Confidential inference, GPU TEEs, private model hosting
Attested AI
Your customer's security team says: prove it. Hand them something they can verify themselves.
Your customers send prompts, documents and weights to a machine they cannot see. Attestation lets them check what runs there before they trust it. Re-attestation keeps that check true after every model swap, driver update and node replacement.
Who this is for
- Confidential AI and private inference vendors, selling to banks, hospitals and public bodies
- Companies buying AI, security and procurement teams who must check a vendor's claim
- Cloud and GPU providers, offering confidential computing to AI customers
When trust has to be proved again
A new model or adapter is loadedThe measured software changes, so the evidence a client checked yesterday no longer describes the service.
GPU driver, firmware or CUDA stack is updatedReference values must be updated and nodes re-attested before they take traffic. Inference nodes scale out and inEvery new node must prove itself before it joins; a node that skipped this is a silent hole.
Keys and sessions rotateSecrets released after attestation must be bound to a fresh proof. Last month's proof is too old.
What we deliver
- An attestation design for CPU and GPU TEEs (Intel TDX, AMD SEV-SNP, NVIDIA confidential computing) end to end, from evidence to key release.
- A re-attestation policy: which events force a new proof, how fast, and what happens to traffic meanwhile.
- Client-side verification your customers can run themselves, so 'confidential' is something they check, not something they read.
Questions worth answering
- Can a customer verify which model version answered their prompt?
- What does your service do when a GPU fails attestation at 3 a.m.?
- Is the key that decrypts user data bound to a measurement, or just to a VM?
Across the life cycle
DesignBuildMake or deployOperateUpdateRetire
The stage that decides here: update. Models and drivers change every few weeks. The update is the stage that decides everything here. See attestation in the product life cycle.
Rules and standards your buyers cite
EU AI ActDORANIS2ISO 27001 and SOC 2
Industries that use this
Healthcare and medical dataPharma and life sciencesBlockchains and decentralised networks
Use case
A private inference service sold to a bank
Go deeper
- How the platforms attest: Intel TDX and SGX, AMD SEV-SNP, NVIDIA GPUs, Microsoft Azure, Google Cloud, Amazon Web Services
- Terms explained: Remote attestation, Attestation freshness, Trusted execution environment (TEE), Confidential AI, GPU attestation, Attested TLS, TCB recovery, Quote, report, token: names for evidence, Attestation across vendors, Enclave, trust domain, realm: one idea, many names, Library operating systems for enclaves
- Read next: What evidence means, and what it does not, Who runs attested AI in production, What broke lately
Who builds in this area
From our public landscape: NVIDIA Confidential Computing, Apple Private Cloud Compute, Tinfoil, Privatemode AI, BeeKeeperAI, Mithril Security, Azure Confidential AI, WhatsApp Private Processing, Google Private AI Compute, NEAR AI Cloud, Confident Security (OpenPCC), Super Protocol, dstack.
Not quite your case? Building agents on top of models? Attested AI agents. Need to show where a model came from? AI passport.
Working on Attested AI? A first call takes 30 minutes and commits you to nothing. Bring your product and the one question that worries you.
Next fixed date: EU Digital Identity Wallets are due in every member state on 24 December 2026, 42 days from now.
Concepts: Re-attestation · Continuous attestation · Remote attestation · IETF RATS architecture