Security Architecture
Hardware and software protection at each step, and where it stops: GPU VMs and models run in Intel TDX, while our relay sees inference requests in clear.
Architecture
Four layers of protection, from silicon to attestation.
Intel TDX Enclaves
Hardware-level isolation: the hypervisor and host OS cannot access enclave memory.
NVIDIA Confidential Computing
In confidential computing mode, data between CPU and GPU travels encrypted and the GPU walls off its memory from the host.
Publicly Verifiable Attestation
Cryptographic proof that the enclave is genuine: you generate it inside a GPU VM, and GET /v1/attestation returns it for the instances serving a model.
What our relay sees
Requests to the inference API, the chat and the agents go through our servers, which see them in clear; the Telegram agent stores your messages to answer them. Security and billing logs are kept per our security page.
Data Protection
- Stored content: the Telegram agent stores your messages in our database; usage records keep the model, token counts and cost of each request.
- Where requests go: the model runs inside an Intel TDX enclave at our inference provider; requests go through our API. Usage records keep the model, token counts and cost.
- No training: your data is never used to train or fine-tune models.
- Isolation, each GPU VM is its own Intel TDX trust domain; inference requests go to the instances serving the model, with no per-user enclave.
- TLS 1.3 encryption for all data in transit.
- No resale: your data is not sold; it reaches only the sub-processors that run the service, listed in our register.
Compliance
Architected to support the requirements regulated industries have to meet.
GDPR Art. 28
Data processing agreement available. French company under EU law.
Intel TDX
Hardware attestation ensures enclave integrity.
What our relay sees
Requests to the inference API, the chat and the agents go through our servers, which see them in clear; the Telegram agent stores your messages to answer them. Security and billing logs retained per our security page.
Data Sovereignty
Customer database in Frankfurt (EU). Compute, payment and monitoring providers include US entities under standard contractual clauses, see the sub-processor register.
HIPAA-aligned, no BAA yet
BAA planned after SOC 2 Type II, not offered yet.
DPA
Standard Data Processing Agreement for enterprise customers.
On the audit roadmap (not yet held)
SOC 2 Type I
NOT YET HELD, auditor selection in progress. Independent attestation of security, availability and confidentiality controls (point-in-time).
SOC 2 Type II
NOT YET HELD, 12-month observation window after Type I. Operating-effectiveness attestation over a continuous monitoring period.
ISO/IEC 27001
NOT YET HELD, gap assessment scheduled. Information security management system certification.
We do not claim certifications we have not earned. The dated targets above are commitments, not promises, they will move if scope or customer demand changes. The Trust Center is the authoritative source for live audit status.
Infrastructure
- Distributed nodes across our GPU providers' data centers for redundancy; locations vary and are not limited to the EU.
- On-demand scaling with automatic resource allocation.
- Encrypted database for account data (AES-256).
- Rate limiting and abuse prevention on all endpoints.
- TEE models only , the inference API serves only models running inside Intel TDX, and our relay sees requests in clear. Standard GPUs, sandboxes and bare metal run without an enclave.
Security Contact
Report vulnerabilities to security@voltagegpu.com. We respond within 24 hours.
security@voltagegpu.com