Enterprises
AI compute your security team can sign off.
Tell us what you need to run and the rules it has to follow. Providers that meet those rules send you a binding quote, and no one else sees your request.
Your request
- Workload
- Your own model
- GPUs
- 4×B300, four weeks
- Location
- UK, ISO 27001
- Protection
- Confidential computing
- Visibility
- Sealed
- Payment
- Escrow
Custom models
Turnkey AI compute for your custom models.
We expect most future AI demand to come from custom models: frontier models refined with reinforcement learning on a company’s own data. Running one means assembling GPUs, a serving stack and the people to operate them.
Our agents streamline the AI ops. You state what your model needs, and you get bespoke AI compute, ready to run.
You can use your own model.
What our agents take care of
- Serving
- vLLM, SGLang or your own container
- Placement
- Across the GPUs you reserved
- Protection
- Confidential computing set up
- Evidence
- Attested on deployment
- Usage
- Metered against the quote
- Settlement
- Escrow released on completion
Cost
You pay a market price.
You pay once it has run as agreed.
Price per B300 GPU-hour
Covextra provider economics model, October 2026. Exchange price of $10 to $12 per GPU-hour against hyperscaler on-demand list prices.
Sovereignty
Your data stays in the country you choose.
Providers outside that jurisdiction never see the request, and your workload can’t be placed with them.
And you can go further than the country.
A data centre in your country can still answer to a foreign parent. You can rule out operators under US ownership or control, or any provider exposed to the US CLOUD Act, and you can require a recognised scheme such as SecNumCloud in France or C5 in Germany.
Privacy and compliance
Ask for confidential computing, and the provider can’t read your data.
Your auditors get evidence.
You decide what happens when the work ends.
Attestation record (illustrative)
- Model digest
- sha256:9f2c…41ab
- CPU
- Intel TDX
- GPU
- NVIDIA B300, CC on
- Measurement
- c1ee9c…70a5
- Signature
- Hardware-signed, verified
What goes in a request
Every rule you need, in one request.
Mark each one as a must or a preference. A provider that can’t meet a must never sees the request.
Workload
- Inference, fine-tuning, training, batch or embedding
- Your own weights, or a model the provider hosts
- vLLM, SGLang or your own container
Protection
- Intel TDX or AMD SEV-SNP
- NVIDIA confidential GPUs with attestation
- Keys released only to an attested instance
- Your own encryption keys at rest
Sovereignty
- Countries allowed and excluded
- Operators under US ownership or control ruled out
- Exposure to the US CLOUD Act ruled out
- Schemes such as SecNumCloud and C5
Facility
- Uptime Institute tier
- ISO 27001 and other certifications
- Physical security and personnel vetting
- Power and cooling redundancy
Sustainability
- Maximum PUE
- Share of renewable energy
- Carbon intensity
- Energy reporting per job
Commercial
- Per GPU-hour or per million tokens
- Spot, on-demand or reserved
- Budget cap, hidden from providers
- Payment in escrow
Available options grow with each release of the exchange.
Questions from security and engineering teams
- Who sets the rules, security or engineering?
- Both, in the same request. Your security team sets the policy once, and your engineers choose the model and the GPUs inside it.
- Do we have to move everything?
- No. You can start with one workload and keep the rest where it is.
- How do we check the attestation ourselves?
- You can verify the record yourselves, or have a third party do it. The request lets you choose who the verifier is.
- Which GPUs support confidential computing?
- NVIDIA data-centre GPUs from the H100 onwards.
Start with the workload your security team has so far said no to.
We launch in March 2027.