Karp is right: controlling your weights is controlling your fate. Here is how Europe does it

On July 1, on CNBC's Squawk Box, Palantir's chief executive walked on set to present his company's new sovereign AI alliance with NVIDIA and ended up delivering the most talked-about interview of the year on financial television. Palantir shares closed the day up more than 9%, clips of the interview went viral on X, and the word 'sovereignty', which Europe has spent years pronouncing almost defensively, became, in the mouth of one of Silicon Valley's most influential executives, the investment thesis of the moment.

At GPU Solutions we have been building exactly what Karp describes since day one. This article covers what he said, why the diagnosis is correct, what he forgot to say about Europe, and how all of it applies if your company operates under European jurisdiction.

The day Wall Street discovered sovereignty

Context matters. The day before, Palantir had published a nine-point manifesto on AI sovereignty on X that spread like wildfire, and had just announced with NVIDIA the deployment of the open Nemotron models inside its sovereign platform for the U.S. government and critical infrastructure: per-customer data isolation, full control of model weights, complete portability and auditability. Against that backdrop, Karp went after generative AI's dominant business model, metered token consumption from closed labs, with a line that carries half the argument:

The manifesto even coined a term for the pathology, 'tokenmaxxing' (maximising token consumption as an end in itself), and left the most quoted line of the week: 'controlling your weights is controlling your fate'. Plus one uncomfortable question for the labs: if tokens created so much value, why do those selling them refuse to charge based on value?

The diagnosis is correct

Stripped of the television spectacle, Karp's technical argument is hard to refute. When a company consumes AI exclusively through a closed lab's API, it is renting intelligence on the landlord's terms. Karp framed it as elementary due diligence: 'clients have to be able to ask and answer very basic questions: are you keeping the data? are you going to enter our business?'. Here is what you do not control when a closed lab's API is your only path to AI:

• The weights: they are not yours and you cannot take them with you. If the provider raises prices or retires the model, you start from zero.
• The data: it travels to someone else's infrastructure, under someone else's jurisdiction, with contractual guarantees you never audit.
• The price: per token, unilaterally revisable, and decoupled from the value it generates in your business.
• The roadmap: the model your product depends on today can be deprecated tomorrow without you having a say.
• The knowledge: every prompt teaches your provider how your business works, the 'alpha' Karp describes.

The irony Karp did not mention

Here is the thing: Karp's sovereignty is American sovereignty. His rhetorical question (whether America is really going to outsource its battlefield to the consensus view of Silicon Valley) answers itself within his borders. Seen from Europe, the irony is double: European governments and regulators have spent years raising about Palantir itself the same dependency concerns Karp raises about the AI labs, and any American provider, whichever it is, remains subject to the CLOUD Act, which compels the surrender of data to U.S. authorities regardless of where the server sits.

The conclusion for a European company is uncomfortable but liberating: sovereignty cannot be outsourced to someone who shares your dependency. If controlling your weights is controlling your fate, the European corollary is that the infrastructure where those weights live must sit under European jurisdiction, operated by European capital, with European certifications. Not as protectionism: as elementary coherence with Karp's own argument.

Tokens, yes, but without surrendering the weights: the ownership ladder

This is where the argument needs sharpening, because we sell tokens too and we are not going to pretend otherwise. The problem is not the token as a billing unit: it is the token as a dead end. A closed-lab token leads nowhere: the more you consume, the more you depend. A token from an open-weights model served on sovereign infrastructure is something else, the first rung of a ladder that ends in ownership.

That is the three-layer architecture we have spent months building and which has already served more than a trillion tokens: the infrastructure layer (NVIDIA B200 GPUs in Madrid's only Tier III datacenter, with ISO 27001 and ENS), the compute-access layer (GPU Flow, in its three variants: tokens, sandboxes and dedicated GPU) and the application layer (the Odiverse suite). All sovereign and, this is the important part, all with the exit door open. Because the sovereignty we defend includes the right to leave: including from us.

'Own the means of production'

Karp closed his argument with a phrase of curious resonance coming from an American CEO: his technical customers want to 'own the means of production', meaning their compute, their models, their data stack and their alpha. We cannot improve on it: that is exactly what we sell, with the difference that here the means of production sit in Madrid, the invoice arrives in euros and the law that governs them is voted in Europe. If your current provider cannot answer Karp's two questions (are you keeping my data? who controls the weights?), perhaps it is time to make the call. The architecture conversation is free.