Source: raw/newsletter-epoch-ai-fa4c3ebf04.md (The Epoch Brief, November 11, 2026, Cathy Bogaart; https://epochai.substack.com/p/the-epoch-brief-november-11-2026), checked against six Epoch pages fetched 2026-11-11: ai-research/epoch-compute-capacity-doubling-9-months-2026-11-11.md, ai-research/epoch-broadcom-tpu-shipments-doubling-2026-11-11.md, ai-research/epoch-ai-chip-imports-explorer-2026-11-11.md, ai-research/epoch-huawei-horinger-china-compute-gap-2026-11-11.md, ai-research/epoch-inference-performance-index-2026-11-11.md, ai-research/epoch-biorxiv-llms-vs-ai-tools-2026-11-11.md.

Creator: Epoch AI · URL: https://epochai.substack.com/p/the-epoch-brief-november-11-2026 · Platform: Epoch AI Substack and epoch.ai

Epoch’s November 11 brief is mostly about the supply of AI compute: how fast it is growing, which countries receive it, and how to count it for inference. The answers are that new capacity doubles about every nine months, the US has received 72% of it since 2022, and raw FLOP/s understates how much newer chips can serve. The brief also records a new Epoch Capabilities Index (ECI) record and a jump in acknowledged LLM use in biology. Its patch-gap item repeats the October report already covered in Open-Weight Models and the Patch Gap.

Key Takeaways

  • New AI compute doubles about every 9 months. Quarterly additions grew from 0.8M H100-equivalents (H100e) in Q1 2024 to 6.3M in Q2 2026, a fitted trend of 2.5× a year (90% CI 2.3× to 2.7×), down from about 6× in 2023.
  • Google’s TPUs are a growing share. Broadcom’s TPU compute shipments roughly doubled in one quarter (fiscal Q2 to Q3 2026) to nearly 2M H100e, on about 800,000 chips; the shipments went to both Google and Anthropic. Google TPUs were about a quarter of all AI compute shipped worldwide from April to June.
  • 72% of the world’s AI compute since 2022 went to the US. Of 27.4M H100e delivered through June 2026, the US got 19.8M; China 10% (2.6M, including domestic chips), the EU 5%, Malaysia 2%. The US share of each year’s new compute rose from about 60-63% (2022-2024) to 72% (2025) and 78% (first half of 2026).
  • China has 18% of the chips but 10% of the compute, because much of its hardware is export-compliant or domestic and slower.
  • One of China’s largest AI data centers holds about 7% of Colossus 2’s compute. Huawei’s Horinger campus has up to 286 MW of IT power but about 81,000 H100e operational (90% CI 36,000-141,000).
  • Count inference capacity in requests served, not FLOP/s. Epoch estimates an Nvidia GB300 serves about 25× the agentic inference load of an H100 with only 2.5× the 8-bit FLOP/s. Its new Inference Performance Index (IPI) scores chips in GB300-equivalents.
  • Claude Fable 5.2 set an ECI record of 174 (previous high GPT-6 Astra, 169), above the post-2024 reasoning-model trend. Model detail is in Claude Fable 5.2 and Mythos 5.2.
  • 17% of biology preprints now acknowledge LLM use, close to the 19% that acknowledge specialised tools such as AlphaFold.

Compute supply: 2.5× a year

From ai-research/epoch-compute-capacity-doubling-9-months-2026-11-11.md (Isabel Juniewicz and Josh You):

  • Method. Quarterly sales estimates for six chip designers (Nvidia, Google, AMD, Amazon, Huawei, Cambricon), converted to H100e, with a log-linear fit over the ten quarters Q1 2024 to Q2 2026. Doubling time 9.1 months (90% CI 8.4-10.0).
  • Quarterly additions by designer, Q2 2026: Nvidia 3.95M H100e, Google 1.61M, Amazon 0.29M, AMD 0.26M, Huawei 0.20M, Cambricon 0.01M; total 6.33M.
  • Slowing, maybe. Some evidence points to further deceleration, but it is not statistically significant and may be a lull between chip generations: Nvidia had not recognised any Rubin sales as of its fiscal quarter ending in July.
  • Cumulative stock grew 2.6× between Q2 2025 and Q2 2026, down from 3.5× the year before, converging on the 2.5× additions trend.

The Broadcom TPU jump (ai-research/epoch-broadcom-tpu-shipments-doubling-2026-11-11.md, Venkat Somala and Josh You):

  • Unit volume rose about 65% quarter on quarter to roughly 800,000 chips; the rest of the doubling came from the mix shifting to TPU v7, about 2.5× the performance of v6e. TPU v8i entered production in fiscal Q3 2026 at about 2.2× v7.
  • In the three previous quarters, shipments grew 2% to 22% a quarter.
  • Broadcom started shipping TPUs for Anthropic in Q3 2026. Epoch’s median TPU v7 price in Q3 is about **25,000 per TPU in finished racks.

Where the chips land: the AI Chip Imports explorer

From ai-research/epoch-ai-chip-imports-explorer-2026-11-11.md (Lucio Melito and Isabel Juniewicz):

  • Method. Customs records from 16 destinations, netting out transit hubs such as Singapore and Hong Kong, used to split Epoch’s global chip-sales estimates by country. China’s domestic production is added from chip-sales data because it never crosses a border.
  • Totals, 2022 to June 2026: US 19.8M H100e (72%), China 2.6M (10%), EU 1.4M (5%), Malaysia 0.7M (2%); no other country above 2%.
  • China’s chips versus compute. 18% of chips, 10% of compute. Export-compliant chips (such as Nvidia’s H20) and Huawei and Cambricon chips make up about 40% of China’s total. Imports alone: 12% of chips, 6% of compute.
  • Uncertainty. China’s total ranges from 1.9M to 4.2M H100e under other assumptions. Smuggled chips count only if they appear in China’s own customs records. Countries without detailed customs data (the UAE, Vietnam) are estimated from what others report sending.

China’s compute gap at one site: Huawei Horinger

From ai-research/epoch-huawei-horinger-china-compute-gap-2026-11-11.md (Ben Cottier):

  • Site. Horinger, Inner Mongolia; up to 286 MW of IT power across four AI buildings (estimated from a cooling model and satellite imagery). Three of five operational buildings are estimated to house Ascend 910C chips.
  • Compute. About 105,000 Ascend 910C and 8,000 910B chips operational, about 81,000 H100e. That is about 11% of all 910C chips sold by August, allowing a four-month lag between sale and deployment.
  • Against Colossus 2. SpaceX’s Colossus 2 in Tennessee, the world’s largest AI data center, has about 14× Horinger’s compute (the brief’s “roughly 7%”).
  • Efficiency. US AI data centers support about 2.5× more AI compute per unit of power, and the largest have reached 1,000 MW.
  • Next phase. A new building is expected to be operational around February 2028; with it, Horinger could reach about 141,000 H100e (90% CI 65,000-227,000), still about the same share of Colossus 2, which Epoch expects to pass 1.8M H100e by 2028.

Added 2026-11-13: details from Epoch’s newsletter version

Source: raw/newsletter-epoch-ai-da92d82c74.md (Epoch AI Substack, Ben Cottier, November 12, 2026; https://epochai.substack.com/p/what-huaweis-horinger-data-center), a condensed version of the report above. It adds no new estimate; these details from it are useful when citing the numbers.

  • Chip conversion. The Ascend 910C performs at about 0.76 H100e; Nvidia’s B200 and B300 at about 2.5 H100e, so one B300 is worth roughly three 910Cs in this measure. The full report notes the 910C lacks native FP8, so 0.76 arguably flatters it for training.
  • What the site can train. 10% of Horinger’s compute (about 8,100 H100e) is enough to train a DeepSeek-V4-Pro-scale model in about a month; all of it combined in one run (optimistic) would take about a year for a GPT-6 Astra-scale model (Epoch’s median estimate of 1e27 FLOP), both at 20% utilisation.
  • Building status. Of the four dense buildings, three are operational and one was complete but not running as of September 2026, expected online by December. The North zone has far fewer cooling towers and is unlikely to hold AI chips.
  • Stock and peers. About 907,000 Ascend 910C and 600,000 910B chips had been sold by the end of March 2026. Only VNET Bayin Ulanqab and Huawei Wuhu are likely to match or slightly exceed Horinger in China; Wuhu is projected at 151,000 H100e by December 2027.
  • Epoch’s conclusion. China’s compute constraints are “unlikely to ease in the next two years”, so Chinese AI capability and deployment capacity are expected to keep lagging the US. For the policy side of the same gap see the Megaspeed article, where David Sacks describes Huawei’s CloudMatrix 384 trading energy for rack-level performance.

Measuring inference: the Inference Performance Index

From ai-research/epoch-inference-performance-index-2026-11-11.md (Jason Li):

  • Why FLOP/s misleads. For inference, memory capacity, memory bandwidth and chip-to-chip links often matter more than arithmetic. Comparing FLOP/s alone would underestimate the GB300 against the H100 by about 10×.
  • What IPI measures. How many requests a chip can serve at once across a range of required output speeds, on agentic workloads, averaged over four representative models. Scores are in GB300-equivalents (GB300 = 1).
  • Examples. Vera Rubin (VR200) scores 2.18; AMD MI355X about 0.55. A GB300 NVL72 rack has about 1.9× the modelled inference capacity per GPU of a B300 HGX, because it links 72 GPUs instead of eight.
  • The trade-off it encodes. Raising concurrency from two to eight requests per chip lowers output speed from about 210 to 80 tokens per second per request in Epoch’s example; throughput is concurrency times speed.
  • The brief says newer chips reach “up to 72×” the H100’s inference capacity; that exact figure was not found on the fetched page.

Capabilities: a new ECI record, scored a new way

From the brief only (the ECI pages were not fetched):

  • Fable 5.2: 174 on the Epoch Capabilities Index, released November 9; previous high GPT-6 Astra at 169. Epoch says the score sits above the trend for reasoning models since late 2024, and cautions that it is “just one data point” in the discussion of AI speeding up AI R&D (see METR on AI R&D Acceleration).
  • Bayesian ECI. Emilia Álvarez and Alexander Barry moved ECI from a frequentist to a Bayesian method, which gives ranges for each score and head-to-head probabilities: Astra has a 72% chance of being more capable than Claude Opus 5.5 and 99% against Claude Fable 5.1. Frontier scores moved about one point on average.

AI in science: biology preprints

From ai-research/epoch-biorxiv-llms-vs-ai-tools-2026-11-11.md (Tara Abrishami):

  • 17% of bioRxiv preprints acknowledge LLM use, against 19% for specialised AI tools (AlphaFold is about half of those).
  • LLM use acknowledged specifically for research or analysis rose from about 1% at the start of 2026 to about 4% in September.
  • Method: GPT-5.6 Sol checked every bioRxiv preprint from January 2020 to September 2026; a GPT-6 Luna classifier sorted LLM uses into research, code, writing and literature review. Preprints may not represent peer-reviewed work, and acknowledgements undercount use.
  • In math, the brief says AI-use acknowledgements on arXiv rose from about 4% in April to 48% by October; the fetched page gives 25% for August, so the October figure is a later reading.

Try It

  1. Plan compute prices on 2.5× a year of new supply, not 6×. If your costs depend on falling inference prices, the supply side is still growing fast but slower than in 2023; combine it with the post-election permitting picture in the data-center slowdown before assuming price cuts.
  2. Ask providers about the hardware generation, not just the GPU count. Epoch’s IPI says a GB300 serves about 25 H100s’ worth of agentic requests. A quote for “N GPUs” means little without the chip and the rack configuration.
  3. Use the AI Chip Imports explorer (https://epoch.ai/data/ai-chip-imports) when a client or report claims a country is “catching up” on compute; check compute share, not chip count.
  4. Cite ECI with its range. With the Bayesian method, quote the probability that one model beats another rather than a one-point gap.

Open Questions

  • “Up to 72×” inference capacity comes from the brief; the fetched IPI report gives 25× for GB300 and 2.18 GB300e for VR200 but does not state 72× in the parts read.
  • The ECI record and Bayesian methodology pages (epoch.ai/models/claude-fable-5-2, epoch.ai/publications/eci-bayesian-methodology) were not fetched.
  • The 48% math figure for October is from the brief only.