AI Strategy

What Alex Karp Got Right About Enterprise AI, and What It Means If You Are Not the Pentagon

By the Praxiron team · Last updated July 5, 2026 · 7 min read

What did Alex Karp actually say on CNBC?

Karp came on CNBC’s Squawk Box to discuss Palantir’s expanded partnership with Nvidia and spent most of the interview on a broader argument. Stripped of the theatrics, it has four parts.

First, a model alone is not a product. In his framing, value comes from three things together: the model, an application layer, and compute. An AI application layer is the software that sits between a company’s knowledge and the AI model: it structures what the model receives, controls what leaves the company, and makes every output traceable back to a source. Palantir’s application layer is called Ontology, and he noted that the profitable parts of Palantir’s business are the layer and the compute, not the models underneath.

Second, enterprises are frustrated with the return on AI spend. CNBC reported that Karp described CEOs as increasingly unwilling to pay for token consumption that does not translate into business results, summarizing the mood with the line “something has gone completely wrong.”

Third, control. Forbes reported that Karp framed the Nvidia partnership around what technical customers want: ownership of their compute, their models, their data stack, and what he calls their alpha, the accumulated know-how that makes a business hard to copy.

Fourth, questions. Karp argued that before connecting AI to their operations, buyers should be able to ask and answer basic questions: who holds the data, where it is cached, whether prompts are secure, and whether value is transferring to a third party.

One thing worth keeping in view: Karp gave this interview two days after announcing the product that answers his own critique. That does not make the critique wrong. It does mean the critique should be separated from the sales pitch, which is exactly what the rest of this article does.

Is he right that enterprises get no value from AI spend?

On the frustration itself, the independent evidence is on his side, and it predates the interview.

MIT’s NANDA initiative studied 300 enterprise AI deployments and found that about 95 percent of generative AI pilots deliver no measurable P&L impact (MIT, The GenAI Divide: State of AI in Business, 2025). The report’s most important finding is the one least quoted: the failure is not driven by model quality. It is driven by integration, by the absence of tools that carry a company’s context, learn its constraints, and fit its workflows.

PwC’s Global CEO Survey points in the same direction from the top of the organization, with a majority of CEOs reporting no significant revenue gain or cost reduction from their AI investments (PwC, 2026).

So Karp is describing something real. Companies bought reasoning power and skipped the part that connects reasoning power to their actual business. The result is exactly what he describes: consumption without return.

The four questions every company should be able to answer

The most useful thing in the interview is not the argument. It is the checklist. Karp’s questions are the right ones for any company connecting AI to its operations, and they deserve straight answers rather than slogans:

A company that can answer all four is in a strong position with any AI engine. A company that can answer none of them has a governance gap, regardless of which model it uses.

Is the answer to distrust the models?

No, and this is where the loudest readings of the interview go wrong.

The frontier models are the most capable reasoning engines ever built. The MIT data quoted above makes this point precisely: the 95 percent failure rate is not a model problem. The same engines that stall inside enterprises are producing extraordinary results wherever they are given structure, context, and clear constraints. Walking away from them means walking away from the reasoning capability itself, which no serious company will do.

The honest conclusion from Karp’s own argument is different: the model was never supposed to carry the whole job. A frontier model is an engine. An engine does not know your rules, your constraints, the reasons behind your past decisions, or what a good outcome looks like in your business. Something has to hold that knowledge, decide what the engine sees, and make the output verifiable. Palantir calls that layer Ontology. The industry calls it the application layer. Whatever the name, the principle is the same.

This is the frame we work from at Praxiron: we do not replace AI engines, we unlock what they were always capable of. The engines are extraordinary. What most companies are missing is the step before them. Karp’s questions are the right questions, but the answer for most companies is not to fear the models. It is to stop handing them raw data and start handing them structured reasoning. The same engine, given your logic instead of your files, is a different product.

What does the application layer mean outside the Pentagon?

Palantir’s answer to these questions is built for governments, militaries, and critical infrastructure: sovereign deployments, air-gapped environments, forward-deployed engineers. If you run a defense ministry, that is the right shape.

Most companies are not the Pentagon. They are engineering firms, funds, insurers, and manufacturers whose hardest asset is judgment: the rules, constraints, and reasoning behind years of decisions, living in a few experts’ heads and scattered files. For them, the same principle takes a different shape:

The knowledge stays yours. The company’s rules, standards, and past decisions are extracted into a structured reasoning model, a decision DNA, that lives in the company’s cloud environment as a company-owned asset. It is not a chat history and not a vendor’s training data.

The model sees the reasoning, not the business. When a question is asked, the engine receives structured, relevant context for that decision, not a wholesale upload of company files. The engine gets what it needs to reason. The alpha, in Karp’s language, stays home.

Every answer can be checked. Outputs arrive ranked, with the reasoning attached, a reference to the exact rule or past project relied on, and a calibrated confidence level. When no sufficient source exists, the platform says so instead of guessing. This is the practical answer to the trust question: not “trust us,” but “check us.”

No engine lock-in. The layer sits above every engine, not inside one. If a better model ships, you switch, and your decision DNA comes with you. This is the same engine-agnostic logic Karp describes, applied at the scale of a company rather than a state.

This is what we build. Praxiron is a decision intelligence platform that extracts a company’s rules, constraints, and past decisions into a structured decision DNA, and sits as a reasoning and control layer above any AI engine. Not sovereign infrastructure, and not another chatbot: the layer that makes the engines you already pay for finally produce results you can act on. You can see how the pieces fit on our platform page and how data handling works on the security page.

How should you apply this in your company?

Three practical moves, in order.

Ask Karp’s four questions internally this week. Not to a vendor, to your own team. Map what data flows to which AI tools today, under which agreements, and whether anyone can trace a given output back to a source. Most companies discover the answer is “we don’t know,” and that discovery is worth more than any tool purchase.

Separate the engine decision from the knowledge decision. Which model to use is a replaceable, low-stakes choice, and it should stay that way. How your company’s knowledge is structured, governed, and connected to models is the durable, high-stakes choice. Companies that fuse the two end up locked in; companies that separate them keep their leverage.

Measure AI on decisions, not usage. Token consumption, seat counts, and adoption rates are the metrics that produced the 95 percent. The metric that matters is whether the output changed a real decision, and whether you could verify it before acting. If you are earlier in this journey, start with should you connect your company files to AI?

Karp shouted the question. The companies that answer it quietly, with structure instead of slogans, are the ones that will get what everyone was promised.

If you want to see what your company’s decision DNA would look like, request access.

Photo by Sebastian Schuster on Unsplash