Every AI company says "we take your privacy seriously." Most of them can read your data and choose not to. We’re building Oluri so we structurally can’t — with the hardware-attested proof going live the moment our Supermicro install lands in November 2026. This page tells you exactly what ships today vs. what ships with the hardware, with no marketing gap between the two.
A privacy claim is only as strong as the layer underneath it. Most consumer AI rests on a contract: we have your data; we promise not to misuse it. Oluri rests on hardware: we don’t have your data in a form we can read.
Oluri runs on a server we own, in a Flexential data center in Nashville. There is no call out to OpenAI. No call out to Anthropic. No call out to Google. When you message Oluri, the AI that thinks about your message is the AI on our box. The thinking never leaves the building.
Message transport — the honest detail. Today, the iMessage relay between your iPhone and our box runs through a third-party partner (the same industry-standard pattern every consumer AI on iMessage uses, including our competitors). It’s the only surface where bytes pass through someone else briefly. We’re closing this gap by running our own Mac-mini iMessage relay in the same Flexential cage as the GPUs — engineering is built and committed (see the migration spec). Target cutover: July 2026. Until then, sensitive content stays on Oluri’s box and reaches you via auth-gated links rather than transiting in plaintext.
The GPUs (NVIDIA RTX PRO 6000 Server Edition) run in a hardware mode called Confidential Compute. In this mode, the GPU itself refuses to expose its working memory to anyone — including us, including our staff, including a person with physical access to the machine — while it’s thinking about your data. It is enforced by the silicon, not by software we wrote.
Every credential you give Oluri (logins, payment methods, OAuth tokens) lives in an encrypted vault. The PIN you set gates access to that vault — three wrong attempts and you’re locked out; lose the PIN and we cannot reset it. Per-user PIN-derived encryption of the full memory store is on the November launch roadmap alongside the Flexential hardware install; until then, the vault is encrypted with a system-managed key that sits inside the same confidential-compute boundary as everything else — protected by the hardware, not exposed to the open internet. The trust page updates when the upgrade ships, with the verifiable proof attached.
Every time the Oluri GPU starts up in production, NVIDIA’s attestation system produces a cryptographically signed report saying "this GPU is running in Confidential Compute mode, this is its firmware hash, this is its certificate." You can read that report and verify it independently against NVIDIA’s public attestation service — we don’t get to fake it. Today, the report below is a transparent placeholder — the Supermicro chassis lands in Nashville the week of November 11 2026, and the real signed values populate the moment the box comes online. This page updates the day the proof goes live.
Refreshed at every boot. The signed report is fetchable at /trust/attestation.json and verifiable against NVIDIA’s Remote Attestation Service (NRAS).
If you can read NVIDIA’s attestation spec, you can verify this yourself. If you can’t, the short version is: the GPU itself signed a report saying its memory is private, and nobody can produce that signature without the actual hardware.
If we ever changed any of this, we’d tell you the day we did, in plain English, with a clear "here’s what changes, here’s what stays the same, here’s what you can do about it." That’s a promise. The rest is hardware.