Shared GPU infrastructure has a specific set of risks. Here is what we do about each of them, and what we do not claim.
Each instance receives a dedicated physical GPU passed through with VFIO. Cards are never time-sliced, MIG-partitioned across accounts, or shared. When your instance ends, the card is reset at the firmware level before it re-enters the pool.
GPU memory is explicitly zeroed during the reset cycle, not merely marked free. This closes the residual-data path that affects naive multi-tenant GPU platforms.
Root volumes are encrypted with a per-instance key held only in the host's memory. Destroying an instance discards the key, which renders the data unrecoverable immediately rather than pending an overwrite pass.
The API and console are TLS 1.3 only, with HSTS. The private compute fabric is physically isolated and does not traverse shared infrastructure.
SSH keys only — password authentication is disabled in every image. API tokens are scoped per project and can be rotated or revoked without downtime.
New instances expose port 22 and nothing else. Opening additional ports is a deliberate act, which prevents the most common accidental exposure in this industry.
We would rather state this plainly than imply more than is true:
| Framework | Status |
|---|---|
| SOC 2 Type II | Observation period in progress |
| ISO 27001 | Gap assessment complete |
| GDPR | DPA available on request |
| HIPAA | Not supported |
| PCI DSS | Not supported |
| FedRAMP | Not supported |
Do not place regulated health or cardholder data on this platform. If your workload requires HIPAA or PCI controls, we are not the right provider today and will tell you so rather than take the business.
We can see instance metadata: which GPU, which region, how long it ran, how much network it moved. We do not inspect the contents of your instances, your volumes or your traffic. Support staff cannot log into a running instance; if you want help debugging inside the box, you grant access explicitly and it is logged.
Email support@gpuai.shop with the subject line "security". We acknowledge within one business day and aim to have a fix or mitigation timeline within five. We do not currently run a paid bounty, and we will not take legal action against good-faith research that stays within your own account.