Sovereign wealth funds are pressing Perplexity to let them run its AI on their own hardware, with a hard condition: the company gets no access to their data. The demand, which hasn't been publicly detailed by the funds themselves, marks a new kind of pressure on AI developers that sell into government and institutional markets.
For Perplexity, the ask is straightforward but technically loaded. The funds want the model running behind their own firewalls, completely disconnected from Perplexity's servers. No telemetry, no queries routed back to the company, no way for the AI firm to see what its models are being asked. In practice, that means Perplexity would have to hand over a system it can't monitor or update remotely.
The data sovereignty squeeze
Data sovereignty isn't a new concept. Governments and large financial institutions have spent years pushing back on cloud services that store or process data outside their borders. What's changing is the target. It's no longer just storage providers and database vendors getting the treatment. AI companies are now in the crosshairs because the models they build ingest everything they're fed.
A sovereign wealth fund managing billions in state assets can't afford to let an outside company see its investment research, its portfolio strategy, or whatever else it types into an AI tool. The legal exposure alone is enough to kill most deals. Then there's the geopolitical angle: some funds are based in jurisdictions that treat foreign data access as a national security matter, not a procurement preference.
So the funds are making the same demand they've made to other vendors, now aimed at Perplexity: either the AI runs locally, or it doesn't run at all.
What Perplexity would have to give up
Running on-premise without data access isn't just a deployment option. It's a different business model. Perplexity's standard product is cloud-based. The company controls the servers, pushes updates, and can see usage patterns. That visibility is how it improves models, detects abuse, and bills customers.
An on-premise version with zero data access strips all of that away. Perplexity wouldn't know how many queries a fund is running, what kinds of questions are being asked, or whether the model is performing well. Updates would have to be shipped as discrete versions, likely installed by the customer's own IT team. Any bug fix becomes a scheduled event, not a silent patch.
The company hasn't said publicly whether it's pursuing this kind of deal. But the demand is coming from a customer segment that doesn't usually take no for an answer.
The same pressure is building across the AI industry. Any company selling large language models to governments, banks, or funds with strict data rules runs into the same wall. The vendors that can offer a fully isolated deployment—no phone home, no telemetry, no vendor access—get the contracts. Those that can't get left out of a lucrative and growing market.
That's the adaptability problem. AI firms built on cloud infrastructure and continuous data collection have to rebuild their architecture if they want sovereign clients. It's not a simple switch. It means separating the model from the service, which cuts into the recurring revenue that makes AI companies valuable.
For now, sovereign wealth funds are the ones setting terms. They have the capital, the legal leverage, and the alternative of building in-house or using open-source models. Perplexity and its peers have to decide how much control they're willing to give up to keep the business.
The next move is Perplexity's. The company hasn't announced an on-premise product or a sovereign deployment program, and it hasn't said whether it's negotiating with any specific funds. Until it does, the demand stays on the table—and the funds keep looking for AI vendors that will meet it.
