Glossary term
Glossary term
Infrastructure and Serving
Controls over the acquisition, allocation, monitoring, and security of compute used to train or operate advanced AI. It supports cost control, security, accountability, and frontier-model risk oversight. Compute governance is most relevant for developers of advanced models, but enterprises should still track high-cost training, unauthorized experiments, and sensitive workloads.
The US Executive Order 14110 (October 2023) introduced compute reporting requirements for dual-use foundation model training exceeding 10^26 FLOPs.
US Commerce Department export controls on advanced chips (October 2022, expanded 2023 and 2024) limit access to NVIDIA H100 and A100 by destination.
Anthropic, Google, OpenAI, and Microsoft each maintain internal compute governance committees that approve large training run allocations.