AI's environmental footprint
Grok 4's training emitted ~72,816 tons of CO₂e; AI data center power hit 29.6 GW.
The inputs to AI keep scaling — but they're concentrating in fewer hands
Chapter 1 of the AI Index 2026 follows the R&D pipeline — from the models themselves through compute, data centers, energy, and open source, out to publications, patents, and talent. Resources are growing; transparency, geographic balance, and gender balance are not. The numbers:
59U.S. notable AI models in 2025 (China 35)
91% of notable AI models now from industry
17million H100-equivalents of global AI compute
5,427U.S. data centers — 10× any other country
131,121AI patents granted worldwide in 2024
89% drop in AI talent moving to the U.S. since 2017
1.1 — Notable models: fewer, more closed, more concentrated
Using Epoch AI's curated dataset, the chapter maps where frontier models come from. In 2025, fewer notable models were released than the year before — and the most capable ones are now the least transparent.
The United States led with 59 notable AI models in 2025, followed by China with 35 and South Korea with 8. New releases declined year over year across every major region. Production stays concentrated: industry now accounts for 91.2% of notable models, with just 2 of 2025's releases coming from academia versus 93 from industry. Within industry, the top contributors were OpenAI (20), Google (14), and Alibaba (11).
The most capable models are the least transparent
- Training code, parameter counts, dataset sizes, and training duration are no longer disclosed for several of the most resource-intensive systems — including those from OpenAI, Anthropic, and Google.
- In 2025, 81 of 102 notable models were released without their training code, versus just 4 released as 'open source.' API access was the single most common release type, at 47 of 102 models.
- Reported parameter counts have stayed near 1 trillion for three years, even as frontier labs stop reporting. Training compute — which can be estimated independently — has kept rising.
Data quality is beating data scale
Synthetic data still isn't replacing real data in pre-training, but data-centric methods are showing promise. OLMo 3.1 Think 32B — with roughly 32 billion parameters, nearly 90× fewer than Grok 4's 3 trillion — reaches comparable results on benchmarks like AIME 2025 through pruning, deduplication, and curation alone. Meanwhile, Graphite estimates that since January 2025, over 50% (51.7%) of newly published online content has been AI-generated.
1.2 — Compute is up 3.3× a year — and it runs through one island
Training compute would be impossible without ever-faster hardware. Aggregate capacity is soaring, but the supply chain has become a single point of failure.
Global AI compute capacity has grown roughly 3.3× per year since 2022, reaching about 17.1 million H100-equivalents by the end of 2025. Nvidia AI chips account for over 60% of total compute, with Google and Amazon supplying much of the rest and Huawei holding a small but growing share. The buildout is driven by hyperscaler data center expansion and sustained demand for frontier training and inference.
One foundry in Taiwan fabricates almost every leading AI chip
- Companies like Nvidia and SK Hynix design chips but don't make them — fabrication is dominated by a single foundry, TSMC, which produces virtually every leading AI chip, including Nvidia's Blackwell GPUs and AMD's MI300X.
- That makes the global hardware supply chain dependent on one foundry in Taiwan, though a TSMC–U.S. expansion began operating in 2025.
- Beyond GPUs, the stack relies on high-bandwidth memory from SK Hynix, Samsung (South Korea), and Micron (USA), plus InfiniBand networking and assembly by firms like ASE Group (Taiwan) and Amkor (USA).
The economics have moved the other way: since 2006, the cost of GPU computation has fallen by more than 99%, which is precisely what made today's scaling economically feasible.
The R&D landscape, in seven moves
From energy and open source through publications, patents, and talent. Tap any card for the full trend and its numbers.
Open source keeps scaling
5.6M AI projects on GitHub; U.S. projects hold 30M cumulative stars.
China leads in publication volume
AI publications hit ~258,000 in 2024; China holds 17.8% of volume and 20.6% of citations.
The top-100 gap is narrowing
China's share of the top-100 most-cited AI papers rose from 33 in 2021 to 41 in 2024.
Patents: China's volume, U.S. influence
131,121 AI patents granted in 2024; China holds 74.2%, but the U.S. earns over half of forward citations.
Talent is no longer flowing to the U.S.
AI talent moving to the U.S. has dropped 89% since 2017 — down 80% in the last year alone.
Gender gaps remain entrenched
No country approaches parity; Saudi Arabia leads female share at 32.3%, with no real progress since 2010.
The chapter in five lines
Headline findings from Chapter 1 · Research & Development.
Industry produced over 90% of notable AI models in 2025 — but the most capable models are now the least transparent.
Global AI compute capacity has grown 3.3× per year since 2022, reaching 17.1 million H100-equivalents.
A single company, TSMC, fabricates almost every leading AI chip, making the global hardware supply chain dependent on one foundry in Taiwan.
China leads in publication volume, citations, and patent grants, while the U.S. produced 59 notable models in 2025 to China's 35.
The number of AI researchers moving to the United States has dropped 89% since 2017, and gender gaps in AI talent have not improved since 2010.
Read the full Research & Development chapter
Chapter 1 (sections 1.1–1.8) with every figure and citation is free from Stanford HAI. Or head back to the 15 takeaways and nine-chapter overview.
Open Chapter 1 · Research & Development →