Evaluators and safety organisations · Article · Epoch AIChanging how we measure AI chip inference performance9 Nov 2026Read at epoch.ai ↗Read at sourceAll rights reservedSearchOutline · 52 paragraphsThe tradeoff between concurrency and output speed¶7How we calculate the IPI¶15Comparing capacity across output speeds¶27Combining the model comparisons¶32Interpreting the IPI¶36Acknowledgements¶42