Enterprises Won't Feed the Labs. Palantir Cleared $1B on It.

TL;DR

Palantir reported Q2 2026 revenue of $1.935 billion, up 93% year over year, and GAAP net income of $1.062 billion. Read those two numbers together and the quarter gets strange: the company earned more profit in the last ninety days than it booked in total revenue in the same quarter a year ago. CEO Alex Karp's explanation is not a product feature. It is a claim about what enterprises now believe, namely that their competitive advantage should never end up inside somebody else's training run, and that they will pay a great deal of money to keep it out.

The number that does the work

Everything here comes from the 8-K exhibit filed August 3, not a summary of it. Revenue: $1,935,464 thousand, against $1,003,697 thousand in Q2 2025. Net income attributable to common stockholders: $1,061,890 thousand, a 55% margin. Diluted GAAP EPS of $0.41, against $0.13 a year earlier.

Companies growing 93% do not usually also print a 55% net margin. Companies printing a 55% net margin do not usually grow 93%. Doing both at a $7.7 billion revenue run rate is the part worth staring at.

The rest of the release tells the same story at different resolutions. US revenue grew 115% to $1.573 billion. US commercial grew 149% to $764 million. US government grew 90% to $809 million. Palantir closed 73 individual deals worth at least $10 million, and a record $2.132 billion of US commercial total contract value, up 153%.

Its Rule of 40 score, the SaaS heuristic where growth plus margin is supposed to clear a pass mark of 40, came in at 155. A pass/fail line is being used as a flex, which tells you where the quarter sits on the distribution.

Karp's actual argument

The quotable part is in the press release itself, which is unusual. Most CEO quotes in an earnings release are wallpaper. This one is a position paper:

"Demand for AI sovereignty has now been unleashed. And Palantir is the only company that has demonstrated it can transform tokens into actual economic value. Our customers trust us to provide them with maximal control over their operations, data, and decisions. Their competitive advantage should never become the training data for future models."

That last sentence is the entire pitch, and it is aimed squarely at the frontier labs. In his quarterly letter to shareholders, Karp went further. Per TechCrunch, he wrote that "there are Marxist overtones and undertones to our business," and that others, "including many of those building large language models, intend, knowingly or otherwise, to capture the means of production of their purported partners." He told CNBC that "we have people trying to drug addict us to a future they believe they control."

Strip the social theory and the mechanism is mundane. When you route your workflows, your exception handling, your pricing logic, and your escalation rules through a lab's API, you are not just buying inference. You are handing over a very high-quality description of how your business actually works, to a counterparty whose stated ambition is to automate work in general. It is like teaching your one great recipe to the caterer who has quietly signed a lease across the street. The food is excellent. The lease is the problem.

Is he right, or is he selling?

Both, and it is worth holding those separately. Palantir sells the application layer: AIP and its Ontology sit on top of whichever models a customer allows, inside the customer's own environment. "Do not hand the labs your institution" is a genuine argument and also, conveniently, the exact shape of the product. A man with a moat business has strong opinions about renting moats.

But the argument is not his alone. On July 24, roughly twenty-five companies including Palantir, Nvidia, Microsoft, Meta, IBM, and Hugging Face signed Open Weights and American AI Leadership, urging US policymakers not to restrict downloadable model weights. As CNBC reported, OpenAI and Anthropic were not on it. The split in that signature list is the same split Karp is monetizing.

What this means if you build with AI

You are unlikely to buy an Ontology license this week. The transferable part is the procurement logic that produced a 149% growth rate, because it is now showing up in the contracts your customers hand you.

• "No training on our data" is table stakes, and it is not the whole question. Zero-retention flags cover the payload. They do not cover aggregate telemetry, eval sets built from your traffic, or what a lab learns about a vertical by serving forty companies in it.
• Open weights are now a procurement answer, not a hobby. A model you can download, pin, and run on your own hardware ends the conversation about data egress rather than mitigating it. vLLM for fleet serving and Ollama for local work are boring, mature, and increasingly what "sovereign" resolves to in practice.
• The moat is the ontology, not the model. Palantir's growth argues that the durable asset is the mapping between a company's messy systems and a schema an agent can act on. Model weights churn every quarter. That mapping does not.
• Expect the question earlier in your sales cycle. If you are an indie shipping on someone's API, the "which model, hosted where, retained how long" question is moving from the security review to the first call.

The caveats, which are real

The 62% margin everyone will quote is the adjusted one. GAAP operating margin was 47%, on $912 million of operating income. The gap is $265 million of stock-based compensation plus $17 million of related payroll tax, added back. That is 14% of revenue in dilution, and it is a real cost even when it is not a cash one.

Second: $809 million of the $1.573 billion US number is government. "Enterprises are refusing to feed the labs" is a clean narrative, but a large share of this quarter is defense and federal spending, which runs on procurement cycles rather than on anyone's feelings about training data.

Third: guidance is guidance. Palantir raised full-year revenue to a range of $8.150 to $8.158 billion, US commercial to at least $3.424 billion, and adjusted free cash flow to $4.5 to $4.7 billion. None of that has happened yet.

Key Takeaways

• Palantir's Q2 2026 GAAP net income of $1.062 billion exceeded its entire Q2 2025 revenue of $1.004 billion, on revenue of $1.935 billion (up 93%).
• US commercial revenue grew 149% to $764 million, with a record $2.132 billion of US commercial contract value closed, up 153%.
• Karp is selling "AI sovereignty" explicitly: customer advantage "should never become the training data for future models," and LLM builders want to "capture the means of production of their purported partners."
• Palantir signed the July 24 open-weights letter alongside Nvidia, Microsoft, Meta, and Hugging Face. OpenAI and Anthropic did not.
• The 62% margin is adjusted; GAAP operating margin was 47%, with $265 million of stock comp added back. Government is still $809 million of the US total.
• For builders, the practical read is that "which model, hosted where, retained how long" is becoming a first-call question rather than a security-review footnote.

Sources: Palantir Q2 2026 earnings release (SEC 8-K, Exhibit 99.1), Palantir Q2 2026 Letter to Shareholders, TechCrunch, CNBC on Karp and the frontier labs, CNBC earnings coverage, Open Weights and American AI Leadership, CNBC on the open-weights letter.