Blog

Don't rent your intelligence

Apple built the most successful platform business in history, and they did it by renting distribution. You built the app, Apple took 30%, and in exchange you got access to a billion customers. The relationship had tension, particularly around that 30%, but the incentive alignment was real: Apple made more money when the ecosystem was healthy than when it wasn't. They deliberately kept their stock apps simple and slow to evolve, taking a decade to add features that third-party apps had shipped years earlier. Developers knew that if they built something too obvious, Apple might eventually replicate it as a default, but the pace was slow enough and the distribution valuable enough that the trade worked for most people.

The emerging relationship between knowledge workers and AI providers follows a superficially similar pattern, but the underlying economics are different in ways that matter enormously.

When you build on Apple's platform, Apple knows what you sell and how many people buy it. When you build on an AI provider's platform, or more precisely when you pour your work into their tools, the provider knows how you think. Your research methods, your analytical frameworks, your strategic reasoning, your client insights, the patterns in your decision-making that you might not even be conscious of yourself. The difference between renting distribution and renting intelligence is the difference between a landlord who knows your address and a landlord who reads your diary.

The platform predation pattern

There's a well-documented pattern in platform businesses: the platform observes where value is being created on top of it, then moves into those verticals. Microsoft used Windows dominance to capture spreadsheets, word processing, and web browsing. Google started by sending users to other websites via search results, then gradually kept more than half of searches on its own properties. Amazon Marketplace let third-party sellers prove demand for product categories, then launched Amazon Basics in those exact categories.

The same pattern is emerging in AI, and it's moving faster than any previous cycle. Cursor became one of Anthropic's biggest API customers, demonstrating that the coding assistant category was enormous. Anthropic launched Claude Code. Revenue accelerated immediately. Then came Claude Design, which reportedly caught Figma off-guard despite the fact that Anthropic's CPO had been sitting on Figma's board and didn't resign until three days before the launch. Then Claude Science. Claude Security. Claude Legal. Claude Financial. Each one moving into a category that was previously served by companies building on Anthropic's own models.

The companies that were building on the API had, in effect, been running market research for their infrastructure provider. They proved which verticals were valuable, trained users to expect AI-powered tools in those categories, and then watched the platform move in with the advantage of owning the model layer underneath.

The argument that open source models should be restricted on safety grounds looks rather different when you notice that a restricted model layer with only a handful of players is exactly the market structure that benefits those players commercially. This is worth sitting with for a moment, because the people making the safety argument and the people who benefit from the competitive outcome are the same people.

The forward-deployed trojan horse

There's a smaller but revealing trend worth paying attention to. Microsoft recently announced $2.5 billion to deploy forward-deployed engineers into enterprise customers. Amazon is spending a billion on the same programme. OpenAI and Anthropic both have them. These are engineers who come into your company, study your workflows, build integrations, and send what they learn back to improve the model.

The stated purpose is to help enterprises adopt AI more effectively. The structural effect is that the AI provider gets an intimate understanding of how your business works, what your competitive advantages are, where the value sits in your workflow, and which parts of your operation could be productised. Every insight these engineers bring back makes the provider's next product better targeted, and the targeting may be aimed directly at the category you operate in.

The life sciences industry is already responding to this dynamic. Anthropic and others have been approaching large pharmaceutical companies, asking them to contribute proprietary experimental data to train specialised models, with the pitch being early access and some proprietary advantage in return. The companies are largely saying no. When you've spent tens of billions generating proprietary data through years of experiments and product development, that data is a core asset, and contributing it to a shared model that also ingests your competitors' data commoditises the very thing that differentiates you.

The pharma companies can see the pattern because they're large enough and experienced enough to recognise it. Individual knowledge workers, small teams, and mid-market companies often can't, or don't think to look, until the pattern has already played out.

The intelligence sovereignty question

There's a deeper layer to this that goes beyond platform risk and commercial predation.

Privacy, in the traditional sense, is about access: who can see your data. Intelligence sovereignty is about something else entirely: who controls the analytical lens applied to your information. When your AI assistant is provided by a company with its own commercial agenda, product roadmap, and competitive strategy, the analysis you receive is filtered through priorities and training choices you have no visibility into and no control over.

This matters most for people whose thinking is their product. A consultant's value is in their analytical frameworks. A researcher's value is in their hypotheses and methodology. A lawyer's value is in their strategic reasoning. A founder's value is in their understanding of a market that others haven't grasped yet. When any of these people pour their work into a tool controlled by a company that might productise their category next quarter, they're sharing the raw material of their competitive advantage with an entity whose interests may not be aligned with theirs.

The historical parallel is less about technology platforms and more about the relationship between thinkers and institutions. The printing press didn't just democratise access to information, the political fights were about who controlled the interpretation. The news wire services of the early twentieth century weren't neutral conduits of information, they shaped what was considered newsworthy. The framing has always mattered as much as the facts, and the AI layer is increasingly the framing layer for knowledge work.

The structural alternative

The alternative to renting your intelligence is owning the infrastructure that supports it. This doesn't mean building your own AI lab or training your own models. It means choosing an architecture where the intelligence layer serves you without absorbing you.

In practice, this looks like a few things working together.

Model choice, including open source. If you can choose which model processes your data, including models that run locally or on your own infrastructure, no single provider has a monopoly on your thinking. Open source models have reached the point where they're competitive with frontier models for the vast majority of tasks, and they can run on hardware you control with zero data flowing to a third party. The argument that these models are too dangerous to use freely deserves scrutiny when the people making that argument benefit directly from a closed market.

Storage you control. Your accumulated knowledge, the context that compounds over time, should live in infrastructure you own. Bring-your-own-storage means the data sits in your S3 bucket or your R2 account, encrypted with your keys, and the application layer reads from it rather than holding it. If you change tools, the data stays exactly where it is.

A boundary layer between you and the models. Fabric sits between your knowledge and the AI models, controlling what any given model can see and ensuring that the interaction serves you rather than the provider's training pipeline. Think of it as a kind of personal Palantir: the intelligence is yours, the models are interchangeable, and the boundary is enforced by architecture rather than by trust.

Portability as a default. Open protocols like MCP mean your knowledge library is accessible to any AI agent on your terms, without being locked into a specific provider's ecosystem. The library outlasts any individual model generation, any individual tool, and any individual vendor. It's your context, and it stays with you.

Who this matters for

The people who should think hardest about intelligence sovereignty are the people whose accumulated thinking is their primary competitive advantage.

Researchers whose unpublished hypotheses and experimental results represent years of work. Consultants whose client insights and analytical frameworks are what they sell. Lawyers whose case strategies would be devastatingly valuable to opposing counsel. Founders whose market understanding is what makes their company worth building. Investors whose deal evaluation and pattern recognition across hundreds of companies is their edge.

For all of these people, pouring their best thinking into a tool controlled by a company that might productise their category, train on their data, or simply change the terms of service is a risk that grows with every day of use, precisely because the value of the accumulated context grows too.

The people building context now, in systems they control, are making an investment that compounds in their favour. The people building context in systems controlled by AI providers are making the same investment, but the compounding may benefit someone else.

Frequently asked questions

Aren't the big AI providers trustworthy enough? Some of them may be, for now. But the question isn't whether you trust the current management team. The question is whether you trust every future management team, every future business decision, and every future competitive move that might be made by a company whose commercial interests may diverge from yours. The history of platform businesses suggests that the answer is no, regardless of how well-intentioned the current leadership might be.

Can open source models really compete with frontier models? For the vast majority of real-world tasks, yes. The gap has narrowed dramatically and continues to narrow. Nvidia's Nemotron, Meta's Llama, and other open source models are competitive for most professional use cases. The model that was state-of-the-art two years ago was already good enough for 90% of tasks; the question was always cost, and running open source models on your own infrastructure is increasingly affordable.

What about teams and enterprises? The argument scales directly. A team's accumulated knowledge, their shared context, their institutional memory, is even more valuable than an individual's, and more vulnerable to platform dependency because the stakes are higher and the data is richer. Team workspaces with controlled storage and model choice apply the same architecture at the organisational level.

Is this just anti-AI? The opposite. This is pro-AI and pro-ownership. The argument is that AI is enormously valuable and getting more so, which is exactly why you should own the intelligence layer rather than renting it from a company whose interests may not align with yours. The more valuable AI becomes, the more the question of who controls it matters.