Inference Performance Index: changing the way we measure inference capacity
Epoch AI's Inference Performance Index, which goes beyond the underestimates of FLOP/s measures and scores chips on memory and communication
Sources / Epoch AI
Epoch AI
Research · Official site ↗ · RSS · JSON Feed
What Huawei’s Horinger data center reveals about China’s compute gap
Huawei’s AI data center in Horinger, Inner Mongolia, uses Ascend chips and has an estimated AI computing capacity that is just 7% that of the world’s largest AI data center, SpaceX’s Colossus 2 in Tennessee.
Broadcom TPU compute shipments doubled from Q2 to Q3 2026, reaching 2 million H100-equivalents
Broadcom’s shipments of TPU by unit volume increased by about 65% quarter over quarter to roughly 800,000 chips, while compute output doubled, driven by a shift to more powerful chips.
Updating the Epoch Capabilities Index methodology to use a Bayesian model
A Bayesian model gives the ECI better-justified uncertainty estimates and lets us compare models directly.
LLM use is acknowledged almost as often as specialized AI tools in biology preprints
17% of biology preprints acknowledge LLM use, nearly matching specialized tools like AlphaFold. Epoch AI analyzes trends in AI acknowledgment in research.
AI can turn security fixes into working exploits before users receive an update
GLM 5.3 turned published-but-unshipped Firefox security fixes into working one-click exploits for six releases, at roughly $1,000 each.
New global AI computing capacity is doubling every 9 months
The computing power of AI chips sold each quarter has grown about 2.5× per year since 2024, down from a surge in 2023.
Introducing the AI Chip Imports explorer
Our explorer uses customs records to estimate national shares. Since 2022, 72% of AI compute has ended up in the US and about 10% in China, including chips China makes itself.
Open-weight models trail closed models in cyber capabilities by 6–7 months
Z.ai’s GLM-5.3 is the top open-weight model on Epoch’s Cyber ECI, scoring where the closed-weight frontier stood 6–7 months earlier.
AI systems are improving rapidly at Slay the Spire and its sequel
Months ago, frontier models could barely beat Slay the Spire. Now they win 90%+ at the easiest setting and up to 20% at the hardest. Results from Epoch AI’s new SpireClimb benchmark.
Can AI automate Epoch?
Frontier models can handle well-defined tasks reliably, but fail in the more open-ended aspects that prevent full work automation. Explore Epoch AI's Automation Reports: realistic, open-ended tasks taken directly from our work, manually graded by our own standards.
US adults are no more likely to face cyber incidents than when Claude Fable 5 launched
June and September polls from Epoch AI/Ipsos show no detectable increase in the share of US adults experiencing cyber incidents
EBR-bench update
Updates on EBR-bench, Epoch AI’s benchmark that tests models’ ability to learn from experience. GPT-6 Astra hit a perfect score on over half its EBR-bench attempts by exploiting one overpowered card, which we now ban. Plus multi-agent scaffold results.
Can AI automate AI R&D yet?
Early evidence from InnovationEval: No
China is more exposed to semiconductor supply-chain disruptions than the US
China is 2.7× as exposed to semiconductor supply shocks as the US, Epoch AI estimates. Epoch traces chips through input-output tables to model the impact of a Taiwan shock and decoupling.
How do Chinese AI companies make money?
China's six leading AI firms earn about 10% of OpenAI and Anthropic's combined AI revenue. Epoch AI examines their five revenue streams: consumer apps, model access, enterprise and government solutions, licensing fees, and indirect monetization through cloud and advertising.
Coding-agent spending by OpenAI researchers is doubling roughly every month
OpenAI researchers’ coding-agent usage, valued at API prices, has recently been doubling roughly every month, according to breakpoint fits.
How many AI agents could run on the AI chips shipped through 2027?
Memory shipped through 2027 could run 33–171 million concurrent frontier-model agents, or billions using efficient open models. Epoch AI estimates inference capacity from HBM supply, serving benchmarks, and agent-hour costs.
Introducing the ChatGPT usage explorer
Median monthly messages per active user rose from 14 in 2023 to 36 in 2025, and the top 10% of users sent 63% of all prompts. Epoch AI's ChatGPT usage explorer draws on 8.3 million messages from 5,000 US YouGov panelists.
AI has improved significantly at reasoning about IKEA furniture assembly
AI model scores on Epoch AI's IKEA furniture assembly benchmark jumped from 28% to 80% in 10 months, with open-weight models trailing closed-weight models by about 7 months. Epoch AI's test of visual and spatial reasoning across 60 photos from three furniture builds.
The plunging price of thought
Epoch AI measures how fast the cost of a given level of AI performance is falling across five benchmarks covering math, science and games of skill: about 47% per quarter, or 13x per year, since 2023, faster than electricity, compute, batteries or DNA sequencing ever fell.
In August, 25% of math preprints acknowledged AI use, up from 4% in April
Acknowledgments of AI use in arXiv math preprints rose from 4% in April 2026 to 25% in August, with 6% crediting AI with a substantial research contribution.
Trade data is consistent with more than $3 billion of chips smuggled into China via Malaysia
Epoch AI's analysis of trade data finds a pattern consistent with $3B worth of chips smuggled to China via Malaysia. China recorded $3.8B of server imports from Malaysia in 2024–25; Malaysia recorded only $0.6B of exports.
GPT-6 Astra leads on math benchmarks, but not on software engineering
OpenAI's GPT-6 Astra tops the Epoch Capabilities Index (ECI) with a score of 166, ahead of Claude Fable 5.1 at 164 and GPT-5.6 Sol at 162. Its Math-ECI of 170 sets a new record, but on software engineering benchmarks its SWE-ECI of 164 still lags behind Fable 5.1's 167.
In the daily digest
9 November 2026,6 November 2026,4 November 2026,30 October 2026,29 October 2026