The Definitive Guide toAI Data Centers
Ask the GuideAboutAccount
GuideGlossaryTest-time compute

Test-time compute

Spending extra inference compute (e.g. chain-of-thought reasoning) to improve answers, shifting cost from training to serving.

Current numbers

5-100xmore tokens/test-time compute per query for reasoning models vs one-shot inference — workload-dependent (NVIDIA cites up to ~100x compute)as of 2025 · register ↗

← All terms