Inside an American government data centre, a system called ImmigrationOS is assembling a stranger’s life.

It reaches for a passport record, then a Social Security file, then a year of tax returns, then a licence-plate read captured by a motorway camera, and it stitches them into a single profile that tells an enforcement officer whom to find next.

The system does not know the person, but it knows the data, and it works with that data regardless, as critics have noted, of the veracity or accuracy of the databases it draws from.

ImmigrationOS was built by Palantir Technologies under a $30 million contract with US Immigration and Customs Enforcement, with a prototype due in September 2025 and the work running through 2027.

It is one of the clearest expressions of what the company actually does for a living. Which is worth holding in mind, because on 3 August 2026, in a letter to shareholders, Palantir’s chief executive, Alex Karp, reached for a much older frame of reference.

“There are Marxist overtones and undertones to our business,” he wrote.

The charge that followed was aimed at the companies building large language models. They intend, “knowingly or otherwise, to capture the means of production of their purported partners.”

Pay a frontier lab to help with your artificial intelligence, the argument runs, and you are really paying “for the right for them to migrate your IP, your know-how, your expertise to their model,” until they can build a competitor that needs neither your business nor your people.

On the earnings call he said it with more relish still: the labs believe “they deserve to colonize your enterprise.”

The quarter behind him was extraordinary by any measure: revenue of $1.9 billion, up 93% on the year, and $1.1 billion in profit, which is to say more profit in three months than the company booked in total revenue in the same period a year earlier.

Full-year guidance was lifted to 82% growth. When a company is compounding like that, its founder gets to philosophise.

The borrowing is at least legible. In Marx, the means of production are the things you need in order to make things: the factory, the machines, the land, the tools.

Whoever owns them owns the process, and can pay the worker for their labour while keeping the gap between what that labour produces and what it costs. Marx called that gap surplus value, and the arrangement exploitation.

Karp’s move is to lay this template over the AI economy. The modern factory, he suggests, is the model, and the frontier labs are quietly enclosing it, turning their own customers into the raw material.

Palantir’s own framing gives the idea a certain poetry. The company promises to let customers keep not just their data but their AI “exhaust,” the prompts, orchestration, and context thrown off as a by-product of using these systems.

It is a revealing phrase. In his telling, the labs are the smokestacks and everyone else is left breathing the fumes; Palantir hands you back the smoke and calls it sovereignty.

There is something to the observation, which is precisely what makes it so useful to Palantir. This is not the first time the company has turned its sales pitch into a manifesto.

Its recent doctrine of AI sovereignty, as we have argued, works as strategic branding dressed up as principle: institutions are urged to hoard their data and own their model weights, which happens to be an exact description of the on-premises software Palantir sells.

The Marx letter is the same manoeuvre in a heavier coat: rebrand your product as the resistance, and every competitor becomes the exploiter.

But take Marx seriously for a moment, more seriously than any shareholder letter can, and the lens does not settle where Karp points it.

In an information economy, the means of production are not only the models. They are the pipes: the integration layer that gathers an institution’s scattered data into one place and makes it legible, searchable, and actionable.

That layer is exactly what Palantir sells. Foundry and Gotham exist to become the operating system of a government or a corporation, the thing without which the daily work no longer runs.

Ownership of that layer is its own kind of enclosure, and dependence on it is its own kind of capture.

Palantir’s customers understand this, which is why the sovereignty talk cuts in both directions. Several European governments have started to treat Palantir itself as the lock-in risk rather than the escape from it.

We have covered about a German military rejection and a broader unease about wiring the machinery of the state to a single American vendor. The company that warns you against migrating your know-how into someone else’s model is also the company you cannot easily walk away from once your institution runs on its software.

Marx had a name for the fear of the tool that ends up owning you back, but Alex Karp has a product line.

Here the analogy turns genuinely uncomfortable, because Marx was not, in the end, writing about quarterly margins. He was writing about power: who holds it, over whom, and by what machinery.

The purest example of Palantir’s business is not a design agency losing its edge to a chatbot. It is ImmigrationOS reading a person’s records in order to decide their fate, or Project Maven, the military targeting programme that Google’s own engineers found too troubling to keep, and which Palantir was reported to have taken on.

Concentrated informational power over human beings is the product. A company that builds the apparatus of the state’s gaze, invoking the philosopher of the powerless, is not producing an overtone. It is a note struck very loudly.

None of which makes him wrong about the labs. The tension he describes is real, and he is not alone in naming it; even Microsoft’s Satya Nadella has warned that the frontier labs increasingly compete with the very partners they depend on, spreading into healthcare, law, and drug discovery.

A customer who pours its proprietary knowledge into someone else’s model may well be feeding a future rival. Wanting to own your data rather than surrender it is a defensible thing to want, and a reasonable thing to sell. The critique is not empty, it is just conveniently one-directional.

Which returns us to the data centre, and the machine still reading a stranger’s life. Marx would recognise it at once, though not as he intends him to.

He would not see a plucky enterprise resisting colonisation, but he would see the means of production humming along nicely, owned by whoever owns the platform, extracting their value from people who never agreed to become the input.

The overtones are Marxist, right enough, I will give that, but it is the casting that is off.

Get the TNW newsletter

Get the most important tech news in your inbox each week.