# Confidential Compute Becomes the Trust Layer for Model Audits

Neither labs nor evaluators will hand over weights or held-out test sets in the clear, so expect audits, private benchmarks and regulated-data inference to keep migrating into hardware enclaves — recurring confidential-GPU pilots, safety-institute participation, and enclave attestation turning into a procurement and compliance requirement rather than a research curiosity.

- Conviction: 38 / 100 (forming)
- Horizon: Emerging (watchlist)
- Tracking since: 2026-08-28T00:00:00.000Z
- Last updated: 2026-08-28T06:28:38.436Z
- Canonical: https://polylog.news/ai/trends/confidential-compute-model-evaluation
- Publisher: Polylog
- Affected regions: United States, Europe

## Recent evidence

- [confirms] Google DeepMind Runs Model Evaluations Inside an Encrypted Enclave So Neither Side Sees the Other's Data (2026-08-28): Google DeepMind tested a Gemini Flash Lite model against confidential benchmarks inside Confidential Space using an Nvidia H100 confidential GPU and Intel memory encryption, with the Singapore AI Safety Institute and MLCommons among the partners. The pilot demonstrates a working two-sided-secrecy audit, the missing mechanism for evaluations against benchmarks that would be ruined by disclosure.