Confidential Computing on GPUs: Can You Trust a Shared AI Cluster With Sensitive Data?
A bank, a hospital or a ministry that wants to use AI runs into the same wall. The most capable hardware is shared: in a cloud, in a colocation hall, or on a cluster used by several departments. The data is not something they are allowed to share.
For years the honest answer was "then you cannot use it". A newer technology changes that answer to "you can, if you can prove the machine is not looking". It is called confidential computing, and it now reaches the GPU.
The gap it closes
| Data isโฆ | Protected by | How long available |
|---|---|---|
| ๐พ At rest | Disk encryption | Decades |
| ๐ In transit | Network encryption | Decades |
| โ๏ธ In use | Confidential computing | Recent On GPUs since the H100 |
The third row was the weak point. To compute on data, a machine has to decrypt it, and at that moment anyone who controls the machine could, in principle, read it. That includes the administrator, the cloud provider and anyone who has compromised either.
How it works, in plain terms
Confidential computing protects data in use by running the computation inside a sealed area of the hardware, called a trusted execution environment, that the rest of the system cannot see into. NVIDIA describes the goal as protecting a user's code and data from even the owner of the computer or cloud service.
Three pieces make it work on a GPU:
- A sealed virtual machine on the processor. NVIDIA's approach relies on a confidential virtual machine provided by the server's processor, using AMD's SEV-SNP or Intel's TDX technology.
- An encrypted link to the GPU. Data moving between the processor and the GPU is encrypted, using AES-GCM 256 on the H100.
- Proof, called attestation. Each GPU carries a unique key fixed at manufacture. Before any sensitive work starts, the GPU produces a cryptographically signed report of its state, which the user checks. Only if the report is valid does the work proceed.
The third piece is what makes this more than a promise. You are not asked to trust the operator. You are given evidence to verify.
Sources: NVIDIA Trusted Computing Solutions, ACM Queue: Creating the first confidential GPUs, Edgeless Systems on the H100
What it protects against, and what it does not
| Threat | Covered? |
|---|---|
| ๐ค A curious or malicious administrator at the provider | Yes This is the main purpose |
| ๐ข Another tenant on the same hardware | Yes |
| ๐ง Tampered GPU firmware | Yes Only NVIDIA-signed firmware runs in this mode |
| ๐ A flaw in your own application | No A leaky app leaks inside the sealed area too |
| ๐ Stolen user passwords or keys | No |
| ๐ฌ A model tricked into revealing its data | No That is an AI safety problem, not a hardware one |
| ๐ Legal and regulatory duties | Helps, does not replace |
The "covered" column is our own summary of the design described in the sources above. It is not a security assessment of any particular deployment.
The limits vendors mention quietly
It works on specific combinations
This is not a switch on every server. NVIDIA's release notes list supported pairings of GPU, firmware, driver and processor technology. One example: the mode that protects an eight-GPU system was listed as available only with Hopper GPUs and Intel processors with TDX, in one particular virtualisation setup. If you need this capability, it has to be designed in from the first quote, not added later.
Source: NVIDIA confidential computing release notes
You move your trust; you do not remove it
The system stops you having to trust the operator. In exchange you trust the chip makers: the keys, the firmware signing and the attestation service are theirs. For most organisations that is a much better position. It is not the same as trusting no one.
It costs something
Encrypting everything that moves between processor and GPU takes work, and guides to the technology discuss the performance trade-off openly. Ask for a measured figure on your own workload.
Someone has to check the proof
Attestation only protects you if the report is verified every time, automatically, before work begins. A confidential system whose attestation nobody checks is an ordinary system with extra steps.
Who needs it
| Your situation | Our read |
|---|---|
| ๐ฆ Regulated data on rented or shared GPUs | Strong case |
| ๐ค Several organisations pooling data to train one model | Strong case No party has to trust another's staff |
| ๐ง Protecting a valuable model you run on someone else's hardware | Good case |
| ๐ข Departments sharing a cluster you own and operate | Depends Ordinary isolation may be enough |
| ๐ Your own data, your own servers, your own room | Rarely needed You already are the operator |
The slightly controversial view
Confidential computing is sometimes presented as the reason regulated organisations can now use any cloud for anything. That overstates it. The technology is real and well engineered, but it covers one threat, the operator, and leaves the others where they were. An organisation that adopts it and relaxes everything else has made itself less safe, not more.
The fairer framing is that it removes one of the oldest objections to shared infrastructure. For a Singapore organisation weighing a local colocation hall against building its own room, that changes the sums. It does not, by itself, satisfy the Personal Data Protection Act or any sector regulator. Those obligations are about governance as well as technology, and are for your compliance team to judge.
Five questions for a provider
- Which exact GPU, processor and software combination supports confidential mode here?
- Does it cover multi-GPU jobs, or single GPUs only?
- Who verifies attestation, and can we run our own verifier?
- What is the measured performance cost on a workload like ours?
- What happens to our data if attestation fails mid-job?
Related on SourceIT
SourceIT Approved
Researched and verified by the SourceIT team for Singapore buyers.




