Explainer

Who is ahead in AI

The File desk · fetched filings and named boards

Public numbers that can be compared. Estimates are marked. We only chart when the unit is the same. A missing figure is omitted, not drawn as zero.

Compiled Wednesday, September 2, 2026

Updated Monday, October 12, 2026

Energy

TWh · Ember yearly electricity · through 2025 · Reported

This is electricity made each year, not chips. China’s line roughly doubled. The U.S. line stayed almost flat, then ticked up. India rose, still far below both. 2025 is Reported (Ember yearly).

More detail

Ember Yearly Electricity Data (Total generation CSV). OWID electricity-generation extract is a secondary check. 2015–2025 Reported.

Other Ember 2025 Reported figures, not on the chart: Russia 1,194 TWh; Japan 1,027 TWh; Brazil 764 TWh; Canada 660 TWh; South Korea 628 TWh.

Do not mix Energy Institute 2025 onto these lines. Energy Institute U.S. 4,772 terawatt-hours is gross, about 250 terawatt-hours above Ember’s U.S. 4,520. Energy Institute China 10,575; India 2,055.

Energy Information Administration check, not a second line: Electric Power Monthly U.S. 2024 4,309 terawatt-hours final; 2025 about 4,430 terawatt-hours preliminary.

Data-centre use, different unit: Energy Department / Lawrence Berkeley National Laboratory U.S. 58 terawatt-hours in 2014 to 176 terawatt-hours in 2023 (Reported). International Energy Agency Estimate, world data centres: about 415 terawatt-hours in 2024; Key Questions Annex A.4, 416 in 2024, 485 in 2025, 945 in 2030 Base Case. Energy and AI also printed about 950 in 2030. Energy Department hub Estimate: 11.8% of U.S. electricity by 2030 (9.5–15.3%).

Compute

TWh · IEA data-centre electricity · 2024 Est. · 2025 Est. · 2030 Base Case proj.

This is electricity used by data centres, not how many chips anyone owns. The U.S. line starts almost twice China’s. The International Energy Agency thinks both rise by 2030. China’s climbs faster in percent. The United States still has the higher bar. 2024 and 2025 are modelled Estimates from Annex A.4. 2030 is their Base Case, not a measurement. Years 2026 through 2029 are not on that table, so they are not drawn.

World 416 terawatt-hours in 2024 on Key Questions Annex A.4 (Energy and AI printed about 415) — caption, not a fourth line. Same annex: 485 in 2025 to 945 in 2030 Base Case (Energy and AI printed about 950). Years 2026–2029 are not on that regional table, so they are not drawn.

More detail

Companies — money for chips and buildings, 2026

Billions of dollars · 0 to the longest bar

This is money these companies plan to spend this year on buildings and machines. Amazon’s bar is tens of billions longer. They do not count every dollar the same way.

Energy Department / Lawrence Berkeley is a different U.S.-only study — not on the International Energy Agency lines: 58 terawatt-hours in 2014 to 176 terawatt-hours in 2023.

Accelerators are not installed stock: Amazon plans more than 1 million NVIDIA GPUs starting 2026 and 2 million more in 2027–28 (Plan); Anthropic has access to more than 220,000 at Colossus 1 (Access); Nvidia’s $89.0 billion Data Center quarter is sold chips.

TSMC second quarter $40.20 billion; 77% of wafer revenue at 7-nanometer and below; SK hynix high-bandwidth memory (HBM4) mass shipments in the second quarter of 2026. Factory dollars and memory shipments are not one scale.

Power contracts, not one scale: Anthropic Amazon up to 5 gigawatts and Google 5 gigawatts of tensor processors from 2027; Microsoft +1 gigawatt this quarter; Oracle more than 1.2 gigawatts delivered in fiscal 2026 (Partially confirmed).

International Energy Agency Estimate, extra data-centre electricity 2024–30: United States about +240 terawatt-hours, China about +175 terawatt-hours, Europe more than +45 terawatt-hours.

Cloud share, Synergy Q2 Estimate: AWS 28, Azure 20, Google Cloud 15. Same-quarter dollars: AWS $42.2 billion vs Google Cloud $24.8 billion, up 82%.

Models

Hard public test (Terminal-Bench 4.0)

Percent of 66 tasks · 0 to the longest bar

Anthropic Opus 5.5 — United States64.8%src
Anthropic Sonnet 5.5 — United States61.8%src
OpenAI GPT-6 Astra — United States58.2%src
OpenAI GPT-6.1 Sol — United States58.2%src
Zhipu GLM-5.3 — China41.8%src

A hard test of 66 computer tasks. Anthropic’s Opus 5.5 now leads this independent Snorkel board. Sonnet 5.5 is second. OpenAI’s Astra and Sol 6.1 tie for third. China’s GLM-5.3 still trails the U.S. leaders but stays on the chart. Longer bar = more tasks solved.

More detail

Opus 5.5 is 64.8% on this independent board. Anthropic’s own post says 66.4%. Sonnet 5.5 is 61.8% here. Anthropic’s own post says 70.6%. Company figures are not on the bars.

Fable 5.1 (57.9%) and Opus 5 (53.9%) are now off the first-screen bars, same board. Fable 5.1 is on this independent Snorkel board at 57.9%. That is not Anthropic’s own 55.8% system-card figure. Mythos 5.1 60.9% is Anthropic-reported (Unverified here) and is not on the independent Snorkel bars.

Grok 4.7 is 37.6% on the same Snorkel Terminal-Bench 4.0 board. GPT-6 Sol 49.4%, GPT-5.6 Sol 37.3%, Grok 4.6 20.3%, Gemini 3.8 Flash 19.1% (Google, United States), and Alibaba Qwen3.8-Max (China) 27% are still on that board. Those scores are off the first-screen bars to keep the chart short. Gemini 4 Argon is not on the board yet.

API list price (input $ per million tokens), not training cost: DeepSeek V4-Pro $1.32; Grok 4.6 $2; Kimi K3 $3; Sol promo $4 through Nov. 21, 2026; Fable 5.1 $10. Gemini list prices did not load.

Snorkel full-run bill, not training: Sonnet 5.5 $7.3k; Fable 5.1 $6.2k; Opus 5 $6.1k; Opus 5.5 $4.7k; Grok 4.7 $3.7k; Astra $3.3k; GLM-5.3 $2.7k; GPT-6.1 Sol $0.6k (same board).

Usage, not capability: Gemini App 950 million monthly users in Q2 2026. OpenAI user posts did not load.