Why Palantir, Microsoft, and JPMorgan are signaling that enterprises should keep their AI in-house

Palantir CEO Alex Karp used a CNBC exclusive Monday to escalate his fight with the frontier AI labs (the small set of companies building the largest, most capable models, including OpenAI and Anthropic), but he is no longer fighting alone. Microsoft CEO Satya Nadella backed Karp on X on July 12, quoting him on "control over compute, models, data stack, and alpha," and JPMorgan CEO Jamie Dimon weighed in on AI spend discipline on CNBC's Squawk Box on July 15. Chamath Palihapitiya said on the same program that Karp "deserves a medal."

The buyer-side argument is hardening from a CEO tantrum into a coalition. What changed underneath is the shape of the enterprise AI deal: a year ago, the question was whether the models were good enough. Now it is who controls the data, IP, and proprietary logic, what Karp calls the "alpha," that ride on top of the model. Karp's critique of frontier labs has settled on a specific point. Enterprises should not have to surrender intellectual property to work with a frontier model, and they should not pay for output on a per-token meter, what Karp has called "tokenmaxx," that lets the lab reprice the relationship at will.

Karp told CNBC the labs are "trying to drug addict us to a future they believe they control," and accused Anthropic CEO Dario Amodei of pushing "a future where we own nothing, where our businesses aren't profitable, where none of us have jobs, and where our adversaries win." He framed Palantir's pitch as an "application layer" play: the software stack that sits between the model and the customer's data, keeping the data in-house and routing it through Palantir's ontology rather than the lab's API. Forbes reported the earlier round of these comments, from a July 1 CNBC appearance, as the moment Karp went viral for calling the AI industry "effing insane" over its token structure.

Anthropic and OpenAI have already conceded the easy half of this argument. Both labs publicly stated over the past month that customer data is not used to train their models, and OpenAI requires explicit enterprise opt-in before any submitted content can be used for training. CNBC noted it reached out to both companies for comment on Karp's renewed attack; as of publication, neither had responded on the record. The remaining fight, by Karp's own account, is not over data exfiltration. It is over pricing power and the application layer: who owns the relationship with the model, and who can reprice it.

The coalition forming around Karp gives the argument weight it did not have in early July, when his "effing insane" comments went viral. Nadella's X endorsement, quoting Karp on "control over compute, models, data stack, and alpha," was a signal that the platform layer is also uncomfortable with where the labs are headed. A DigitalApplied analysis read the position as a "means of production" argument: whoever owns the data stack and the application logic owns the enterprise AI economy. Dimon's appearance on Squawk Box, focused on AI spend discipline, was a buyer's signal that the largest enterprises are watching what they pay the labs. Palihapitiya's "deserves a medal" line read as a venture capital tell that the buyer-side case is being received on the same terms Karp is pitching it.

The next signal is whether the buyer-side coalition produces a procurement or product event. Karp's "application layer" pitch is structural: Palantir sits on top of the enterprise stack, not beside it. The frontier labs can match it by releasing enterprise products that keep the application logic inside the customer's perimeter, or they can concede the layer and compete on raw model quality. The open thread is whether either lab's next enterprise launch keeps the application logic inside the customer, or pushes it back into the lab's perimeter. CNBC's outreach to OpenAI and Anthropic is where the next on-record answer would land.