Data stories
AI in numbers
Who builds AI models, who opens the weights, and what companies actually use. Each chart answers, in one line, what it means for your industry — with the source, licence and data date right under it.
Story 1 · Countries
The US still makes the most notable models; China's share has grown
Of the 107 notable AI models Epoch lists for 2025, 64 came from US organisations and 35 from Chinese ones.
What it means for your industry
Frontier models mostly come from US and Chinese organisations. A shipping or manufacturing company choosing a model is in practice choosing between two ecosystems whose regulation differs (export controls, data residency). That is a procurement question before it is a technical one.
- Epoch's list is a curation, not a census. The country is the developer organisation's, not where the people work.
- The AI Index 2026 (April 2026) counted 59 US, 35 China and 8 South Korea out of 102 for 2025 (PDF p.15, p.21). Epoch has added models since the report's cut-off, so this page shows Epoch's live figure with its date. Stanford HAI, AI Index 2026 (CC BY-ND 4.0 — quoted, not redrawn)
Story 2 · Open weights
The open frontier is Chinese
China releases most of its notable models with open weights, and the US increasingly does not. On the download side, models built on Chinese bases lead too.
In 2025, 28 of China's 35 notable models had open weights; in the US it was 11 of 64.
- US — open weights
- China — open weights
- Not open weights
- * 2026 = partial year (hatched)
Share of US notable models with open weights: 43% in 2023 → 17.2% in 2025. Both panels share one y-scale.
Show the data table
| Year | China: open / all | US: open / all |
|---|---|---|
| 2023 | 13 / 23 (56.5%) | 37 / 86 (43%) |
| 2024 | 16 / 31 (51.6%) | 20 / 57 (35.1%) |
| 2025 | 28 / 35 (80%) | 11 / 64 (17.2%) |
| 2026 YTD | 20 / 29 (69%) | 4 / 49 (8.2%) |
File: knowledge/data/derived/open_weight_notable_by_country_year.csv
Models built on Chinese base models took 54.4% of downloads among Hugging Face's top 1,000 text models; Qwen alone took 42.29%.
- China 54.4%
- US 26.3%
- Other countries 1.4%
- Not mapped 18%
“Downloads” are config-file requests over 30 days, so CI pipelines count; they are not users. Origin follows one base_model hop, and the 18% we could not map is shown on its own. This is not a count of deployments in industry.
Show the data table
| Base-model origin | 30-day downloads | Share | Models in top 1,000 |
|---|---|---|---|
| China | 210,640,305 | 54.4% | 402 |
| United States of America | 101,722,708 | 26.3% | 272 |
| (origin not mapped) | 69,719,796 | 18% | 299 |
| France | 3,766,185 | 1% | 21 |
| South Korea | 974,457 | 0.3% | 5 |
| Japan | 380,797 | 0.1% | 1 |
File: knowledge/data/derived/hf_downloads_by_origin_country.csv, hf_downloads_by_origin_org.csv
What it means for your industry
On-premise and edge deployments use open weights — factories with air-gapped OT networks, ships with limited bandwidth, utility control rooms. For an operator, the weights' country of origin becomes a supply-chain question: who trained it, on what, and under which licence.
- The author-to-organisation map is hand-kept, so the unmapped share is always shown. Some of the download volume is Hugging Face's own CI test fixtures.
Story 3 · The gap
Has the open-weight gap closed? It depends on the ruler
Two rulers give two answers to the same question. We show both, side by side, on one shared time axis.
On Epoch's benchmark index (ECI), the closed–open gap narrowed to 1.2 points in Jul 2024 and is back to 11.0 in Oct 2026.
- Shaded = under ~4 pts, within the 90% CI (noise)
The gap between the best scores among models released up to each month. ECI has no meaningful zero, so we plot the gap, from zero, not the levels.
Show the data table
| Month | Best closed (ECI) | Best open (ECI) | Gap (pts) |
|---|---|---|---|
| 2023-03 | GPT-4 (Mar 2023) 125.74 | LLaMA-65B 109.97 | 15.77 |
| 2023-04 | GPT-4 (Mar 2023) 125.74 | LLaMA-65B 109.97 | 15.77 |
| 2023-05 | GPT-4 (Mar 2023) 125.74 | LLaMA-65B 109.97 | 15.77 |
| 2023-06 | GPT-4 (Mar 2023) 125.74 | LLaMA-65B 109.97 | 15.77 |
| 2023-07 | GPT-4 (Mar 2023) 125.74 | Stable Beluga 2 117.02 | 8.72 |
| 2023-08 | GPT-4 (Mar 2023) 125.74 | Stable Beluga 2 117.02 | 8.72 |
| 2023-09 | GPT-4 (Mar 2023) 125.74 | Stable Beluga 2 117.02 | 8.72 |
| 2023-10 | GPT-4 (Mar 2023) 125.74 | Stable Beluga 2 117.02 | 8.72 |
| 2023-11 | GPT-4 (Mar 2023) 125.74 | Yi-34B 117.33 | 8.41 |
| 2023-12 | GPT-4 (Mar 2023) 125.74 | Mixtral 8x7B 118.44 | 7.3 |
| 2024-01 | GPT-4 Turbo (Nov 2023) 126.43 | Mixtral 8x7B 118.44 | 7.99 |
| 2024-02 | Claude 3 Opus 126.9 | Mixtral 8x7B 118.44 | 8.46 |
| 2024-03 | Claude 3 Opus 126.9 | Mixtral 8x7B 118.44 | 8.46 |
| 2024-04 | GPT-4 Turbo (Apr 2024) 127.27 | Llama 3-70B 123.01 | 4.26 |
| 2024-05 | GPT-4o (May 2024) 128.99 | DeepSeek-V2 (MoE-236B, May 2024) 124.8 | 4.19 |
| 2024-06 | Claude 3.5 Sonnet 130 | Qwen2-72B 125.28 | 4.72 |
| 2024-07 | Claude 3.5 Sonnet 130 | Llama 3.1-405B 128.77 | 1.23 |
| 2024-08 | Claude 3.5 Sonnet 130 | Llama 3.1-405B 128.77 | 1.23 |
| 2024-09 | o1-mini 135.7 | Qwen2.5-72B 128.97 | 6.73 |
| 2024-10 | o1-mini 135.7 | Qwen2.5-72B 128.97 | 6.73 |
| 2024-11 | o1-mini 135.7 | Qwen2.5-72B 128.97 | 6.73 |
| 2024-12 | o1 141.72 | DeepSeek-V3 132.3 | 9.42 |
| 2025-01 | o1 141.72 | DeepSeek-R1 138.82 | 2.9 |
| 2025-02 | o1 141.72 | DeepSeek-R1 138.82 | 2.9 |
| 2025-03 | Gemini 2.5 Pro (Mar 2025) 143.96 | DeepSeek-R1 138.82 | 5.14 |
| 2025-04 | o3 146.73 | Qwen3-235B-A22B 139.15 | 7.58 |
| 2025-05 | o3 146.73 | DeepSeek-R1 (May 2025) 141.13 | 5.6 |
| 2025-06 | o3-pro 147.33 | DeepSeek-R1 (May 2025) 141.13 | 6.2 |
| 2025-07 | o3-pro 147.33 | Qwen3-235B-A22B-Thinking (Jul 2025) 143.76 | 3.57 |
| 2025-08 | GPT-5 150 | Qwen3-235B-A22B-Thinking (Jul 2025) 143.76 | 6.24 |
| 2025-09 | GPT-5 150 | DeepSeek-V3.2-Exp 144.91 | 5.09 |
| 2025-10 | GPT-5 Pro 150.59 | DeepSeek-V3.2-Exp 144.91 | 5.68 |
| 2025-11 | Gemini 3 Pro 153.07 | Kimi K2 Thinking 145.82 | 7.25 |
| 2025-12 | GPT-5.2 Pro 156.33 | DeepSeek-V3.2 146.09 | 10.24 |
| 2026-01 | GPT-5.2 Pro 156.33 | Kimi K2.5 148.07 | 8.26 |
| 2026-02 | GPT-5.3 Codex 157.09 | Kimi K2.5 148.07 | 9.02 |
| 2026-03 | GPT-5.4 Pro 159.57 | Kimi K2.5 148.07 | 11.5 |
| 2026-04 | GPT-5.5 Pro 162.91 | Kimi K2.6 150.94 | 11.97 |
| 2026-05 | GPT-5.5 Pro 162.91 | Kimi K2.6 150.94 | 11.97 |
| 2026-06 | Claude Fable 5 163.42 | GLM-5.2 151.94 | 11.48 |
| 2026-07 | Claude Opus 5 163.94 | Kimi K3 157.86 | 6.08 |
| 2026-08 | Claude Opus 5 163.94 | Kimi K3 157.86 | 6.08 |
| 2026-09 | Claude Opus 5.5 168.83 | Kimi K3 157.86 | 10.97 |
| 2026-10 | Claude Opus 5.5 168.83 | Kimi K3 157.86 | 10.97 |
File: knowledge/data/derived/eci_best_closed_vs_open_monthly.csv
On human-preference votes (Arena, style-controlled) the best open model trails the best closed one by 2.0% (Oct 22, 2026).
Only snapshots that contain OpenAI, Anthropic and Google models are computed. “Open” means any licence other than Proprietary, including non-commercial and custom licences.
Show the data table
| Snapshot | Top closed | Top open | Gap (% of open) |
|---|---|---|---|
| 2024-08-28 | chatgpt-4o-latest 1288.8 | llama-3.1-405b-instruct 1249.7 | 3.13% |
| 2024-09-04 | chatgpt-4o-latest 1288.7 | llama-3.1-405b-instruct 1250 | 3.09% |
| 2024-09-15 | gpt-4o-2024-05-13 1261.6 | llama-3.1-405b-instruct-bf16 1250.9 | 0.86% |
| 2024-09-17 | o1-preview 1325.7 | llama-3.1-405b-instruct-fp8 1249.8 | 6.07% |
| 2024-09-27 | o1-preview 1308.7 | llama-3.1-405b-instruct-bf16 1253.5 | 4.4% |
| 2024-10-07 | o1-preview 1300.2 | llama-3.1-405b-instruct-bf16 1252.2 | 3.83% |
| 2024-10-15 | o1-preview 1300.5 | llama-3.1-405b-instruct-fp8 1251.7 | 3.9% |
| 2024-10-23 | o1-preview 1302.9 | llama-3.1-405b-instruct-fp8 1251.4 | 4.11% |
| 2024-10-28 | o1-preview 1301.6 | llama-3.1-405b-instruct-fp8 1251.5 | 4.01% |
| 2024-11-04 | o1-preview 1300.5 | llama-3.1-405b-instruct-fp8 1252.2 | 3.85% |
| 2024-11-12 | o1-preview 1300.1 | llama-3.1-405b-instruct-fp8 1252.4 | 3.81% |
| 2024-11-13 | o1-preview 1299.7 | llama-3.1-405b-instruct-fp8 1252.5 | 3.77% |
| 2024-11-20 | o1-preview 1300.6 | llama-3.1-405b-instruct-fp8 1252.3 | 3.86% |
| 2024-11-21 | o1-preview 1300.5 | llama-3.1-405b-instruct-fp8 1252.2 | 3.86% |
| 2024-11-22 | o1-preview 1300.5 | llama-3.1-405b-instruct-fp8 1252.2 | 3.86% |
| 2024-12-01 | o1-preview 1301.5 | llama-3.1-405b-instruct-fp8 1252.7 | 3.89% |
| 2024-12-05 | o1-preview 1301.9 | llama-3.1-405b-instruct-fp8 1252.4 | 3.95% |
| 2024-12-10 | o1-preview 1301.7 | llama-3.1-405b-instruct-fp8 1252.2 | 3.95% |
| 2024-12-15 | o1-preview 1301.8 | llama-3.1-405b-instruct-fp8 1252.3 | 3.95% |
| 2024-12-18 | o1-preview 1302.4 | llama-3.1-405b-instruct-fp8 1252.3 | 4% |
| 2024-12-22 | o1-preview 1302.6 | llama-3.1-405b-instruct-fp8 1252.4 | 4% |
| 2024-12-30 | o1-2024-12-17 1327.9 | deepseek-v3 1274.2 | 4.22% |
| 2025-01-05 | o1-2024-12-17 1322.8 | deepseek-v3 1277.5 | 3.55% |
| 2025-01-15 | o1-2024-12-17 1321.6 | deepseek-v3 1278.6 | 3.36% |
| 2025-01-19 | o1-2024-12-17 1323 | deepseek-v3 1278.9 | 3.45% |
| 2025-01-21 | o1-2024-12-17 1322.5 | deepseek-v3 1278.7 | 3.42% |
| 2025-01-22 | o1-2024-12-17 1322.5 | deepseek-v3 1278.7 | 3.42% |
| 2025-01-24 | o1-2024-12-17 1322.9 | deepseek-r1 1316.2 | 0.51% |
| 2025-01-28 | o1-2024-12-17 1322.3 | deepseek-r1 1316.1 | 0.47% |
| 2025-02-03 | o1-2024-12-17 1322.6 | deepseek-r1 1319 | 0.27% |
| 2025-02-05 | o1-2024-12-17 1322.8 | deepseek-r1 1319.7 | 0.23% |
| 2025-02-06 | o1-2024-12-17 1322.6 | deepseek-r1 1319.8 | 0.21% |
| 2025-02-09 | o1-2024-12-17 1322 | deepseek-r1 1319.2 | 0.22% |
| 2025-02-11 | o1-2024-12-17 1323.1 | deepseek-r1 1319.8 | 0.25% |
| 2025-02-14 | o1-2024-12-17 1322.7 | deepseek-r1 1319.7 | 0.23% |
| 2025-02-17 | o1-2024-12-17 1323.5 | deepseek-r1 1320 | 0.26% |
| 2025-02-21 | o1-2024-12-17 1322.2 | deepseek-r1 1319.5 | 0.21% |
| 2025-02-27 | o1-2024-12-17 1323 | deepseek-r1 1320.3 | 0.21% |
| 2025-03-03 | gpt-4.5-preview-2025-02-27 1369.8 | deepseek-r1 1320.3 | 3.75% |
| 2025-03-11 | gpt-4.5-preview-2025-02-27 1362.3 | deepseek-r1 1318.1 | 3.35% |
| 2025-03-15 | gpt-4.5-preview-2025-02-27 1358.4 | deepseek-r1 1316.7 | 3.16% |
| 2025-03-17 | gpt-4.5-preview-2025-02-27 1357.9 | deepseek-r1 1316.8 | 3.12% |
| 2025-03-20 | gpt-4.5-preview-2025-02-27 1356.8 | deepseek-r1 1317.1 | 3.01% |
| 2025-03-25 | gpt-4.5-preview-2025-02-27 1356.4 | deepseek-r1 1316.4 | 3.04% |
| 2025-03-26 | chatgpt-4o-latest-20250326 1358.8 | deepseek-r1 1316.5 | 3.21% |
| 2025-03-30 | chatgpt-4o-latest-20250326 1358.8 | deepseek-v3-0324 1319.2 | 3% |
| 2025-04-05 | chatgpt-4o-latest-20250326 1362.5 | deepseek-v3-0324 1319.2 | 3.29% |
| 2025-04-08 | chatgpt-4o-latest-20250326 1362.5 | deepseek-v3-0324 1319.2 | 3.29% |
| 2025-04-09 | chatgpt-4o-latest-20250326 1359.5 | deepseek-v3-0324 1320.6 | 2.95% |
| 2025-04-16 | chatgpt-4o-latest-20250326 1359.8 | deepseek-v3-0324 1321.8 | 2.87% |
| 2025-04-23 | o3-2025-04-16 1380.3 | deepseek-v3-0324 1323.1 | 4.32% |
| 2025-05-05 | o3-2025-04-16 1372.8 | deepseek-v3-0324 1322.1 | 3.84% |
| 2025-05-06 | o3-2025-04-16 1373.3 | deepseek-v3-0324 1323 | 3.8% |
| 2025-05-11 | o3-2025-04-16 1375.2 | deepseek-v3-0324 1323.3 | 3.92% |
| 2025-05-19 | o3-2025-04-16 1434.4 | deepseek-v3-0324 1382.4 | 3.77% |
| 2025-05-20 | o3-2025-04-16 1434.4 | deepseek-v3-0324 1382.4 | 3.77% |
| 2025-05-22 | o3-2025-04-16 1435.2 | deepseek-v3-0324 1382 | 3.85% |
| 2025-06-01 | o3-2025-04-16 1441.8 | deepseek-v3-0324 1385.5 | 4.06% |
| 2025-06-03 | o3-2025-04-16 1441.4 | deepseek-v3-0324 1385.8 | 4.01% |
| 2025-06-05 | o3-2025-04-16 1442.8 | deepseek-v3-0324 1388.4 | 3.92% |
| 2025-06-11 | o3-2025-04-16 1447.1 | deepseek-v3-0324 1393.5 | 3.85% |
| 2025-06-17 | o3-2025-04-16 1448.9 | deepseek-r1-0528 1410.2 | 2.75% |
| 2025-06-18 | o3-2025-04-16 1449.1 | deepseek-r1-0528 1411 | 2.7% |
| 2025-06-24 | gemini-2.5-pro 1466.6 | deepseek-r1-0528 1412.8 | 3.81% |
| 2025-07-01 | gemini-2.5-pro 1463.4 | deepseek-r1-0528 1412.7 | 3.59% |
| 2025-07-07 | gemini-2.5-pro 1465.4 | deepseek-r1-0528 1414.1 | 3.63% |
| 2025-07-15 | gemini-2.5-pro 1461.8 | deepseek-r1-0528 1414.4 | 3.35% |
| 2025-07-17 | gemini-2.5-pro 1461.7 | kimi-k2-0711-preview 1419.8 | 2.95% |
| 2025-07-25 | gemini-2.5-pro 1460.2 | kimi-k2-0711-preview 1418.1 | 2.97% |
| 2025-07-28 | gemini-2.5-pro 1462.8 | kimi-k2-0711-preview 1419.8 | 3.03% |
| 2025-08-01 | gemini-2.5-pro 1458.5 | qwen3-235b-a22b-instruct-2507 1432.9 | 1.78% |
| 2025-08-04 | gemini-2.5-pro 1458.6 | qwen3-235b-a22b-instruct-2507 1429.5 | 2.04% |
| 2025-08-07 | gpt-5-high 1481.4 | qwen3-235b-a22b-instruct-2507 1428.1 | 3.73% |
| 2025-08-14 | gpt-5-high 1463.3 | qwen3-235b-a22b-instruct-2507 1426.2 | 2.6% |
| 2025-08-18 | gpt-5-high 1460.7 | qwen3-235b-a22b-instruct-2507 1423.2 | 2.64% |
| 2025-08-21 | gemini-2.5-pro 1457.1 | qwen3-235b-a22b-instruct-2507 1422.5 | 2.43% |
| 2025-08-27 | gemini-2.5-pro 1456.4 | kimi-k2-0711-preview 1421.2 | 2.48% |
| 2025-08-28 | gemini-2.5-pro 1455.9 | kimi-k2-0711-preview 1420.9 | 2.46% |
| 2025-09-08 | gemini-2.5-pro 1455 | kimi-k2-0711-preview 1419.6 | 2.49% |
| 2025-09-17 | gemini-2.5-pro 1456 | qwen3-235b-a22b-instruct-2507 1419.6 | 2.56% |
| 2025-09-18 | gemini-2.5-pro 1456.1 | qwen3-235b-a22b-instruct-2507 1419.1 | 2.61% |
| 2025-09-30 | gemini-2.5-pro 1452.9 | qwen3-vl-235b-a22b-instruct 1420.5 | 2.28% |
| 2025-10-01 | gemini-2.5-pro 1452.1 | qwen3-235b-a22b-instruct-2507 1418.8 | 2.35% |
| 2025-10-03 | claude-sonnet-4-5-20250929-thinking-32k 1453 | glm-4.6 1425.7 | 1.92% |
| 2025-10-07 | gemini-2.5-pro 1451.8 | glm-4.6 1421.4 | 2.14% |
| 2025-10-16 | gemini-2.5-pro 1450.7 | glm-4.6 1421.7 | 2.04% |
| 2025-10-17 | gemini-2.5-pro 1450.8 | glm-4.6 1421.6 | 2.05% |
| 2025-11-05 | gemini-2.5-pro 1451.6 | glm-4.6 1427.8 | 1.67% |
| 2025-11-06 | gemini-2.5-pro 1451.9 | glm-4.6 1429.1 | 1.6% |
| 2025-11-09 | gemini-2.5-pro 1452 | glm-4.6 1428.7 | 1.63% |
| 2025-11-16 | gemini-3-pro 1500.6 | kimi-k2-thinking-turbo 1431.8 | 4.8% |
| 2025-11-17 | gemini-3-pro 1502.2 | kimi-k2-thinking-turbo 1431.7 | 4.92% |
| 2025-11-18 | gemini-3-pro 1497.9 | kimi-k2-thinking-turbo 1430.4 | 4.72% |
| 2025-11-20 | gemini-3-pro 1495 | kimi-k2-thinking-turbo 1429.5 | 4.58% |
| 2025-11-26 | gemini-3-pro 1492.4 | kimi-k2-thinking-turbo 1430.8 | 4.3% |
| 2025-12-01 | gemini-3-pro 1491.4 | kimi-k2-thinking-turbo 1429.1 | 4.36% |
| 2025-12-04 | gemini-3-pro 1490.8 | kimi-k2-thinking-turbo 1427.7 | 4.42% |
| 2025-12-05 | gemini-3-pro 1491.3 | kimi-k2-thinking-turbo 1427.6 | 4.46% |
| 2025-12-08 | gemini-3-pro 1490.3 | kimi-k2-thinking-turbo 1427.1 | 4.42% |
| 2025-12-11 | gemini-3-pro 1491.8 | kimi-k2-thinking-turbo 1425.5 | 4.65% |
| 2025-12-15 | gemini-3-pro 1491.2 | kimi-k2-thinking-turbo 1427 | 4.49% |
| 2025-12-16 | gemini-3-pro 1492.1 | kimi-k2-thinking-turbo 1428.3 | 4.47% |
| 2025-12-18 | gemini-3-pro 1491.4 | kimi-k2-thinking-turbo 1426.3 | 4.56% |
| 2025-12-19 | gemini-3-pro 1490.1 | kimi-k2-thinking-turbo 1428.1 | 4.34% |
| 2025-12-21 | gemini-3-pro 1490 | kimi-k2-thinking-turbo 1428.4 | 4.31% |
| 2025-12-23 | gemini-3-pro 1489.9 | kimi-k2-thinking-turbo 1428.4 | 4.31% |
| 2025-12-30 | gemini-3-pro 1490.2 | glm-4.7 1438.8 | 3.57% |
| 2026-01-09 | gemini-3-pro 1489.6 | glm-4.7 1442.6 | 3.26% |
| 2026-01-27 | gemini-3-pro 1487.9 | glm-4.7 1440.8 | 3.27% |
| 2026-01-29 | gemini-3-pro 1487.3 | kimi-k2.5-thinking 1449.6 | 2.6% |
| 2026-02-06 | claude-opus-4-6-thinking 1503.5 | kimi-k2.5-thinking 1448.8 | 3.78% |
| 2026-02-09 | claude-opus-4-6-thinking 1504.4 | kimi-k2.5-thinking 1446.7 | 3.99% |
| 2026-02-10 | claude-opus-4-6-thinking 1505 | glm-5 1470.4 | 2.36% |
| 2026-02-11 | claude-opus-4-6-thinking 1505.9 | glm-5 1451.5 | 3.74% |
| 2026-02-16 | claude-opus-4-6-thinking 1507 | glm-5 1455 | 3.57% |
| 2026-02-17 | claude-opus-4-6-thinking 1506 | glm-5 1455.2 | 3.49% |
| 2026-02-19 | claude-opus-4-6 1505.1 | glm-5 1454.5 | 3.48% |
| 2026-02-20 | claude-opus-4-6 1505.1 | glm-5 1454.7 | 3.46% |
| 2026-02-21 | claude-opus-4-6-thinking 1505.5 | glm-5 1452.4 | 3.65% |
| 2026-02-22 | claude-opus-4-6-thinking 1505.5 | glm-5 1452.4 | 3.65% |
| 2026-02-24 | claude-opus-4-6 1504.2 | glm-5 1453 | 3.52% |
| 2026-02-25 | claude-opus-4-6 1503.8 | glm-5 1453.8 | 3.44% |
| 2026-02-26 | claude-opus-4-6-thinking 1503.4 | glm-5 1455.4 | 3.3% |
| 2026-03-03 | claude-opus-4-6 1505.1 | glm-5 1454.4 | 3.49% |
| 2026-03-04 | claude-opus-4-6 1504.1 | glm-5 1453.2 | 3.5% |
| 2026-03-05 | claude-opus-4-6 1504 | glm-5 1452.6 | 3.53% |
| 2026-03-06 | claude-opus-4-6 1503.5 | glm-5 1451.7 | 3.57% |
| 2026-03-11 | claude-opus-4-6 1503.4 | glm-5 1453.9 | 3.4% |
| 2026-03-16 | claude-opus-4-6-thinking 1501.4 | glm-5 1453.5 | 3.29% |
| 2026-03-18 | claude-opus-4-6-thinking 1502.1 | glm-5 1456.6 | 3.13% |
| 2026-03-19 | claude-opus-4-6-thinking 1502.1 | glm-5 1456.4 | 3.14% |
| 2026-03-20 | claude-opus-4-6-thinking 1502.1 | glm-5 1455.3 | 3.22% |
| 2026-03-26 | claude-opus-4-6-thinking 1504.1 | glm-5 1455.8 | 3.32% |
| 2026-03-31 | claude-opus-4-6-thinking 1504.2 | glm-5 1455.7 | 3.33% |
| 2026-04-02 | claude-opus-4-6-thinking 1503.7 | glm-5 1455.4 | 3.32% |
| 2026-04-07 | claude-opus-4-6-thinking 1502.8 | glm-5.1 1467.4 | 2.41% |
| 2026-04-09 | claude-opus-4-6-thinking 1503.5 | glm-5.1 1470.7 | 2.24% |
| 2026-04-10 | claude-opus-4-6-thinking 1503.8 | glm-5.1 1470.9 | 2.24% |
| 2026-04-14 | claude-opus-4-6-thinking 1502.3 | glm-5.1 1471.2 | 2.12% |
| 2026-04-17 | claude-opus-4-7-thinking 1504.5 | glm-5.1 1471.6 | 2.24% |
| 2026-04-20 | claude-opus-4-7-thinking 1504 | glm-5.1 1468.8 | 2.39% |
| 2026-04-22 | claude-opus-4-7-thinking 1502.6 | glm-5.1 1468.5 | 2.33% |
| 2026-04-23 | claude-opus-4-7-thinking 1503.1 | glm-5.1 1469.6 | 2.28% |
| 2026-04-27 | claude-opus-4-7-thinking 1503.2 | glm-5.1 1469.9 | 2.26% |
| 2026-04-28 | claude-opus-4-7-thinking 1503.2 | glm-5.1 1469.9 | 2.26% |
| 2026-04-29 | claude-opus-4-7-thinking 1502.9 | glm-5.1 1469.1 | 2.3% |
| 2026-04-30 | claude-opus-4-7-thinking 1504 | glm-5.1 1470.3 | 2.29% |
| 2026-05-01 | claude-opus-4-7-thinking 1502.7 | glm-5.1 1470.5 | 2.19% |
| 2026-05-06 | claude-opus-4-7-thinking 1503.3 | glm-5.1 1471.2 | 2.18% |
| 2026-05-07 | claude-opus-4-7-thinking 1503 | glm-5.1 1471.3 | 2.16% |
| 2026-05-12 | claude-opus-4-6-thinking 1501.7 | glm-5.1 1471.1 | 2.08% |
| 2026-05-14 | claude-opus-4-6-thinking 1501.5 | glm-5.1 1472.3 | 1.99% |
| 2026-05-18 | claude-opus-4-6-thinking 1501.8 | glm-5.1 1472.2 | 2.01% |
| 2026-05-19 | claude-opus-4-6-thinking 1502 | glm-5.1 1472.2 | 2.02% |
| 2026-05-27 | claude-opus-4-6-thinking 1502.2 | glm-5.1 1473.9 | 1.92% |
| 2026-06-03 | claude-opus-4-6-thinking 1503.3 | glm-5.1 1475.2 | 1.91% |
| 2026-06-04 | claude-opus-4-6-thinking 1503.5 | glm-5.1 1475.4 | 1.9% |
| 2026-06-05 | claude-opus-4-6-thinking 1503.6 | glm-5.1 1475 | 1.94% |
| 2026-06-10 | claude-fable-5 1510.4 | glm-5.1 1475.2 | 2.38% |
| 2026-06-16 | claude-fable-5 1507.6 | glm-5.1 1475.3 | 2.19% |
| 2026-06-25 | claude-fable-5 1508.2 | glm-5.1 1473.4 | 2.36% |
| 2026-07-02 | claude-fable-5 1508.5 | glm-5.1 1472.2 | 2.47% |
| 2026-07-10 | claude-fable-5 1508.6 | glm-5.1 1472.1 | 2.48% |
| 2026-07-12 | claude-fable-5 1507.5 | glm-5.1 1471.9 | 2.42% |
| 2026-07-14 | claude-fable-5 1507.5 | glm-5.1 1471.7 | 2.44% |
| 2026-07-16 | claude-fable-5 1507.5 | glm-5.1 1470.8 | 2.49% |
| 2026-07-20 | claude-fable-5 1506.8 | glm-5.1 1469.9 | 2.51% |
| 2026-07-21 | claude-fable-5 1507.3 | glm-5.1 1469.6 | 2.56% |
| 2026-07-26 | claude-fable-5 1508 | kimi-k3-max 1485.3 | 1.53% |
| 2026-07-27 | claude-fable-5 1507.6 | kimi-k3-max 1485.8 | 1.47% |
| 2026-07-30 | claude-fable-5 1507.6 | kimi-k3-max 1485.8 | 1.47% |
| 2026-08-02 | claude-fable-5 1508.6 | kimi-k3-max 1485.3 | 1.56% |
| 2026-08-03 | claude-fable-5 1508.6 | glm-5.2-max 1469.5 | 2.66% |
| 2026-08-06 | claude-fable-5 1507.3 | kimi-k3-max 1485.4 | 1.48% |
| 2026-08-10 | claude-fable-5 1505.7 | kimi-k3-max 1487.3 | 1.23% |
| 2026-08-11 | claude-fable-5 1506.4 | kimi-k3-max 1486.8 | 1.32% |
| 2026-08-12 | claude-fable-5 1506.6 | kimi-k3-max 1488.6 | 1.21% |
| 2026-08-19 | claude-fable-5 1507.3 | kimi-k3-max 1489.7 | 1.19% |
| 2026-08-21 | claude-fable-5 1507.8 | kimi-k3-max 1489.1 | 1.25% |
| 2026-08-27 | claude-fable-5 1507.5 | kimi-k3-max 1488.8 | 1.25% |
| 2026-09-01 | claude-fable-5 1507.4 | kimi-k3-max 1489 | 1.23% |
| 2026-09-02 | claude-fable-5 1507.2 | kimi-k3-max 1488.7 | 1.24% |
| 2026-09-11 | claude-fable-5 1505.5 | glm-5.3-max 1485.6 | 1.34% |
| 2026-09-13 | claude-fable-5 1505.7 | kimi-k3-max 1484.8 | 1.41% |
| 2026-09-25 | claude-opus-5.5-high 1508.6 | kimi-k3-max 1488 | 1.38% |
| 2026-09-30 | gemini-4-argon-high 1524.8 | kimi-k3-max 1488 | 2.47% |
| 2026-10-02 | gemini-4-argon-high 1525.2 | kimi-k3-max 1488.4 | 2.48% |
| 2026-10-08 | gemini-4-argon-high 1525.4 | kimi-k3-max 1488.1 | 2.51% |
| 2026-10-13 | gemini-4-argon-high 1524.8 | kimi-k3-max 1488.1 | 2.46% |
| 2026-10-15 | Gemini 4 Argon (High) 1523.9 | Kimi K3 (Max) 1487.1 | 2.48% |
| 2026-10-20 | Gemini 4 Argon (High) 1523.9 | Kimi K3 (Max) 1488.1 | 2.41% |
| 2026-10-22 | Gemini 4 Argon (High) 1517.2 | Kimi K3 (Max) 1487.4 | 2% |
File: knowledge/data/derived/arena_top_closed_vs_open.csv
Since 2025, each new best open-weight model matched a closed model's score a median 5.4 months later; the fastest was DeepSeek-R1 at 1.1 months.
- US lab
- Chinese lab
- Other country
- Rule = median since 2025, 5.4 months
LLaMA-65B has a negative lag, an artefact of ECI coverage starting 2023-02-24, so it is not drawn (it is in the table). Measured on one index, against the first closed model to reach the score.
Show the data table
| Open model | Country | Released | First closed model at that score | Lag (months) |
|---|---|---|---|---|
| LLaMA-65B | United States of America | 2023-02-24 | GPT-4 (Mar 2023) (2023-03-14) | -0.6 |
| Llama 2-70B | United States of America | 2023-07-18 | GPT-4 (Mar 2023) (2023-03-14) | 4.1 |
| Stable Beluga 2 | United Kingdom | 2023-07-20 | GPT-4 (Mar 2023) (2023-03-14) | 4.2 |
| Yi-34B | China | 2023-11-02 | GPT-4 (Mar 2023) (2023-03-14) | 7.7 |
| Mixtral 8x7B | France | 2023-12-11 | GPT-4 (Mar 2023) (2023-03-14) | 8.9 |
| Mixtral 8x22B | France | 2024-04-17 | GPT-4 (Mar 2023) (2023-03-14) | 13.1 |
| Llama 3-70B | United States of America | 2024-04-18 | GPT-4 (Mar 2023) (2023-03-14) | 13.2 |
| DeepSeek-V2 (MoE-236B, May 2024) | China | 2024-05-07 | GPT-4 (Mar 2023) (2023-03-14) | 13.8 |
| Qwen2-72B | China | 2024-06-07 | GPT-4 (Mar 2023) (2023-03-14) | 14.8 |
| Llama 3.1-405B | United States of America | 2024-07-23 | GPT-4o (May 2024) (2024-05-13) | 2.3 |
| Qwen2.5-72B | China | 2024-09-19 | GPT-4o (May 2024) (2024-05-13) | 4.2 |
| Phi-4 | United States of America | 2024-12-12 | o1-mini (2024-09-12) | 3 |
| DeepSeek-V3 | China | 2024-12-26 | o1-mini (2024-09-12) | 3.4 |
| DeepSeek-R1 | China | 2025-01-20 | o1 (2024-12-17) | 1.1 |
| Qwen3-235B-A22B | China | 2025-04-28 | o1 (2024-12-17) | 4.3 |
| DeepSeek-R1 (May 2025) | China | 2025-05-28 | o1 (2024-12-17) | 5.3 |
| Qwen3-235B-A22B-Thinking (Jul 2025) | China | 2025-07-25 | Gemini 2.5 Pro (Mar 2025) (2025-03-31) | 3.8 |
| DeepSeek-V3.2-Exp | China | 2025-09-29 | o3 (2025-04-16) | 5.5 |
| Kimi K2 Thinking | China | 2025-11-06 | o3 (2025-04-16) | 6.7 |
| DeepSeek-V3.2 | China | 2025-12-01 | o3 (2025-04-16) | 7.5 |
| Kimi K2.5 | China | 2026-01-27 | GPT-5 (2025-08-07) | 5.7 |
| GLM-5.1 | China | 2026-04-07 | GPT-5 (2025-08-07) | 8 |
| Kimi K2.6 | China | 2026-04-20 | Gemini 3 Pro (2025-11-18) | 5 |
| GLM-5.2 | China | 2026-06-16 | Gemini 3 Pro (2025-11-18) | 6.9 |
| Kimi K3 | China | 2026-07-16 | GPT-5.4 Pro (2026-03-05) | 4.4 |
File: knowledge/data/derived/eci_open_weight_lag.csv
What it means for your industry
The time for an open model to reach a closed model's score (a median 5.4 months since 2025) is shorter than most enterprise procurement cycles. A terminal operator that standardises on open weights for on-prem use gets a model about as capable as last year's closed frontier.
- Arena ratings are relative within one snapshot, so compare gaps, not levels, across years. Single records have wide confidence intervals; trust the median, not one dot.
- The AI Index 2026 says GPT-4-0314 led Vicuna-13B by 174 points (15.2%) in May 2023 (PDF p.76). The Arena archive we use has no GPT-4 rows in snapshots from that period, so we do not compute the gap there. Stanford HAI, AI Index 2026 (CC BY-ND 4.0 — quoted, not redrawn)
Story 4 · Release type
The best models are increasingly not released as weights at all
API-only release became the most common route. Counts by release type and the open-weight share are drawn as two charts, not one dual-axis chart.
In 2025, API-only release (46 models) overtook open weights (45) for the first time.
- Open weights
- API access
- Hosted or limited, no API
- Unreleased or not recorded
- * 2026 = partial year
Epoch's list is a curation, not a census, so the mix also moves with what Epoch chooses to include.
Show the data table
| Year | Open weights | API access | Hosted, no API | Limited | Unreleased | Not recorded / other | Total | Open share |
|---|---|---|---|---|---|---|---|---|
| 2019 | 31 | 0 | 0 | 0 | 37 | 0 | 68 | 45.6% |
| 2020 | 28 | 1 | 0 | 0 | 21 | 0 | 50 | 56% |
| 2021 | 41 | 4 | 5 | 0 | 29 | 0 | 79 | 51.9% |
| 2022 | 41 | 8 | 5 | 0 | 36 | 1 | 91 | 45.6% |
| 2023 | 59 | 20 | 4 | 0 | 36 | 0 | 119 | 49.6% |
| 2024 | 39 | 30 | 13 | 0 | 16 | 0 | 98 | 39.8% |
| 2025 | 45 | 46 | 9 | 0 | 7 | 0 | 107 | 42.1% |
| 2026 YTD | 29 | 43 | 5 | 3 | 4 | 0 | 84 | 34.5% |
File: knowledge/data/derived/notable_models_by_access_year.csv
What it means for your industry
API-only means the operator's data leaves the site. In regulated or safety-critical operations — OT networks, vessel systems, grid control — the release type is a security control, not a pricing detail.
- The AI Index 2026 counted 47 of 102 models as API-only for 2025 (PDF p.21), and notes that “the most capable models are now the least transparent” (PDF p.15). Stanford HAI, AI Index 2026 (CC BY-ND 4.0 — quoted, not redrawn)
Story 5 · Compute
The compute gap is roughly 47×
The largest training run per year, US and China. Marker fill shows how confident Epoch's estimate is. The axis is a log scale.
The largest US training run of 2026 used about 47× the compute of China's largest — both are Epoch estimates, so read it as “an order of magnitude or more”.
- US
- China
- Filled = Confident
- Faded = Likely
- Open ring = Speculative
In 2026 only 7 US and 9 Chinese notable models have a compute figure; the rest are empty, not zero.
Show the data table
| Year | Country | Model | FLOP | Confidence | Models with / without compute |
|---|---|---|---|---|---|
| 2019 | United States of America | OpenAI Five | 6.7×10²² | Confident | 32 / 24 |
| 2020 | United States of America | GPT-3 175B (davinci) | 3.1×10²³ | Confident | 26 / 11 |
| 2021 | United States of America | FLAN 137B | 2×10²⁴ | Confident | 42 / 12 |
| 2022 | United States of America | Minerva (540B) | 2.7×10²⁴ | Confident | 45 / 23 |
| 2023 | United States of America | Gemini 1.0 Ultra | 5×10²⁵ | Speculative | 44 / 42 |
| 2024 | United States of America | Llama 3.1-405B | 3.8×10²⁵ | Confident | 26 / 31 |
| 2025 | United States of America | Grok 4 | 5×10²⁶ | Speculative | 15 / 49 |
| 2026 YTD | United States of America | GPT-6 Astra | 9.4×10²⁶ | Likely | 7 / 42 |
| 2019 | China | trRosetta | 3.8×10¹⁹ | Confident | 3 / 3 |
| 2020 | China | KEPLER | 1.7×10²¹ | Confident | 5 / 3 |
| 2021 | China | ERNIE 3.0 Titan | 1×10²⁴ | Confident | 17 / 4 |
| 2022 | China | GLM-130B | 3.5×10²³ | Confident | 7 / 5 |
| 2023 | China | Qwen-72B | 1.3×10²⁴ | Confident | 15 / 8 |
| 2024 | China | Doubao-pro | 2.5×10²⁵ | Likely | 16 / 15 |
| 2025 | China | Qwen3-Max | 1.5×10²⁵ | Speculative | 23 / 12 |
| 2026 YTD | China | Kimi K3 | 2×10²⁵ | Speculative | 9 / 20 |
File: knowledge/data/derived/max_training_compute_us_china_year.csv
What it means for your industry
The China data points to doing more with less compute — the same constraint every on-prem industrial deployment lives under.
- Most frontier compute figures are Epoch estimates, many of them Speculative. Compute and ECI come from the same publisher: consistent, but not independent confirmation.
Story 6 · Money
The US–China funding gap is far wider than the capability gap
Money and capability have different units. The two charts sit side by side, never on one axis — the point is the ratio.
On capability, the best US model leads the best Chinese model by 10.97 ECI points, about 6.5% of the US score (Oct 2026).
- Shaded = under ~4 pts, within the 90% CI (noise)
The percentage divides an interval-scale difference by a level, so it is a rough comparison only.
Show the data table
| Month | Best US (ECI) | Best China (ECI) | Gap (pts) |
|---|---|---|---|
| 2023-06 | GPT-4 (Mar 2023) 125.74 | Baichuan1-7B 90 | 35.74 |
| 2023-07 | GPT-4 (Mar 2023) 125.74 | Baichuan1-7B 90 | 35.74 |
| 2023-08 | GPT-4 (Mar 2023) 125.74 | Baichuan1-7B 90 | 35.74 |
| 2023-09 | GPT-4 (Mar 2023) 125.74 | Qwen-14B 112.88 | 12.86 |
| 2023-10 | GPT-4 (Mar 2023) 125.74 | Qwen-14B 112.88 | 12.86 |
| 2023-11 | GPT-4 (Mar 2023) 125.74 | Yi-34B 117.33 | 8.41 |
| 2023-12 | GPT-4 (Mar 2023) 125.74 | Yi-34B 117.33 | 8.41 |
| 2024-01 | GPT-4 Turbo (Nov 2023) 126.43 | Yi-34B 117.33 | 9.1 |
| 2024-02 | Claude 3 Opus 126.9 | Yi-34B 117.33 | 9.57 |
| 2024-03 | Claude 3 Opus 126.9 | Yi-34B 117.33 | 9.57 |
| 2024-04 | GPT-4 Turbo (Apr 2024) 127.27 | Yi-34B 117.33 | 9.94 |
| 2024-05 | GPT-4o (May 2024) 128.99 | DeepSeek-V2 (MoE-236B, May 2024) 124.8 | 4.19 |
| 2024-06 | Claude 3.5 Sonnet 130 | Qwen2-72B 125.28 | 4.72 |
| 2024-07 | Claude 3.5 Sonnet 130 | Qwen2-72B 125.28 | 4.72 |
| 2024-08 | Claude 3.5 Sonnet 130 | Qwen2-72B 125.28 | 4.72 |
| 2024-09 | o1-mini 135.7 | Qwen2.5-72B 128.97 | 6.73 |
| 2024-10 | o1-mini 135.7 | Qwen2.5-72B 128.97 | 6.73 |
| 2024-11 | o1-mini 135.7 | Qwen2.5-72B 128.97 | 6.73 |
| 2024-12 | o1 141.72 | DeepSeek-V3 132.3 | 9.42 |
| 2025-01 | o1 141.72 | DeepSeek-R1 138.82 | 2.9 |
| 2025-02 | o1 141.72 | DeepSeek-R1 138.82 | 2.9 |
| 2025-03 | Gemini 2.5 Pro (Mar 2025) 143.96 | DeepSeek-R1 138.82 | 5.14 |
| 2025-04 | o3 146.73 | Qwen3-235B-A22B 139.15 | 7.58 |
| 2025-05 | o3 146.73 | DeepSeek-R1 (May 2025) 141.13 | 5.6 |
| 2025-06 | o3-pro 147.33 | DeepSeek-R1 (May 2025) 141.13 | 6.2 |
| 2025-07 | o3-pro 147.33 | Qwen3-235B-A22B-Thinking (Jul 2025) 143.76 | 3.57 |
| 2025-08 | GPT-5 150 | Qwen3-235B-A22B-Thinking (Jul 2025) 143.76 | 6.24 |
| 2025-09 | GPT-5 150 | DeepSeek-V3.2-Exp 144.91 | 5.09 |
| 2025-10 | GPT-5 Pro 150.59 | DeepSeek-V3.2-Exp 144.91 | 5.68 |
| 2025-11 | Gemini 3 Pro 153.07 | Kimi K2 Thinking 145.82 | 7.25 |
| 2025-12 | GPT-5.2 Pro 156.33 | DeepSeek-V3.2 146.09 | 10.24 |
| 2026-01 | GPT-5.2 Pro 156.33 | Kimi K2.5 148.07 | 8.26 |
| 2026-02 | GPT-5.3 Codex 157.09 | Kimi K2.5 148.07 | 9.02 |
| 2026-03 | GPT-5.4 Pro 159.57 | Kimi K2.5 148.07 | 11.5 |
| 2026-04 | GPT-5.5 Pro 162.91 | Kimi K2.6 150.94 | 11.97 |
| 2026-05 | GPT-5.5 Pro 162.91 | Qwen3.7-Max 154.04 | 8.87 |
| 2026-06 | Claude Fable 5 163.42 | Qwen3.7-Max 154.04 | 9.38 |
| 2026-07 | Claude Opus 5 163.94 | Kimi K3 157.86 | 6.08 |
| 2026-08 | Claude Opus 5 163.94 | Kimi K3 157.86 | 6.08 |
| 2026-09 | Claude Opus 5.5 168.83 | Kimi K3 157.86 | 10.97 |
| 2026-10 | Claude Opus 5.5 168.83 | Kimi K3 157.86 | 10.97 |
File: knowledge/data/derived/eci_best_us_vs_china_monthly.csv
OpenAI has raised $182.8B in equity and Anthropic $139.4B; the best-funded Chinese lab in the data, DeepSeek, $22.5B.
- United States
- China
- Other country
Cohere, Google, Meta, Moonshot AI have no figure, so no bar is drawn (not zero). DeepSeek's figure rests on press reports of two 2026 rounds and should be re-checked.
Show the data table
| Company | Country | Cumulative equity funding (USD) |
|---|---|---|
| OpenAI | United States of America | 182,750,000,000 |
| Anthropic | United States of America | 139,404,000,000 |
| xAI | United States of America | 37,000,000,000 |
| DeepSeek | China | 22,500,000,000 |
| Mistral AI | France | 6,494,300,000 |
| MiniMax | China | 619,000,000 |
| Z.ai (Zhipu) | China | 558,000,000 |
| Cohere | Canada | no figure |
| United States of America | no figure | |
| Meta | United States of America | no figure |
| Moonshot AI | China | no figure |
File: knowledge/data/derived/frontier_company_funding.csv
What it means for your industry
Capital is concentrated in a handful of labs. Industry buyers should expect pricing power and lock-in on the closed side, and treat a cheap open-weight alternative as a real option.
- Google and Meta have no figure: they are not standalone labs.
- The AI Index 2026 puts 2025 US private AI investment at $285.9B, 23× China's $12.4B (PDF p.11, data: Quid). The report warns these figures likely understate China's total AI spending, given its government guidance funds. Stanford HAI, AI Index 2026 (CC BY-ND 4.0 — quoted, not redrawn)
Story 7 · Inside companies
Data science is not gone: its weight differs by industry
In the EU enterprise survey, text generation overtook machine learning — yet industries such as energy, where forecasting and maintenance matter, still lean on ML.
In EU firms, text generation overtook machine learning in 2024 (5.41% vs 4.24%); outside ICT, energy uses ML most, at 13.07%.
- Machine learning for data analysis
- Natural-language generation (incl. code)
EU-27, % of enterprises with 10+ employees. Energy ML use (13.07%) is about 2.6× the all-sector 5.1%. Shipping sits inside “Transportation & storage” and cannot be separated.
Show the data table
| Activity (NACE) | ML 2023→2024→2025 | NLG 2023→2024→2025 | Any AI 2025 |
|---|---|---|---|
| All activities (excl. agriculture, mining, finance) (C10-S951_X_K) | 2.6% → 4.24% → 5.1% | 2.1% → 5.41% → 8.76% | 19.95% |
| Manufacturing (C) | 1.73% → 2.73% → 3.66% | 1.19% → 3.53% → 7.09% | 17.27% |
| Electricity, gas, steam (energy) (D35) | 8.51% → 12.59% → 13.07% | 4.2% → 8.91% → 17.17% | 33.57% |
| Transportation and storage (H) | 1.35% → 1.98% → 2.69% | 1.28% → 3.32% → 3.92% | 11.15% |
| Information and communication (J) | 16.28% → 25.66% → 28.58% | 11.14% → 25.83% → 42.23% | 62.52% |
File: knowledge/data/derived/eurostat_ai_tech_by_industry.csv
What it means for your industry
Data science has not gone; its weight differs by sector. This data does not show that classical methods outperform LLMs. Transport and storage is the lowest adopter shown here.
- This is EU data only; we found no comparable global, Korean or Gulf series. A firm can tick both technologies, and Eurostat does not say whether the ML is classical or deep learning.
How this page is made
- Every number comes from a row in knowledge/data/derived/, and the charts are drawn from those CSVs at build time — no outside requests, no chart library.
- Missing values stay missing — never drawn as zero.
- Counts and percentages start at zero. ECI and Arena ratings have no meaningful zero, so we plot the gap. Compute uses a log scale.
- The Stanford AI Index is CC BY-ND, so its figures are quoted with page numbers and never redrawn. Hugging Face data appears as aggregates only. Artificial Analysis is not used, because its terms forbid it.
- What this data cannot say: which models industries actually run, anything about shipping on its own (it sits inside “transportation & storage”), or the country of use (every country field is the developer's).