In July, Palantir chief executive Alex Karp went on CNBC to promote a sovereign AI partnership with NVIDIA, and spent most of the interview laying into the business model of the frontier labs. It was combative and at times hard to follow, and the clip is still doing the rounds on LinkedIn months later. Strip out the theatre and one line holds up, and it is the whole case for sovereign AI inference.
What technical customers want is control over their compute, their models, their data stack, and their alpha.
He is right. It is worth taking seriously precisely because of who said it. Palantir is not a sovereignty start up with a point to prove. It is one of the largest software suppliers to the US government, and it is telling the market that renting intelligence from somebody else’s API carries a cost that never appears on the invoice.
Four questions every buyer should ask
Karp’s prescription for rebuilding trust was simple. Everyone should be able to ask and answer basic questions: who owns the data, where is it cached, are the prompts secure? Add the one he kept returning to, is any of it being transferred to the provider?, and you have a checklist to put to every AI vendor. We set out our own version in five questions for your AI vendor.
Most vendors will answer. The trouble is how. The answers come back as policy: we do not train on your data, we keep it for 30 days for abuse monitoring, we process it in an EU region. Each of those is a promise about behaviour. None of them describes who can read the content, or under which law. We looked at how much a provider can see in practice when we read Anthropic’s own threat report.
Your prompts are the asset
Karp’s word for what leaks is alpha. In plainer terms, your prompts are a map of your business. They carry your customers’ questions, your internal documents, your code and the workflows you have spent years refining. Send them to a provider that also builds products, and you are describing your business to a company that may one day compete with it.
That is no longer a theoretical concern. The frontier labs are not only model suppliers any more. They ship coding tools, design tools, security tools and agents, moving into markets once served by the companies building on their models. Nobody has to break a contract for that to matter. Knowing which workloads are growing and what users ask for is valuable whether or not a single prompt is ever used for training.
The true cost is not on the invoice
Karp’s sharpest line was about pricing. If a model can make you a billion dollars, why charge for tokens rather than take a share of the value? His answer was that the true cost is what you make minus what you lose, including the value of your business that walks out of the door.
We charge for tokens too, and we think that is the honest model. Inference is a utility, like power or bandwidth, and it should be metered like one. The problem is not the meter. It is what else leaves the building with every request.
Where the argument stops short
Karp framed his answer around American infrastructure: open models, an application layer and compute, kept onshore. For a US defence customer that is coherent. For a European business it skips the part that matters most.
Under the US CLOUD Act, American authorities can compel a US provider to hand over data in its possession, custody or control, wherever in the world that data is stored. A European data centre run by a US company is still within reach. An EU region answers where data sits. It does not answer whose law decides who can read it. We covered that gap in why an EU region is not sovereignty.
Control is not a contract clause, and it is not a region drop down. It is a property of the architecture: who runs the hardware, what is retained, and which courts can order the operator to hand something over.
Sovereign AI inference: our answers to his questions
Who owns the data? You do. TensorX is an Irish company with no parent outside the EU, so no foreign government has a legal hook into us.
Where is it cached? Prompts and completions are held in volatile memory for the lifetime of the request and are never stored, on GPUs in Dublin and Helsinki. There is no prompt log to cache. The detail is on our trust and security page.
Are the prompts secure? They are encrypted in transit, and we do not log, read or train on them. We do not need a promise about inspection, because nothing is kept to inspect.
Is anything transferred to us? Nothing that outlives your request. We sell inference. We do not build applications that compete with our customers, and no customer data is shared with the companies that publish the models.
Who controls the weights? We serve open weight models from DeepSeek, Z.ai, Moonshot AI, Qwen and MiniMax. The weights are published, so nobody can withdraw them or quietly change them under you. If you want the hardware to yourself, dedicated inference gives you reserved GPUs in the EU.
The 90 day test
The LinkedIn post that brought the interview back around ends on a good question: could you switch AI provider in 90 days without losing your edge? It is the right test, because lock in is the quiet half of the control problem.
With TensorX the API is OpenAI compatible. Moving in is a change of base URL and key, and so is moving out. Your prompts, evaluation sets and workflows stay yours, because they never lived with us in the first place. A provider you can leave in an afternoon is one that has to keep earning the work.
Karp asked the right question. For Europe, the answer is not a more trustworthy vendor. It is a provider that sits outside the reach of the laws you are worried about, and keeps nothing worth asking for.
You do not want a provider who promises not to look. You want a provider who structurally cannot.
Frequently asked questions
What did Alex Karp say about AI control?
On CNBC in July 2026, Palantir’s chief executive said technical customers want control over their compute, their models, their data stack and their alpha. He argued buyers should be able to ask who owns their data, where it is cached and whether their prompts are secure.
What is sovereign AI inference?
Sovereign AI inference means running AI models with a provider that sits under the jurisdiction you choose, so no foreign law can compel access to your prompts or outputs. For European buyers that means an EU operator with no parent outside the EU, not only an EU data centre.
Does an EU cloud region protect data from the US CLOUD Act?
No. The CLOUD Act lets US authorities compel a US provider to hand over data in its possession, custody or control, wherever that data is stored. Moving the data to Europe does not move the provider out of US jurisdiction.
Does TensorX store or train on prompts?
No. Prompts and completions are held in volatile memory for the lifetime of the request and are never stored, logged or used for training.
Can I switch to TensorX without rewriting code?
Yes. The TensorX API is OpenAI compatible, so moving is a change of base URL and API key.
Related reading: Europe is sleepwalking into US AI dependence, and Ireland just set the alarm. What the EU AI Act enforcement wave means for where your data actually runs.
