Compute access of AI Safety Institute

Question

The UK AI Security Institute publicly states it has priority access to over 1.5 billion of compute. What dedicated compute capacity does AISI have for frontier model evaluations? Has AISI assessed whether Australian sovereign compute is sufficient to run pre-deployment evaluations on models of the scale of Mythos or GPT-5.5? Has AISI published a Work Programme for 2026-27? If not, when will it be published? Will it identify the frontier models AISI intends to evaluate?

Answer

See attached response.

Documents

• Question with answer (PDF) (opens in a new tab)
• BI-015_Shoebridge_Compute access of AI Safety Institute_DISR.pdf (opens in a new tab)