Palantir CEO Alex Karp spent the first of July on CNBC arguing that enterprises adopting AI want to "own the means of production." Whatever you make of his delivery, the question has gone mainstream. The frontier models are the most capable tools I've seen in my career in tech. They are also interchangeable in a way nothing else in your stack is: today's best model will not be next year's, and prices move every quarter. The model cannot be where your firm's memory lives.

The logic is the alpha

For a private fund, that knowledge is the fund's logic: how the waterfall tiers in practice, the side letter you gave your anchor LP in Fund II cascades and impacts fees for everyone else, and which distributions are recallable and until when. Your track record is the output of that logic. Every cut of the numbers an LP asks for during diligence is that logic, re-run. The same history is how you underwrite the next deal, which assumptions held and where you were consistently wrong.

Today, that logic is not codified as the operational data layer for the future of the fund. It lives in the CFO's head and a few dozen workbooks. You can't version it, you can't query it, and when the person holding it leaves, it leaves with them.

The reverse information paradox

Microsoft CEO Satya Nadella argues that AI runs the paradox in reverse: a buyer can't know what information is worth until they have it, and once they have it, they've "in effect acquired it without cost." The buyer risks giving knowledge away just to use what they bought, paying for intelligence twice, once in money and again in the proprietary knowledge revealed to make it useful. The seller risks giving knowledge away just to sell it. The better you want the model to perform, the more of that knowledge you have to feed it.

For a fund, the leak is quieter than a data breach. AI products improve on what Nadella calls exhaust: the prompts your team writes, the tools its agents call, and especially the corrections people make when the model is wrong. Every time your controller fixes a model's reading of a fee offset or rejects a track-record cut that doesn't tie, that correction is distilled know-how a competitor couldn't buy, leaking trace by trace. As your fund experiments with LLMs running on Excel, those workflows, prompts, and corrections accumulate inside one provider's tools until the firm's intelligence is inseparable from that product, and the learning flows one direction: the seller learns more about your firm with every use, while you learn very little about what they've learned.

You can see the early cost on the token bill. Most of what you pay for is rediscovery: the model re-reads the same LPA and rebuilds the same waterfall understanding it built last week before it can even attempt the question you asked. And most teams measure activity rather than time to accuracy. If the close isn't measurably faster, the technology will not provide alpha.

Confidence in the answer is the blocker

We have run this with off-the-shelf tools: a frontier model pointed at a capital-statement workbook. It reported $30.8 million of accrued carry. The real number was $1.4 million. The tool had matched a transaction-type label to the wrong cells and returned the wrong number with complete confidence. A capital call, an allocation, a waterfall cannot be probably correct; it has to reconcile, trace to source, and come out the same way every time. The model helps your people ask better questions and move through complexity faster; the number itself has to come from a governed system that computes it from the fund structure, the ownership, and the allocation rules.

Own that layer and the collective intelligence accumulates. An LP asks a follow-up about their net position, and you can answer it in the meeting instead of two weeks later. A prospective anchor wants their own terms, and you can price them before you concede them. You walk into the next raise able to cut the track record any way the room asks, same day, tied to the penny. Each interaction gets recorded, the workflows improve, and the next decision starts further ahead. That loop, not a prompt library, is the firm's new intellectual property.

The AI sovereignty test

Maybern is betting that logic ownership is the essential foundation to an agentic fund. Your data stays yours, your rules live in configuration your team can read, versioned and queryable. The numbers come from deterministic rails with no live model call touching a fund calculation. Success gets measured in closes and error rates rather than usage.

Logic captured with agents running on top drives alpha that deepens with every close and every raise. Funds that give up their AI sovereignty to publicly available tech are relinquishing their alpha in exchange for tokens.