That second payment is the one that should keep operators up at night. The better you want a model to perform on your problem, the more of your hard-won knowledge you have to hand over: the prompts your experts write, the traces of how your teams actually work, and – most valuable of all – the specialized know-how you feed in to get useful work out: the synthesis routes, the sourcing logic, the way your code is built. It's the kind of knowledge a competitor could never buy, and exactly the kind that leaks out one interaction at a time, almost imperceptibly.

We think Nadella is right, and we think the implications run deeper for companies whose entire advantage lives in specialized, hard-won knowledge. For us, that advantage takes two forms. One is the chemistry: how to source a molecule, route a synthesis, and build a supply chain that holds when the world shifts. The other is the software itself – the codebase our product is built on, and the engineering decisions embedded in it. When developers use AI to write, debug, or refine that code, the same paradox applies: the model sees how our system is architected, where its hard problems are, and how we chose to solve them. Chemistry knowledge and source code are both IP, and both leak the same quiet way. Neither is exhaust. Both are the business.

Why we take this seriously

Occam is an AI-enabled company. We use commercial models in our workflows because they make our people faster and sharper, and we'd be foolish to pretend otherwise. But using that intelligence and surrendering our own are two very different things, and we are deliberate about the line between them.

The reason is simple: the value we create for clients rests on knowledge that is genuinely ours. The judgment our chemists and engineers apply, the patterns we've learned across engagements, the corrections that make our systems smarter over time –  that accumulated intelligence is the whole point of working with us. If that learning flowed in only one direction, out of our hands and into someone else's model, the economics of the entire field would drift toward whoever owns the learning infrastructure, rather than the people who created the knowledge in the first place. We're not willing to let that happen – not with our knowledge, and not with our clients'.

So we treat intelligence sovereignty as a first-order concern, not an afterthought. It shapes how we adopt AI: which tools we use, where our proprietary data is allowed to go, and what we insist stays inside our own boundary. The goal is to capture every bit of leverage these models offer while making sure the knowledge that makes Occam Occam compounds for us and our clients – not for a third party we happened to buy software from.

The provider can become the competitor

There's a sharper edge to this than ordinary data hygiene, and it's the part that most deserves attention. The company selling you the model can turn around and use what it learns from you to compete with you. In one moment, it's a vendor of tools and services. In the next, armed with a pattern of how your industry actually works, it's a direct rival. Those two roles sit inside the same relationship, and the switch between them can happen quietly.

For chemicals, that isn't a hypothetical. This is a $5 trillion industry underpinning nearly everything the world makes – far too large for the frontier AI companies to overlook. We've already watched specialized, verticalized models emerge for law and for healthcare, purpose-built for domains that were until recently considered too complex and too expert to automate. It would be naive to assume chemicals won't be next. And the raw material for building that vertical model is exactly the knowledge a chemical company feeds in while trying to get useful work out. The more of ourselves we pour into a general model, the more we help assemble the specialized one that could one day be pointed back at us.

That is a large part of why we take this so seriously. We intend to draw every advantage we can from commercial AI, and, as much as we're able, to avoid handing these companies the means to verticalize into our space and compete against us with our own knowledge.

From protecting data to protecting learning

There's a useful way to frame the shift Nadella is pointing at. In the cloud era, enterprises learned to protect their data. In the AI era, that's no longer enough. The thing worth protecting now is the mechanism of learning itself – the loop through which an organization gets better at what it does over time. Data is a snapshot. Learning is the compounding asset.

Patents solved one half of Arrow's original paradox: they let an inventor disclose an idea without simply giving it away. The reverse paradox needs its own equivalent – a way for a company to draw on the world's best models without quietly handing over the institutional know-how that makes it distinct. We don't think that's a problem any single vendor solves for you. It's a discipline a company has to build for itself: clear boundaries, deliberate choices about where knowledge accumulates, and a refusal to let convenience quietly erode ownership of your own edge.

Our takeaway

We're grateful for essays like this one, because they name a risk that's easy to feel and hard to articulate. Nadella gave it a name. Our job is to act on it – to keep using the best intelligence available to us while making sure that, in consuming it, we go on creating and keeping intelligence of our own.

In a volatile world, the companies that win will be the ones that stay in command of what they know. That's true of supply chains, and it's true of the knowledge underneath them. Advantage: you – and it should stay that way.