Covextra Exchange
Enterprises post sealed requests, and GPU providers answer with binding quotes, on spot or forward contracts. Payment is held in escrow until the work is done.
The exchange for enterprise AI compute
A new frontier stack for enterprise AI compute.
Enterprises get GPU capacity their security team can sign off. GPU providers get enterprise customers at premium prices. Covextra’s exchange and agents handle everything in between.
Verifiable governance
Workloads run in confidential computing environments, with attested evidence of where and how they ran.
GPU mutualisation
Capacity from independent providers, pooled on one exchange and priced in real time.
Orchestration for AI agents
Every workload placed by policy: jurisdiction, certification, cost, and performance.
Enterprises
AI compute your security team can sign off.
Get turnkey AI compute for your custom models. Choose the country, the certifications and the level of protection. Qualifying providers compete for your workload, and each confidential deployment gives you a hardware-signed record for your auditors.
How it works for enterprises
GPU providers
Enterprise customers for your GPUs, at premium prices.
Hourly rental is becoming a commodity. Our agent integrates with your cloud stack and runs the AI ops, so you can sell compliant AI compute to regulated enterprises.
How it works for GPU providers
How it works
What happens when a bank needs GPUs
A bank wants to run its own model on four B300s in the UK for four weeks, in an ISO 27001 facility.
Sealed request
The bank sends one request. Only UK providers with the right certifications can see it.
Request for compute
- Workload
- The bank’s own model
- GPUs
- 4×B300, four weeks
- Location
- UK, ISO 27001
- Protection
- Confidential computing
- Visibility
- Sealed
- Payment
- Escrow
Binding quotes
They reply with binding quotes, each with a price and the time they need to be ready.
Running and proven
The bank accepts one. Our agents deploy the model to that provider’s GPUs inside confidential computing, where the provider can’t read it.
Confidential computing
The bank’s modelB300B300B300B300Settled on proof
The hardware attests that the right model ran. Escrow pays the provider, and both sides keep the evidence.
AttestationEscrow releasedProvider paid- The next workload starts here, under the same policy.
Back to step 1. The next workload starts here, under the same policy.
The stack
Covextra sits between your workload and the hardware.
A workload arrives with its policy attached. Covextra finds a provider that meets it, places the workload, and returns the evidence to both sides.
Workloads
AI agents, inference, fine-tuning, training and batch jobs
Covextra
Exchange, harness, agents, metering and settlement
Confidential computing
Attested environments on CPUs and NVIDIA GPUs, on the Privasys stack
Infrastructure
GPU providers’ data centres and clouds, such as Canonical OpenStack
Products
What you get with Covextra
Covextra Harness
Where an enterprise sets its policy and runs its agents. The agent, the model and each tool run in attested confidential environments, and the browser reaches them over a channel the operator can’t read.
Covextra agents
The market agent matches requests to capacity and confirms settlement. The management agent, inside the provider’s stack, orchestrates workloads across machines. The compute agent configures each GPU node and its trusted execution environment.
Two ways to buy
Buy the GPUs, or buy the tokens.
Dedicated capacity
GPUs provisioned for you alone, with your choice of model, weights and serving configuration. Use it for your own models, fine-tuning and training.
Shared inference
A standard model the provider already runs, priced per million tokens. You can take a posted price directly, under the same compliance filters, and be live in minutes.
Questions we hear most
- What does Covextra do?
- Covextra runs an exchange for AI compute. Enterprises state what they need to run and the rules it must follow. GPU providers that meet those rules quote for the work, and our agents deploy it and settle the contract.
- What is confidential computing?
- It keeps data and code encrypted in memory while they are being processed, inside a hardware-protected environment. The hardware can then sign an attestation of what is running.
- Which GPUs support confidential computing?
- NVIDIA data-centre GPUs from the H100 onwards. Other GPUs can still join the exchange, for workloads that don’t ask for it.
- When can we start?
- We launch in March 2027, and we’re talking to enterprises and GPU providers now about early access.
We launch in March 2027.