Skip to main content

Workloads you can run

In one breath: inference is only part of the story. The same trust domain you defined wraps training and federated learning too, so the whole model lifecycle stays inside your rules.

Private inference​

● Live

The workhorse. The compiler matches the technique to how sensitive the data is: masked, secret-shared, or fully homomorphic. The strongest option never decrypts your inputs at all, anywhere. One key, four modalities: vision, voice, text and time-series.

Confidential-hardware inference​

Models served inside attested enclaves, with NVIDIA confidential computing on the GPUs doing the serving. Near-native speed and chip-signed evidence, for teams whose auditors already accept confidential computing.

Federated learning​

◐ Active R&D

Train one model across hospitals, branches or borders without the data ever pooling. Each site keeps its records where they are; only the learning moves. This is how a network of clinics gets a model none of them could train alone, without any of them handing over a single patient record.

Private training & fine-tuning​

◐ Active R&D

Bring the model to your corpus instead of the other way around. Fine-tune on proprietary data inside the perimeter you defined, receipts included.

You don't pick the technique. You pick the outcome.​

It bears repeating: you never have to become an expert in MPC versus FHE versus federated averaging. You tell the compiler what the data is, where it may live, and what your auditors accept. It picks what holds up, and gives your engineers the same OpenAI-compatible endpoint either way.

Where this stands

Private inference on Standard and Private tiers serves production traffic. Confidential-hardware serving is assembled per engagement on attested infrastructure. Federated learning and in-perimeter training are active R&D, run with design partners. The Status board is the source of truth; your agreement defines what applies to you.