AI Sovereignty for Deal Firms: Own the Context, Rent the Model

There is a loud version of this argument and a useful one. The loud version is Palantir’s Alex Karp calling frontier AI pricing a wealth tax: enterprises paying for tokens while the labs learn from their data. You don’t have to buy the theatre to take the point underneath it. When an AI vendor charges you per token rather than per outcome, the price structure itself tells you what they believe about their product. Electricity is priced per kilowatt-hour because nobody’s kilowatt is special. Commodity pricing means interchangeable goods, and interchangeable goods can’t be anyone’s edge.

For a deal firm that has a specific consequence. I’ve argued before that the model doesn’t matter and your data layer does, and the market has spent two years proving it. But interchangeable models raise a sharper question than “which one,” and it’s the question worth spending a partner’s attention on: who ends up owning the layer that everything else depends on?

Everyone is converging on the same ground

Watch where the largest players are actually moving. The frontier labs are climbing from models up into applications, launching products aimed squarely at the industries their customers work in. The biggest software companies are moving into embedded services, putting engineers inside enterprises to build AI capability on proprietary data. Sovereignty vendors sell the whole stack inside your perimeter. And open source keeps resetting the price floor underneath all of it.

Four different strategies, one destination: the layer that connects an organisation’s knowledge to the machines that work on it. When every serious player converges on the same piece of ground, that is the market telling you where the value settles. It also tells you that control of that layer is now a live question for any firm that would rather be the owner than the terrain.

The three risks of renting everything

Price is the visible risk. Frontier rates run well above open alternatives that decent engineering can match, so a firm that rents the premium model as its entire strategy pays a premium for no differentiation, since every competitor can rent the same thing.

Continuity is the second. Access to a hosted model changes on someone else’s timetable, whether that’s a price rise, a deprecation, or a decision taken several boardrooms away from yours. Anything you have wired to run on exactly one vendor’s model inherits that vendor’s roadmap and its risk appetite.

Competition is the least discussed and, lately, the best evidenced. Frontier labs are entering the application layer of the industries they serve, and recent history shows partnership agreements buying incumbents notice measured in days. Any profitable workflow sitting on top of a frontier model is a candidate for that lab’s own product roadmap, and the data you feed a vendor today can help train the competitor you meet tomorrow. That belongs in your vendor reviews, and if you invest in software, in your diligence.

What a deal firm actually owns

Strip a fund to its durable assets and the list is short: capital relationships, reputation, and the accumulated record of judgment. Every deal sourced, priced, passed on and closed, and the reasons behind each call. That record is the one input no model vendor can supply and no competitor can copy.

Most firms hold it in the worst possible form, scattered across drives, inboxes and CRMs that don’t talk to each other. We saw this with a lower-middle-market PE firm whose partners couldn’t query their own history: every “have we seen this before?” meant half a day of folder archaeology, and institutional memory walked out the door whenever someone left. We connected five years of their deals, documents, emails and models into one queryable foundation, put their existing AI assistant on top of it, and every answer came back cited to source. A decision on a new deal got roughly ten times faster. As one of their partners put it, they could finally ask why the firm passed on a deal three years ago and get a real answer, with the sources.

The point there is not the hours saved on any one question. It’s that the firm’s judgment stopped living in individual heads and started compounding. Every deal the team looks at now makes the next one faster to assess, because the reasoning and the comparables are one query away and always current. That is what owning the context layer buys you, and no model swap can take it away.

Sovereignty is a spectrum

The maximal version, open-weight models on your own hardware inside your own perimeter, exists for firms whose regulators or LPs demand it. Most don’t need it. The practical version has three parts.

Own your data and its shape. Your deal record lives in systems you control, structured so any model can use it and exportable without a vendor’s permission. The mechanics of building that record are a subject in themselves: why AI needs a system of record, not just another folder of files.

Rent the intelligence. Use the best commercial model for each job and assume you’ll swap it within the year. The moment your workflows only run on one vendor’s model, you’ve traded your optionality for someone else’s roadmap.

Read the data terms. Know what each AI contract lets the vendor retain, learn from, or train on, and treat that as diligence rather than procurement.

Where to start

Not with a platform decision. Start with an inventory: what does the firm know, where does it live, and how much of it could an AI system actually reach today? Then audit the data terms across every AI contract in the firm and the portfolio. Then build the record, one structured home for deals, decisions and outcomes, made to outlast whichever model happens to be best this quarter.

Ask your own firm the question you’d ask any software target in diligence: what do we own that survives a model swap? If the honest answer is a stack of subscriptions, the work starts with the record. The firms doing this are compounding their own judgment. The rest are paying the wealth tax and calling it a strategy.