Last week, Alex Karp of Palantir did what he often does. He went on television, sounded half philosopher and half street fighter, and that made a lot of people uncomfortable.
His point deserves closer attention. Companies are rushing to use frontier AI models, and in doing so they may be handing over the very thing that makes them valuable: their data, their workflows, their judgement, their alpha.
In the US, this has already become a debate about sovereign AI. No surprise. Platforms start by helping an ecosystem, learn from it, and eventually move into its most valuable parts. It is simply what platforms do.
For Indian enterprises, the simple question is this: as our companies and public systems aggressively adopt AI, where should the intelligence layer sit?
Because our intelligence is not just our data. It is how a credit officer reads weak signals in a borrower’s behaviour. How a fund manager interprets management tone across twenty meetings. None of this appears in a database schema. It sits in notes, calls, judgement, exceptions, habits and scars.
AI becomes powerful precisely when it gets access to this context. That is also when the risk begins.
Indian finance alone holds enormous institutional intelligence: underwriting behaviour, investor preferences, fraud patterns, compliance judgement. Most of it is earned knowledge, not public data. If every bank, AMC, insurer and fintech pipes it into the same few external platforms, we gain short-term productivity while losing differentiation.
Rejecting frontier models would be foolish. Use them where they add value. But we should not confuse access with strategy.
A serious AI strategy needs an operating layer between the enterprise and the models: an AI control plane. Some work goes to frontier models, some to open models, some stays local, and some should never leave the company at all. It is basic governance.
Data by itself is not the moat. Interpreted data is, along with the judgement built on top of it. AI is now touching that layer. Which is why token pricing is the wrong metric. Did cycle time improve, did error rates fall, did risk reduce, did we preserve the intelligence that makes us different?
A model can summarise an earnings call. It cannot be allowed to become the investment committee.
We built services, digital public infrastructure, payments at population scale. We are now building AI capacity. Whether we build only usage, or also control, remains open.
Anyone can buy usage. Control has to be built.
A useful principle: use the strongest model available, but never let the model own the memory of the institution.
The entities that win with AI will not always be the ones with the most powerful model. Most often, they will be the ones who plan for what not to send to it.
Once the architecture hardens, we will not be choosing our intelligence layer. We will be renting it.
Better to have this debate today.
Originally posted on LinkedIn on July 5, 2026.
Srinivas Jain 