Xanadu
questions buyers ask

The questions, answered.

Real questions from real diligence calls, including the ones a marketing site would skip. Missing yours? Email hello@xanadu.run and it usually appears here within a week.

Product

  • What model runs inside the enclave?

    An open-weight frontier model whose weights are content-addressed and pinned at attestation time. Current target is a 1T-parameter open-weight model (Kimi K2.6 leading; we plan automatic fallback to DeepSeek V4 Pro and Qwen 3 Coder Next as tier alternates). Customers can verify the running weights match the published hash at session establishment.
  • Can I self-host the cloud?

    Not from us. The architectural protections (Swiss legal home, Iceland hardware, our build pipeline) are part of what you're buying. The open-source dispatcher is fully self-hostable today under MIT.
  • What metadata does the operator see?

    Token counts, dispatch timestamps, and tenant id. Prompt and output contents stay inside the enclave. The threat-model page lists what is retained and for how long.

Billing

  • How do you bill us if you can't see what we did?

    Metering is computed inside the enclave via signed accumulators over token counts. The enclave emits billing telemetry: usage counts, dispatch ids, cost. Nothing else. The ledger row contains exactly what is necessary to invoice, plus a customer-readable dispatch id you can quote back. Anything more would be a leak.
  • Why publish prices instead of gating them behind a sales call?

    Hidden pricing signals the price is whatever the buyer will tolerate. A sovereignty product cannot ship that signal. We publish brackets on /cloud. Design partners lock their price through GA.

Architecture

  • What does “attestation” verify?

    Three things: (a) the code running is the published code (matched by reproducible build hash); (b) the hardware is a genuine NVIDIA confidential-compute GPU (matched by silicon root of trust); (c) the channel is bound to that specific GPU instance (so a MITM swap is cryptographically visible). The customer SDK refuses sessions where any of these fail.
  • What's the performance overhead of confidential compute?

    Measured at 4 to 8 percent over comparable non-confidential inference at our target model size on Blackwell-class hardware. The buyer should not feel a difference at p50 latency. We will publish independent benchmarks alongside the public beta.
  • Why Iceland for the hardware and Switzerland for the legal home?

    Iceland: geothermal power, sub-zero ambient cooling, EU-GDPR-equivalent data-protection law layered with Iceland's own. Switzerland: Article 271 SCC, constitutional privacy under Art. 13, no CLOUD Act, strong rule of law for the buyer segment. Hardware lives where it's economical and lawful; entity lives where the legal protections compound.

Compare

  • How is this different from OpenAI Enterprise zero-retention?

    Zero-retention removes their database. Their ability to read in flight is unchanged. A US court can still compel anything they can read. We remove the operator's ability to read at all.
  • How is this different from running open weights yourself?

    Self-hosting gives you full control and full operational cost. You become the operator and inherit your home jurisdiction's exposure. You build the attestation chain yourself. We are the build-it-once version. Same model class. The operating entity is forming where the legal protection compounds (Swiss Stiftung, in formation).

Operations

  • When does the cloud open?

    Private beta H2 2026, a small design-partner cohort onboarded as the first Iceland nodes attest. Public beta follows after the first transparency audit. The roadmap with concrete dates lives in the design-partner program brief shared after waitlist acceptance.
  • How do I report a vulnerability?

    Email security@xanadu.run or read the full disclosure policy at /security. Reports are read by a human within 24 hours, acknowledged within 72, and patched on a published timeline.