Epoch’s recent research highlights interconnected trends shaping AI capabilities and risks: open-weight models can autonomously generate harmful code, global AI compute is growing fast and concentrating geographically, and newer chips can deliver far greater real-world inference throughput than FLOP/s alone indicates.
Open-weight models and exploiting the patch gap
In controlled offline sandbox tests, the open-weight model GLM 5.3 autonomously produced working exploits against six recent Firefox versions. Firefox has over 200 million active users worldwide; Epoch reports the model created usable exploits at roughly $1,000 per successful attack.
The technique leverages a well-known software vulnerability: the time window between publication of security patches and when those fixes reach users — the so-called “patch gap.” Previously, turning patch information into functioning exploits required substantial expertise, time, or privileged access to frontier models. Epoch’s experiments show a widely available open-weight model can now perform serious patch-gap attacks without human intervention.
Epoch did not replicate the result against Chrome, but notes that improved prompting or a stronger model could change that outcome. Patch gaps are common across software ecosystems — from other browsers to phone OSes and router firmware — and Epoch expects patch-gap attacks to increase as more capable open-weight models are released.
Claude Fable 5.2 and the Epoch Capabilities Index (ECI)
Anthropic released Claude Fable 5.2 on November 9; it set a new Epoch Capabilities Index record with a score of 174 (the previous high was GPT-6 Astra at 169). The ECI aggregates results from many benchmarks into a single general-capability scale that allows cross-model comparisons.
Epoch notes that model capabilities have risen across generations, and since AI labs added reasoning capabilities to models in late 2024, the rate of improvement has accelerated. Claude Fable 5.2’s score is above what the pre-Astra trend would have predicted. Epoch cautions that while there is debate about AI’s contributions to accelerating AI R&D, Fable 5.2 represents a single data point in that debate.
Upgrading how ECI is calculated
Researchers Emilia Álvarez and Alexander Barry changed ECI’s calculation from a frequentist approach to a Bayesian one. The new method yields better-justified score ranges and enables probabilistic statements about how likely one model is to outperform another. For example, Epoch can now say Astra has a 72% probability of being more capable than Claude Opus 5.5, and a 99% probability of being more capable than Claude Fable 5.1.
The methodological update did not materially change frontier model scores: for models stronger than OpenAI’s GPT-4, scores shifted on average by about 1 ECI point.
Growth in global AI compute and chip shipments
Epoch estimates that new global AI computing capacity is doubling roughly every nine months. Quarterly additions of compute have grown at about 2.5× per year since 2024, slower than the roughly 6× annual surge observed in 2023. Some evidence suggests further deceleration during this period, possibly reflecting a lull between chip generations, but that signal is not statistically significant.
A notable contributor to recent compute growth is Broadcom, which produces Google’s custom AI chips (TPUs). Broadcom doubled its TPU compute shipments between its fiscal Q2 and Q3 of 2026, reaching nearly 2 million H100-equivalents, according to Venkat Somala and Josh You. In the prior three quarters, shipments rose 2–22% quarter-to-quarter.
Much of Broadcom’s increase came from shipping 65% more chips (about 800,000 additional units) and from shifting toward newer, higher-performance chips, mainly the TPU v7. Shipments went to both Google and Anthropic. Epoch’s chip sales data indicate Google’s TPUs accounted for about a quarter of all AI compute shipped worldwide from April to June 2026.
Where chips end up: AI Chip Imports explorer
Epoch’s AI Chip Imports explorer uses trade data to estimate how much AI compute has arrived in 16 countries by quarter since 2022. Lucio Melito and Isabel Juniewicz report that 72% of the roughly 27.4 million H100-equivalent units delivered through June 2026 landed in the United States. China received about 10%, the EU 5%, and Malaysia 2%.
It is important to distinguish between shares of compute and shares of chips. China’s share of physical chips is larger (about 18%), suggesting many of those chips are lower-performance. Chips that comply with U.S. export controls and domestically produced chips make up roughly 40% of China’s total supply.
Case study: Huawei’s Horinger AI data center
Senior researcher Ben Cottier’s case study of Huawei’s AI campus in Horinger, Inner Mongolia, underscores China’s chip-performance constraints. The campus has up to 286 megawatts of IT power capacity across four AI buildings, making it likely one of China’s largest AI data centers. Epoch estimates it holds roughly 7% of the computing capacity of SpaceX’s Colossus 2 in Tennessee, currently the world’s largest AI data center.
China has ample energy for data centers, but U.S. facilities support about 2.5× more AI compute per unit of power, and the largest U.S. sites have reached 1,000 MW. Horinger is expected to add over 70% more compute by early 2028, reaching about 141,000 H100-equivalents; that would still be roughly the same share relative to Colossus 2, which Epoch expects to surpass 1.8 million H100-equivalents.
This case study draws on Epoch’s broader AI data center research, which uses high-resolution satellite imagery, permits, and construction knowledge to estimate data center power, compute capacity, and cost.
How scientists use AI
Researcher Tara Abrishami finds that acknowledgements of LLM use in biology preprints have risen to about 17%, nearly matching the 19% that acknowledge specialized AI tools such as AlphaFold. In September 2026 roughly 4% of biology preprints explicitly acknowledged using LLMs for research or analysis, up from about 1% at the start of the year.
This complements Epoch’s analysis of math arXiv preprints, where acknowledgements of AI use rose from roughly 4% in April to 48% in October.
Inference Performance Index (IPI): a different measure of inference capacity
Jason Li developed the Inference Performance Index (IPI) to estimate how many simultaneous requests a chip can serve across a range of latencies, capturing differences that FLOP/s misses (inter-chip communication, memory behavior, etc.).
Using the IPI, Epoch estimates the Nvidia GB300 provides about 25× the inference capacity per chip of an H100 on agentic workloads, despite only having about 2.5× the 8-bit FLOP/s. Against the GB300 benchmark, some newer chips show up to 72× higher inference capacity than the Nvidia H100.
IPI offers a common unit for estimating an individual data center’s total inference capacity or the world’s aggregate inference capacity.
Careers and closing
Epoch is hiring across multiple roles: Researcher / Senior Researcher, Designer / Senior Designer, Design Lead / Head of Design, Product Manager (Website), Talent Acquisition Coordinator, and Software Engineer (Cybersecurity Benchmarks). Applications are rolling and candidates are encouraged to apply soon.
In sum, Epoch’s work documents accelerating and concentrated growth in compute, evolving measurement methods (Bayesian ECI, IPI), record-setting model releases (Claude Fable 5.2), and emergent risks such as open-weight models weaponizing patch gaps.
