Last week, one of the most watched interviews in enterprise software was not about a product. It was a warning. Palantir’s Alex Karp went on television and said out loud what a lot of enterprises have been muttering privately: companies are paying for AI by the token while handing over something far more valuable in return. Their data. Their workflows. Their edge.
Strip away the theatrics and the point underneath is solid. The real question in enterprise AI is no longer which model is smartest. It is who ends up knowing your business.
What your AI vendor learns about you
Think about what actually leaves your walls when your teams use a hosted AI service. Not just documents. Prompts that describe how you price, how you underwrite, how you resolve disputes, how your engineers structure systems. Feed a model your daily work for a year and it has observed something no consultant ever could: the full operating logic of your company.
Now add the part most people missed. The large model providers are no longer just selling models. They are moving up the stack, building applications, agents and design tools of their own. We have already watched providers launch products that compete directly with companies built on top of them. If your provider can build products, and it can see how your business works through everything you send it, the question rather answers itself: why keep sending it?
You do not need to accuse anyone of wrongdoing. The risk is structural. The incentive exists, the access exists, and once your data leaves your walls you cannot verify what happens to it. For a regulated business, that combination should be enough.
Five questions to ask any AI vendor
Karp’s rant contained a genuinely useful checklist, buried under the noise. Put these to any AI provider before you send them another token of your business:
- Who owns the data once it leaves our systems, and can you prove it?
- Where is it cached, in which country, under whose jurisdiction?
- Are our prompts retained, logged, or used to improve your models?
- Do you build products in our space, or could you tomorrow?
- Who controls the model weights, and what happens to our workflows if access changes overnight?
If a vendor cannot answer these plainly, the answer is the answer.
The market is already voting
This is not a hypothetical shift. Coinbase’s CEO said publicly that the company cut its internal AI spend by nearly half by defaulting engineers to open-weight models, GLM and Kimi among them, served through an internal gateway they control. Microsoft is reported to be weighing open models as an engine for parts of Copilot. The pattern is the same everywhere: keep the intelligence, take back the pipe.
Enterprises are discovering that the top open-weight models are within touching distance of the frontier for most real work, and that running them on infrastructure you trust removes the whole category of questions above. Not answered. Removed.
How TensorX answers the five questions
We built TensorX so that a European enterprise can answer that checklist without blinking:
- Who owns the data? You do. Always. We serve open-weight models and we do not train on your traffic, full stop.
- Where is it cached? In the EU, on hardware we own, under EU jurisdiction. Your requests are processed on our own NVIDIA B300s on European soil and nowhere else.
- Are prompts retained? No. Zero data retention is the default on every request, not an enterprise upsell. We do not store your prompts or your outputs, ever.
- Do we build products in your space? No. We run inference. Your workflows are not our roadmap, and never will be.
- Who controls the weights? Open-weight models mean nobody can switch your stack off, reprice your access overnight, or absorb your edge into their next release.
Trust the architecture, not the promise
Every vendor will promise to be careful with your data. Promises are policy, and policy changes with a funding round, an acquisition, or a strategy memo. Architecture does not.
With sovereign, zero-retention, open-weight inference there is nothing to promise, because there is nothing retained, nothing offshore, and nothing proprietary sitting between you and your own intelligence. Your data cannot be harvested by a provider that never holds it.
You do not have to trust a vendor’s promise not to look. You can choose an architecture where they structurally cannot.
See the open-weight models we serve, or read why zero data retention is the future of AI.
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.
