Palantir's Karp Called the AI Industry 'Marxist' the Same Day He Posted a $1B Profit Quarter. Here's the Actual Test for Whether His Warning Applies to Your SaaS.
Follow-up to: 2026-07-05: "Palantir's CEO Called AI a 'Wealth Tax' on Your Business. He's Half Right, and the Half He's Right About Tells You What to Sell."
On August 3, Palantir reported Q2 2026 revenue of $1.935 billion, up 93% year over year, with net income crossing $1 billion for the first time in the company's history. On the same earnings call, CEO Alex Karp told CNBC that frontier AI labs are "trying to drug addict us to a future they believe they control," and TechCrunch's headline quoted him calling the AI industry "Marxist." I covered Karp making a similar argument back in July, when it was a CNBC rant with no numbers behind it. This time he has the numbers, and it changes what's worth taking seriously.
What's actually new since July
The July post was built around Karp calling AI a "wealth tax" during a CNBC interview: a strong quote, weak evidence. This one lands on an earnings call where Palantir just posted a Rule of 40 score of 155%, raised its full-year revenue guidance to $8.15 billion to $8.158 billion, and grew US commercial revenue 149% year over year. Karp isn't complaining from the sidelines anymore. He's making the same argument from a company that just had, by any normal measure, an extraordinary quarter, which gives the argument a credibility it didn't have three weeks ago, even if the underlying claim hasn't changed.
Karp's actual position, stripped of the CNBC theater: enterprises are handing frontier AI labs their operational data, their workflows, and their institutional knowledge in exchange for API access, and the lab captures the compounding value of that relationship while the customer just gets a monthly bill. He's framed this as labs trying to "colonize" or "capture the means of production of their purported partners," which is dramatic language for what's a fairly ordinary platform-dependency argument dressed up for television.
Where the argument actually holds up
Here's the part worth taking seriously if you're a solo operator: if your entire product is a prompt wrapped around one vendor's API, with no data moat, no evaluation set you own, and no realistic fallback provider, you have taken on exactly the dependency Karp is describing. You're just doing it at a much smaller scale, with none of the negotiating leverage a company posting $1.9 billion in quarterly revenue has when it sits across the table from Anthropic or OpenAI.
The mechanism is straightforward. Your product's actual value (the thing customers pay for) increasingly lives in three places: the raw model capability (which you don't control and which improves or gets deprecated on someone else's schedule), your prompts and system instructions (which are usually thin and easy to reverse-engineer from your product's outputs), and whatever data or workflow you've wrapped around the model (which is often the only part that's actually yours). If a platform update changes model behavior overnight (and this has happened to real products, not hypothetically), you find out how much of your value proposition was actually the model doing the work versus your layer on top of it.
The test that actually matters
Forget the "wealth tax" framing. The concrete question is: can you swap your model provider in an afternoon? If your prompts, your evals, and any fine-tuning are all built around one vendor's specific quirks (the exact phrasing that gets Claude to behave a certain way, or a fine-tune that only runs on one lab's infrastructure), swapping providers isn't an afternoon, it's a rewrite. That's the actual risk, and it's a risk you can audit and reduce this week: build your prompts against a provider-agnostic interface, keep a small eval set you control, and actually test your product against a second provider's model at least once a quarter, even if you never switch. The point isn't redundancy for its own sake. It's knowing what breaks and what doesn't before a pricing change or a model deprecation forces you to find out live.
The honest counter-take
Karp has an obvious incentive to make this argument. Palantir's entire business model is selling the software layer that sits on top of frontier models (the "outcome," in the language of my July post) specifically so enterprises don't have to build that dependency directly on OpenAI or Anthropic. Every time Karp tells an enterprise that going straight to a model API is reckless, he's also pitching them Palantir's platform as the safer alternative. That's not automatically wrong, but it means his argument and his sales pitch are the same sentence, which should lower your trust in the framing even where the underlying mechanism is real.
It's also worth saying plainly: there's no evidence any frontier lab is literally using customer prompts to train competing products against them. Anthropic, OpenAI, and Google all publish policies stating they don't train on API customer data by default. Karp's "capture the means of production" language implies something closer to theft than what's actually happening, which is a more mundane, more real risk: ordinary platform dependency, the same kind that's existed since the first company built a product on top of AWS or Stripe.
What I'd actually do
Don't panic about vendor lock-in, and don't rebuild your stack around Karp's framing. Do the concrete audit instead: pull up your system prompts and count how many contain vendor-specific tricks that wouldn't survive a provider swap. Check whether you have an eval set (even ten test cases) that you could run against a second provider today. If the answer is "no eval set, prompts are vendor-specific, never tested a fallback," spend one afternoon this month fixing that, not because the labs are colonizing you, but because model deprecations and pricing changes happen on a schedule you don't control, and you'd rather find out what breaks on your own timeline than theirs.
Author
Lukas
@lukcombinator