llms.txt
Compare with other sites' llms.txt files →# Confidential AI
> We sell private inference and confidential GPU VMs. Buy tokens, rent confidential GPUs, or license the stack for your own GPUs. Verified by Intel, AMD and NVIDIA signatures, not by us.
Two ways to run Confidential AI. Cloud: our infrastructure, usage-based. Licensed: your infrastructure, per hardware unit, per year.
Full pricing: https://confidential.ai/pricing
## Confidential Inference (Cloud)
OpenAI-compatible endpoint. Open-weight models in TEEs. Your client verifies the model, the code and the hardware before the first token. Overhead: about 1 to 3% token throughput on B200s (https://arxiv.org/abs/2608.26575), negligible impact on time to first token.
| Model | Status | Input (per 1M tokens) | Input Cached (per 1M tokens) | Output (per 1M tokens) |
| --- | --- | --- | --- | --- |
| Kimi K3 | Request capacity | $3.00 | $0.30 | $15.00 |
| GLM 5.3 | Request capacity | $1.40 | $0.26 | $4.40 |
| Qwen3.8 27B | Request capacity | $0.30 | $0.03 | $1.80 |
| DeepSeek V4.1 Flash | Available | $0.30 | $0.03 | $1.20 |
Cached input: prompt-prefix tokens served from cache, billed at the cached rate.
Billing: metered per token, invoiced monthly in arrears. Card or bank transfer.
How to get access: email hello@confidential.ai with models, expected volume and use case, or use the form at [request access](https://confidential.ai/request-access). DeepSeek V4.1 Flash is available; other models on request. Provider integrations for inference routers: hello@confidential.ai.
Docs: https://confidential.ai/docs/inference-api/get-started
## Try it live: /try
https://confidential.ai/try is our live production inference cluster. Verify it in your browser: Intel TDX quotes, image measurements, the admission allowlist and mesh identity, checked client-side. No Confidential AI server takes part in the check.
Model served: DeepSeek V4.1 Flash. Endpoint: https://api.confidential.ai
## Confidential VMs (Cloud)
Rent a full node or individual GPUs; each rental is one confidential VM with root over SSH. Blackwell nodes are Intel TDX, up to 8 GPUs per node in NVIDIA Confidential Computing mode. CPU and GPU attestation from inside the guest.
GPU VMs, per GPU-hour:
| GPU | VRAM | Host CPU TEE | GPU CC mode | Status | Per GPU-Hour |
| --- | --- | --- | --- | --- | --- |
| B300 | 288 GB HBM3e | Intel TDX | Multi-GPU, NVLink encrypted | Available | $7.50 |
| B200 | 192 GB HBM3e | Intel TDX | Multi-GPU, NVLink encrypted | Available | $6.50 |
| H100 | 80 GB HBM3 | Intel TDX | Protected PCIe, NVLink not encrypted | Request capacity | $3.25 |
Available: allocated within a few days of a signed order. Request capacity: send configuration and duration and we confirm lead time.
Single GPU pass-through is available on every listed GPU. Multi-GPU with encrypted NVLink: B200 and B300. Protected PCIe (multi-GPU, NVLink and NVSwitch unencrypted): H100.
CPU VMs, per vCPU-hour (1 core = 2 vCPUs) plus per GB-hour of RAM:
| TEE Backend | Per vCPU-Hour | Per GB-Hour (RAM) |
| --- | --- | --- |
| AMD SEV-SNP | $0.05 | $0.012 |
| Intel TDX | $0.05 | $0.012 |
How to get CVMs:
Confidential VMs are provisioned by us. A console and API are on the roadmap.
1. Email hello@confidential.ai with GPU class, GPU count, region, start date and duration.
2. We confirm availability and send an order form.
3. Return it signed, with the legal name and billing address of the individual or company, a technical contact, and the SSH public key to install. We run basic KYC on all dedicated-hardware orders.
4. First invoice: 50% of the order, by bank transfer. On payment we allocate your VM and send SSH access.
5. Final invoice: 50% on completion. Longer engagements are invoiced monthly in arrears.
Minimum order $500. Minimum term 3 months; a short evaluation of up to a week is available first. Rent one, two, four or eight GPUs; each rental is one confidential VM. Billing runs for the reserved period whether the VM is running or not. USD by bank transfer; card payment is not accepted for dedicated hardware. No annual commitment required. Longer engagements: rates slightly below list, depending on term. [Ask](mailto:hello@confidential.ai).
Regions: United States (South Central) and Central America. EU coming.
SLA and support terms are set in the order form or master agreement. Ask for current terms.
Details: https://confidential.ai/pricing
## Licensed
The full stack on your own infrastructure: on-prem, bare metal, and all major clouds. One license per GPU per year, or per CPU core per year; NVIDIA CC licensing included. Annual license, invoiced monthly in arrears. No per-component line items. Software updates and standard support included.
What you get:
- Confidential Metal: attestable, verifiable confidential VMs on bare metal.
- C8s: Confidential Kubernetes. Scale AI workloads to data center scale.
- AI Workload Services: confidential inference, training and fine-tuning.
- Confidential OS: hardened VM guest OS for development and production.
- Client Libraries & SDKs: clients verify confidentiality claims themselves.
GPU licenses, per GPU per year:
| GPU class | Per GPU-Year |
| --- | --- |
| RTX PRO 6000 | $2,789 |
| H100 | $3,876 |
| H200 | $4,974 |
| B200 | $7,625 |
| B300 | $9,884 |
CPU-core license: $48.85 per core-year. Physical cores, not vCPUs. Not required for cores in GPU machines.
Support:
| | Standard | Enterprise |
| --- | --- | --- |
| Price | Included | 25% of total license price |
| Coverage | Business hours | 24/7 for production outages |
| Response, production outage | 1 business day | 1 hour |
| Contact | Shared queue | Dedicated technical account manager |
Example: 500 B200s: 500 x $7,625 = $3,812,500/year, invoiced monthly at about $317,708. That is $635 per GPU per month, with the full stack, NVIDIA CC licensing, updates, and support included.
Contact: hello@confidential.ai
## Private model weights
Serve your weights on Confidential Inference or on the licensed stack. Weights run inside the TEE: the operator cannot read them, and downstream clients verify the exact model before the first token. Model owners and their users are both protected.
## Private training and fine-tuning
AI Workload Services (licensed) runs training and fine-tuning inside confidential VMs on Intel TDX or AMD SEV-SNP, with NVIDIA GPUs in Confidential Computing mode. Training data and checkpoints stay encrypted in use.
## How it is verified
Every workload runs inside a confidential VM on Intel TDX or AMD SEV-SNP, with NVIDIA GPUs in Confidential Computing mode. The host, the hypervisor and Confidential AI cannot read memory in use. CPU overhead is about 1.5%.
Hosted inference: before you send anything, your client fetches a hardware-signed attestation and verifies the Intel TDX evidence: exact firmware, kernel, root filesystem, container digests and model. That evidence includes the boot policy, which only admits nodes whose GPUs pass NVIDIA attestation. The NVIDIA GPU report itself is not included in the evidence today. If anything does not match, nothing is sent.
Confidential VMs: you collect CPU and GPU attestation from inside the guest, an Intel TDX quote and the NVIDIA GPU attestation report, and verify both yourself.
The verifier is open source and the builds are reproducible. Run the check yourself.
## FAQ
### General
What Confidential AI is, what it costs, and how it compares.
What is Confidential AI?
Licensed software and cloud services for running AI infrastructure on trusted execution environments (TEEs). We sell Confidential Inference, Confidential VMs, and the licensed stack for your own GPUs, and your client verifies the attestation before it sends anything.
What does it cost?
Confidential Inference is metered per token: DeepSeek V4.1 Flash is $0.30 input and $1.20 output per 1M tokens. Confidential VMs rent by the GPU-hour: B300 at $7.50 per GPU-hour, B200 at $6.50. Licensed deployments are per GPU per year, from $2,789. Full pricing: pricing (https://confidential.ai/pricing).
How does Confidential AI compare to AWS Bedrock and Azure?
Bedrock's zero retention mode has abuse detection exceptions, and Azure Foundry stores flagged prompts and completions for human review unless you are approved for modified abuse monitoring. Confidential Inference rests on an attestation your client verifies before sending anything. The tradeoff is model choice: we serve open-weight models only.
Can I serve my own model weights privately?
Yes, on Confidential Inference or on the licensed stack. Weights run inside the TEE: the operator cannot read them, and downstream clients verify the exact model before the first token. Model owners and their users are both protected.
Can I train or fine-tune privately?
Yes, on the licensed stack. AI Workload Services runs training and fine-tuning inside confidential VMs on Intel TDX or AMD SEV-SNP, with NVIDIA GPUs in Confidential Computing mode. Training data and checkpoints stay encrypted in use.
Can I run confidential agents?
Yes, on top of our Confidential VMs and Confidential Inference offering. We have done so: the agent runtime and inference both run in TEEs. If you want to run agents confidentially, talk to us (mailto:hello@confidential.ai).
### Confidential Inference
Our hosted inference endpoint runs on C8s, our confidential Kubernetes. The same answers apply if you license the stack and run it on your own GPUs.
How do I verify the endpoint from my client?
The quickest check needs no code: open /try (https://confidential.ai/try) and verify the live cluster in your browser. For inference from your own client, follow the attestation guide (https://confidential.ai/docs/inference-api/attestation) and its current release verification profile. Verify the signed release, pin the reviewed allowlist and deployment identity, and send requests through TEErminator (https://github.com/confidential-dot-ai/TEErminator). This local proxy checks the upstream connection before forwarding data and blocks requests that fail verification.
Do you verify NVIDIA GPU evidence?
Yes, at boot: before a Kubernetes node with GPUs can run workloads, it checks that each of its GPUs is in NVIDIA Confidential Computing mode and passes NVIDIA attestation, and it powers off if any check fails. That check is part of the measured image your client verifies; your client receives the result of the check, not each GPU's raw report.
Is attestation per request or per session?
Verification is bound to the upstream connection, rather than performed for every inference request. TEErminator checks the accepted measurements, policy and identity before forwarding data. Certificate and verdict-cache lifetimes apply; an allowlist update does not guarantee immediate disconnection of every open stream. Review and approve new release and deployment pins before accepting an upgrade.
What does attestation prove?
That every node is genuine confidential-computing hardware that booted the exact measured image, and that only allowlisted workloads can run. An application allowlist entry pins the image digest, command line, environment and mount restrictions, and specifies secret grants. A secret grant does not pin the secret contents.
Can I inspect all the code?
The hosted endpoint runs open, reproducible builds. Dedicated, performance-optimised deployments may include proprietary components. Attestation still proves exactly which image runs, but not all of it is open source.
Is the prompt cache shared between customers?
On the shared hosted endpoint, prefix caching is not isolated per account today. Per-account or per-tenant cache isolation is available on dedicated deployments.
Which OpenAI API features are supported?
Chat completions, streaming, structured output and tool calling, all OpenAI-compatible. DeepSeek V4.1 Flash is text-only.
What happens to model IDs when a model is upgraded?
V4.1 is the improved replacement for V4. We now route all V4 requests to V4.1 automatically. Your existing model ID continues to work. Both deepseek-ai/DeepSeek-V4-Flash-0731 and deepseek-ai/DeepSeek-V4.1-Flash use the V4.1 backend. Responses retain the model ID you requested.
What are the rate limits? Can I load test?
Each key has rate limits; above them the endpoint returns 429 with a Retry-After header, which clients should honour. Tell us before you load test. Reserved capacity is available for guaranteed throughput.
### Confidential VMs
Confidential VMs give you root inside a measured confidential VM, so you collect and verify attestation yourself.
How do I get GPU attestation?
From inside your VM: you can fetch both the CPU attestation (Intel TDX or AMD SEV-SNP) and the NVIDIA GPU attestation report, and verify them with our open-source tools or your own.
What do I control?
Root over SSH inside a measured confidential VM, with no access to the bare-metal host. Custom guest images are part of Confidential Metal (licensed), not rented VMs.
Is my VM on a shared host?
8 GPUs is a whole node, nothing shared. With 1, 2 or 4 GPUs, other tenants' confidential VMs may run on the same host. The hardware (Intel TDX or AMD SEV-SNP) keeps your memory unreadable to them, to the host and to us.
Is NVLink encrypted?
Yes on B200 and B300: multi-GPU with encrypted NVLink. H100 multi-GPU uses Protected PCIe, so NVLink is not encrypted.
Does storage persist across restarts?
Not yet. Encrypted persistent storage across VM restarts is not available. Treat each start as a cold load.
Can I get bare metal?
No. We provision the confidential VM and keep the host layer. To run your own hypervisor, use a bare-metal provider, or license Confidential Metal for your own hardware.
### Licensed
The full stack on your own infrastructure: on-prem, bare metal, and all major clouds.
What is C8s?
C8s is Confidential Kubernetes: our licensed stack for running AI workloads at data center scale on TEEs. Each node runs in its own confidential VM, and only allowlisted workloads can run.
What do I get with a licensed deployment?
The full stack on your own infrastructure. One license per GPU per year, NVIDIA CC licensing included, with software updates and standard support included.
Can I bring my own image?
Yes, on a licensed deployment. Bring-your-own-image is part of Confidential Metal. Rented Confidential VMs run our measured images; custom guest images and kernels are not available on rented VMs.
### Shared (all products)
These answers apply across Confidential Inference, Confidential VMs and licensed deployments.
What do you retain?
We don't store prompts, outputs or workload data. Metadata, such as usage for billing, is covered by our Terms of Service (https://confidential.ai/legal/terms-of-service) and Privacy Policy (https://confidential.ai/legal/privacy-policy).
Can Confidential AI or the hosting provider see my data?
Memory in use is encrypted by the hardware, in the VM and on the GPU. For the hosted endpoint, see our threat model (https://github.com/confidential-dot-ai/confidential-inference/blob/main/docs/threat-model.md) for the full trust boundary, including what our operator key can and cannot do.
Do you support HIPAA, GDPR and NDAs?
Regulated data (HIPAA PHI, payment card data) is permitted under a written agreement. DPA for GDPR and UK GDPR on request. Happy to sign an NDA: send yours or ask for ours.
Are you SOC 2 compliant?
Not yet. We're working with our compliance partner toward SOC 2 Type I.
Who are your subprocessors?
They're listed in our DPA, which we sign with each customer. Ask for it (mailto:hello@confidential.ai).
What is the minimum term?
Confidential Inference has no minimum: pay per token. Reserved GPU nodes have a 3-month minimum term. Before committing, we offer a short evaluation of up to a week: ask (mailto:hello@confidential.ai).
Where does it run?
United States (South Central) and Central America. EU capacity coming. Licensed deployments run wherever you run them.
Full FAQ: https://confidential.ai/faq
## Company
Inexorable, Inc. d/b/a Confidential AI. Links: [team](https://confidential.ai/team), [careers](https://confidential.ai/careers), [blog](https://confidential.ai/blog).
## Documentation
- Get started, Confidential Inference: https://confidential.ai/docs/inference-api/get-started
- Get started, Confidential VMs: https://confidential.ai/docs/confidential-vms/get-started
- Inference API reference: https://confidential.ai/docs/inference-api/reference
- C8s (Confidential Kubernetes): https://confidential.ai/docs/c8s
- Attested Builds: https://confidential.ai/docs/attested-builds
- Concepts: https://confidential.ai/docs/concepts/confidential-computing-primer