AI Index Report 2025
arrow_back Overview open_in_new Stanford HAI translate 中

In 2024, AI research got bigger, costlier — and almost entirely corporate

Chapter 1 of the AI Index 2025 follows where AI knowledge is made: papers, patents, notable models, the chips underneath them, and the energy they burn. The pattern is consistent — output keeps climbing, but the frontier is moving into a smaller number of very well-funded hands. The numbers:

90% of notable AI models that came from industry in 2024 (60% in 2023)
40notable AI models from US institutions in 2024 (China 15, Europe 3)
242thousand AI publications in computer science in 2023 (102 thousand in 2013)
122,511AI patents granted worldwide in 2023, up from 3,833 in 2010
280-fold drop in the cost of GPT-3.5-level inference in about 18 months
8,930tons of CO₂ emitted to train Llama 3.1 405B (GPT-3 emitted 588)

1.1 — More AI papers than ever, and China writes the most of them

Between 2013 and 2023, AI publications in computer science venues more than doubled, from roughly 102,000 to over 242,000 — a 19.7% rise in the last year alone. AI now accounts for 41.8% of all computer science publications, up from 21.6% a decade earlier.

The growth is not just more AI researchers writing more AI papers. Fields across computer science — hardware, software engineering, human-computer interaction — are now contributing to AI, so the rise reflects a broader migration of the discipline toward AI methods. Journals carried the largest share of AI publications in 2023 (41.8%), followed by conferences (34.3%), while preprint repositories like arXiv keep gaining ground.

Where the papers come from

  • China led the world in AI article publications in 2023 with 23.2% of the total, ahead of Europe (15.2%) and India (9.2%). China also led on citations, taking 22.6% of all AI publication citations, ahead of Europe (20.9%) and the United States (13.0%).
  • By region, East Asia and the Pacific produced 34.5% of AI publications and attracted 37.1% of citations — a share that has risen sharply since 2017, when it was roughly level with North America.
  • Academic institutions remain the engine room, producing 84.9% of AI publications in 2023, versus 7.1% from industry, 4.9% from government and 1.7% from nonprofits.
  • The mix differs sharply by country: 16.5% of US AI publications come from industry, against 8.0% in China, where 84.5% originate in the education sector.
  • By topic, machine learning appeared in 75.7% of 2023 AI publications, followed by computer vision (47.2%), pattern recognition (25.9%) and natural language processing (17.1%), with a sharp jump in generative AI papers over the past year.

The most-cited papers tell a different story

New this year, the AI Index also tracked the 100 most-cited AI publications of 2021, 2022 and 2023. The United States topped that list every year — 64 papers in 2021, 59 in 2022, 50 in 2023 — with China second each time, though the US share has been gradually declining. Academia produced the most top-cited work (42 papers in 2023), while industry's contribution collapsed from 17 in 2021 and 19 in 2022 to just 7 in 2023: as competition intensified, corporate labs began publishing less often and disclosing less when they do. Google led each year but tied with Tsinghua University in 2023, both contributing eight papers to the top 100.

East Asia and the Pacific now produce a third of the world's AI research

AI publications in computer science as a share of the world total, 2023 (%). East Asia and the Pacific also took 37.1% of all AI publication citations that year.

East Asia and the Pacific now produce a third of the world's AI researchEast Asia & Pacific: 34.534.5East Asia & PacificEurope & Central Asia: 18.218.2Europe & Central AsiaNorth America: 10.310.3North AmericaMiddle East & N. Africa: 5.25.2Middle East & N. Africa

1.2 & 1.3 — China owns the patents, the US ships the models

Two very different maps of AI capability. On granted patents, China holds 69.7% of the world total. On notable models — the systems that actually move the frontier — the United States produced 40 in 2024, against China's 15 and Europe's combined three.

Patents: growth, and a widening geographic gap

  • Granted AI patents worldwide grew from 3,833 in 2010 to 122,511 in 2023 — a 29.6% rise in the last year alone.
  • As of 2023, 82.4% of the world's granted AI patents originated in East Asia and the Pacific, with North America the next largest contributor at 14.2%. The gap has widened steadily since 2010.
  • By country, China accounts for 69.7% of grants and the United States for 14.2% — down from a US peak of 42.8% in 2015.
  • Per capita, the picture changes: South Korea led in 2023 with 17.3 granted AI patents per 100,000 inhabitants, followed by Luxembourg (15.3) and China (6.1). Sweden recorded the greatest increase in AI patenting over the decade.

Notable models: fewer of them, and almost none from universities

In 2024 the United States led with 40 notable AI models, followed by China with 15 and France with three. Every major geographic group released fewer notable models than the year before — plausibly a consequence of ever-larger training runs, growing technical complexity, and the difficulty of finding genuinely new modeling approaches. The sectoral split is starker still: Epoch AI counted 55 notable models from industry in 2024 and identified none at all from academia, pushing industry's share to 90.2%, up from about 60% in 2023.

  • The top model producers in 2024 were Google (7), OpenAI (7) and Alibaba (6). Since 2014 the cumulative leaders are Google (187), Meta (82) and Microsoft (39).
  • Among academic institutions since 2014, Carnegie Mellon (25), Stanford (25) and Tsinghua (22) have been the most prolific.
  • API access was the most common release type in 2024, covering 20 of 61 models — 32.8% of releases, a share that has risen steadily since 2020.
  • Openness stops short of code: 60.7% of notable models released in 2024 came without corresponding training code, even when weights were published.
  • A count of zero academic models needs care: it means Epoch AI identified none as notable, not that universities stopped building models. Academic work also takes longer to be recognized.

A decade of notable models, and who built them

Cumulative number of notable AI models by organization, 2014–24. Google alone accounts for more than twice as many as the next contributor; the leading universities sit an order of magnitude behind.

A decade of notable models, and who built themGoogle: 187187GoogleMeta: 8282MetaMicrosoft: 3939MicrosoftCarnegie Mellon: 2525Carnegie MellonStanford: 2525StanfordTsinghua: 2222Tsinghua

What a frontier model costs, from AlexNet to DeepSeek-V3

Training compute for notable models doubles roughly every five months, LLM dataset sizes every eight months, and training power annually. Here is what that compounding looks like in dollars, watts and carbon.

  1. 2012 · AlexNet

    Five days of training, and negligible emissions

    One of the first models to use GPUs for training, AlexNet finished in roughly five to six days on hardware that would look primitive today. Its estimated training emissions were about 0.01 tons of CO₂.

  2. 2017 · Transformer

    $670 to train the architecture behind every modern LLM

    The original Transformer was trained on roughly 2 billion tokens, required around 7,400 petaFLOP, drew an estimated 4,500 watts, and cost about $670 to train in inflation-adjusted terms.

  3. 2019 · RoBERTa Large

    Six figures, and state of the art on comprehension

    RoBERTa Large, which posted state-of-the-art results on canonical benchmarks such as SQuAD and GLUE, cost roughly $160,000 to train — a 240-fold jump over the Transformer in two years.

  4. 2020 · GPT-3 175B

    374 billion tokens, and 588 tons of carbon

    One of the models underpinning the original ChatGPT was trained on an estimated 374 billion tokens and reportedly emitted around 588 tons of CO₂ during training — roughly 33 times what an average American emits in a year.

  5. 2023 · GPT-4

    Around $79 million, and 5,184 tons of carbon

    The AI Index estimates GPT-4's training cost at around $79 million based on cloud compute rental prices; OpenAI's Sam Altman has said training exceeded $100 million. Estimated training emissions were 5,184 tons of CO₂.

  6. 2024 · Llama 3.1 405B

    90 days, 25.3 million watts, $170 million

    Meta's flagship took roughly 90 days to train — a typical window by today's standards — at an estimated $170 million and a power draw of 25.3 million watts, over 5,000 times the original Transformer. Its training emitted 8,930 tons of CO₂. Llama 3.3, released the same year, was trained on roughly 15 trillion tokens.

  7. Dec 2024 · DeepSeek-V3

    High performance, reportedly for about $6 million

    DeepSeek's V3 drew attention for achieving strong results with far fewer computational resources; its reported training cost was about $6 million, and its emissions are estimated to be comparable to GPT-3 five years earlier. Some reports dispute that figure once salaries, capex and research expenses are counted. Epoch AI finds China's top 10 models by training compute have scaled about 3× per year since late 2021, against roughly 5× per year elsewhere since 2018.

1.4–1.6 — The chips get more efficient, the models eat the gains

Machine learning hardware has improved on every axis that matters: 43% more performance a year, 30% cheaper a year, 40% more energy efficient a year. And yet the power required to train a frontier model is doubling annually.

Faster, cheaper, more efficient

  • Measured in 16-bit floating-point operations, machine learning hardware performance grew about 43% annually from 2008 to 2024, doubling every 1.9 years — driven by transistor counts, semiconductor manufacturing, and AI-specific silicon.
  • Price performance improves about 30% a year. Nvidia's H100, announced in March 2022, delivers 22 billion FLOP per second per dollar — roughly 1.7 times the A100 and 16.9 times the P100 of 2016.
  • Energy efficiency improves about 40% a year. The B100, released in March 2024, reaches 2.5 trillion FLOP/s per watt, against 74 billion for the 2016 P100 — 33.8 times better.
  • The A100 remains the most commonly reported training chip, used by 64 notable models, with the H100 climbing fast at 15 models by the end of 2024.

And yet total power keeps doubling

Efficiency gains have been swallowed whole by scale. The original Transformer drew an estimated 4,500 watts; PaLM, one of Google's first flagship LLMs, drew 2.6 million watts — almost 600 times as much. Llama 3.1 405B required 25.3 million watts, over 5,000 times the Transformer. Epoch AI finds the power needed to train frontier models is doubling every year, and carbon follows: AlexNet's emissions were negligible, GPT-3 emitted around 588 tons, GPT-4 around 5,184 tons, and Llama 3.1 405B around 8,930 tons. For scale, the average American emits 18.08 tons of carbon a year, and a single passenger flying New York to San Francisco and back accounts for 0.99 tons.

The community around the models

  • AI conference attendance rose 21.7% from 2023 to 2024, and has grown by more than 60,000 attendees since 2014. NeurIPS remains the largest, drawing almost 20,000 participants in 2024.
  • GitHub AI projects grew from 1,549 in 2011 to roughly 4.3 million in 2024 — a 40.3% rise in the last year alone.
  • US-based developers accounted for 23.4% of GitHub AI projects in 2024, with India close behind at 19.9% and Europe at 19.5%. The US share has been declining since 2016 and appears to have stabilized.
  • New GitHub stars for AI projects rose from 14.0 million in 2023 to 17.7 million in 2024, with US-based projects receiving 21.1 million stars in total.

Five questions the chapter actually answers

The open arguments in AI research, with what the 2025 data says about each.

Is AI about to run out of training data?
Not immediately, but the window is visible. Epoch AI estimates Common Crawl holds a median of 130 trillion tokens, the indexed web roughly 510 trillion, and the entire web around 3,100 trillion; images add about 300 trillion and video about 1,350 trillion. With an 80% confidence interval, Epoch projects the current stock of training data will be fully utilized somewhere between 2026 and 2032. That is later than last year's forecast of high-quality text running out by 2024, because new research showed that filtered web data works better than curated corpora and that models can be trained on the same data multiple times. Overtraining — pushing past the point of diminishing returns to gain cheaper inference — depletes the stock faster.
Can synthetic data fill the gap?
Partly, and with caveats. Early research warned of model collapse: repeatedly training on synthetic data causes models to lose the tails of distributions and degrade output quality, observed across VAEs, Gaussian mixture models and LLMs. Newer work found that if synthetic data is layered on top of real data rather than replacing it, collapse does not occur — though accumulation does not necessarily improve performance either. A Slovenian team found most synthetic relational data is still systematically detectable and generally underperforms real data, with occasional exceptions. There are real wins in specific domains: synthetically augmented healthcare datasets raised F1 or AUROC scores by 5%–10% on minority classes, and Stanford and UNC Chapel Hill researchers used automated fact-checking to build FactTune-FS, which outperformed other RLHF and decoding-based methods on factuality.
Why has using a model got so cheap?
Because smaller models keep catching up to yesterday's frontier. Holding performance fixed, the cost of querying a model scoring GPT-3.5's 64.8 on MMLU fell from $20.00 per million tokens in November 2022 to $0.07 by October 2024 with Gemini-1.5-Flash-8B — a more than 280-fold reduction in roughly 18 months. The same pattern shows up on harder tests: for models scoring above 50% on GPQA, inference fell from $15 per million tokens in May 2024 to $0.12 by December 2024 with Phi-4. Epoch AI estimates that, depending on the task, LLM inference prices have been falling anywhere from 9 to 900 times per year. State-of-the-art models remain priced at a premium — the point is that yesterday's state of the art becomes nearly free.
Who actually builds the models now?
Industry, almost exclusively. Nearly 90% of notable models in 2024 came from industry, up from about 60% in 2023, and Epoch AI identified no notable academic models at all that year. Creating a cutting-edge model now demands data, compute and money at a scale universities do not have. But the picture inverts on influence rather than output: academia produced 42 of the 100 most-cited AI papers in 2023, while industry's share of that list fell from 19 papers in 2022 to just 7 in 2023 — corporate labs are publishing less, and disclosing less when they do.
How much does training cost the climate?
It is rising steadily, and the reference points are becoming uncomfortable. AlexNet's 2012 training run emitted about 0.01 tons of CO₂. GPT-3 in 2020 emitted around 588 tons, GPT-4 in 2023 around 5,184 tons, and Llama 3.1 405B in 2024 around 8,930 tons. The average American emits 18.08 tons of carbon per year, so Llama 3.1 405B's training run is roughly equivalent to 494 Americans' annual emissions. Notably, DeepSeek V3 — comparable in performance to OpenAI's o1 — is estimated to have emissions in the range of GPT-3, released five years earlier, which suggests efficiency is still available if it is prioritized.

Read the full Research and Development chapter

Chapter 1 (sections 1.1–1.6) — publications, patents, notable models, hardware, conferences and open-source software — with every figure and citation is free from Stanford HAI.

Open the AI Index Report 2025 →

中文版