opposite. The next generation of life-sciences companies will be built by exactly these three, at one table, owning their stack, owning their data, and structuring the whole thing so it holds up years later when a buyer's counsel finally comes to look.
Same three people. Same problem. One table.
If that is a table you are trying to build, I would like to hear from you.
A month after this went up, the federation showed up in two forms, and they are almost opposites.
In July, GSK, Boltz, Cradle, Dyno Therapeutics and A-Alpha Bio launched the Atlas Consortium. Members pay a subscription, A-Alpha runs the experiments, and every quarter each member receives the same pooled dataset of antibody-antigen measurements. The data moves. Everyone gets an identical copy. Meanwhile AbbVie, Johnson & Johnson, Bristol Myers Squibb, Takeda and Astex are training OpenFold3 the opposite way: each company's protein structures stay behind its own walls, and a federated platform carries the learning out without carrying the data. The data never moves. Only the model improves.
Both are answers to the question this essay asked. Neither press release answers the harder one. When a model gets better because of what five companies fed it, who owns the improvement? Who can take it into a merger, license it, or walk away with it when the consortium ends?
That is a contract question, and the public announcements do not say how those rights were allocated.
And on the choice of where to point all of this first: on September 16, Novo Nordisk announced it will put Anthropic's Claude into its research and development workflows, with data governance and human oversight named as design principles. Novo chose a model-layer partner and put the governance language in the first paragraph. The table is being set. The seats are still open.