Palantir's 149% US commercial quarter: the stack is the product

Palantir reported Q2 on Monday and the number that should interest anyone running a business — not a portfolio — is US commercial revenue up 149% year over year. Not government. Not a defense line item. Ordinary American companies, more than doubling what they spend with a vendor that sells them a stack rather than a model. That's a signal about where enterprise AI budgets are actually landing, and it's worth reading if you're deciding how to buy.

What actually happened

Per Palantir's earnings release, Q2 revenue was $1.935 billion, up 93% year over year. US commercial revenue hit $764 million, up 149%; US government came in at $809 million, up 90%. GAAP net income was $1.062 billion — a 55% margin. The company raised full-year 2026 guidance to roughly $8.15 billion in revenue, implying 82% growth, and lifted US commercial guidance to more than $3.42 billion, or at least 134% growth. CNBC reported shares jumped double digits after hours.

CEO Alex Karp used the quarter to attack the frontier labs directly, arguing they intend to "capture the means of production of their purported partners" — per TechCrunch, his claim is that model vendors absorb a client's intellectual property and workflows and then compete with them. He's selling against that, so read it as a pitch. It's also not a new observation: we've watched Microsoft compete with its own AI suppliers and dev-tool vendors chase their customers' customers all year.

Why enterprise AI spending matters for your business

You are not buying Palantir. Their smallest commercial deals are larger than most companies' entire software budget. But the shape of what those buyers are paying for is copyable, and it's the opposite of how most small operators buy AI.

They're not paying for a model. They're paying for an ontology — their own data, mapped into their own objects, with the model as a swappable component sitting on top. That's why 149% growth is happening at the same time model prices are collapsing. The model got cheap. The thing wrapped around it did not.

The small-business version of that is unglamorous and cheap: your customer, order, job, and inventory records living in a database you control, with a defined schema, reachable by an API. Once that exists, the model is a config line, and you can move from Sonnet to Haiku to an open-weight model on your own hardware without renegotiating your business. Without it, you're renting someone else's ontology, and every AI feature you add makes their product better and your position weaker.

Karp is right about the mechanism even if he's talking his book. The defense isn't picking a nobler vendor. It's owning the layer underneath all of them.

Sources: Palantir Q2 2026 earnings release (BusinessWire), CNBC, TechCrunch.