The AI boom in numbers

Training compute doubles every 5.7 months

Each dot is one of the 543 notable AI models for which Epoch AI has an estimate of training compute: the number of floating-point operations (FLOP) — additions and multiplications — that went into training it. They run from Theseus, Claude Shannon's maze-solving mechanical mouse of 1950, trained with 40 FLOP, to the largest so far, GPT-6 Astra by OpenAI (Sep 2026), at 9.4×10²⁶ FLOP.

The scale is logarithmic, so a straight line means steady exponential growth. A least-squares line through the 481 models published since 2010 rises ×4.3 a year: training compute doubled every 5.7 months, and ten years along the line multiply it 2.15M times. Through the 62 models before 2010 it rose ×1.5 a year, a doubling every 20.5 months. Frontier models — the five largest by compute when they came out — doubled every 5.4 months since 2010 (72 models).

The numbers are estimates. Epoch rates 59% of them confident (right within 3 times either way), 28% likely (within 10 times) and 12% speculative (within about 31 times); speculative ones are hollow dots. Without them, the trend since 2010 is a doubling every 5.6 months.

The largest model of each year

A new compute record was set 24 times since 2010. The largest model of 2010, RNN LM, was trained with 5.4×10¹⁶ FLOP; GPT-6 Astra used 17.4B times as much.

Epoch AI also estimates the cost of the training compute for 178 of the models, in 2023 dollars. The highest estimate is $388M, for Grok 4 (xAI, 2025); over $100M: 3 models. The most power in the data, about 233 megawatts, was drawn by the 100,000 chips that trained GPT-6 Astra, by Epoch's estimate.

The 10²³ FLOP club

Epoch AI keeps a list of large-scale models: those trained with over 10²³ FLOP, a fixed line some AI regulations use. It has a compute estimate for 326 of them. The first was AlphaStar in 2019; in 2025, 104 models crossed the line, after 101 in 2024.

The first models above the higher lines were Wu Dao 2.0 over 10²⁴ (2021), GPT-4 over 10²⁵ (2023) and Grok 3 over 10²⁶ (2025). Models over 10²⁵ FLOP published in 2025: 9. Of all 104 in the club that year, 90 were released with open weights — anyone can download and run them — and 10 only through an API or a hosted service; 4 were not released.

2026 is not over: to October 28, 2026 the list has 27 models with an estimate. Epoch adds new models within weeks and compute estimates later, so recent months fill in over time. It also lists 208 large-scale models without a compute estimate; they are not counted here.

Who builds them

Organizations in the United States worked on 135 of the 326 models over 10²³ FLOP and organizations in China on 139; next come South Korea (17), France (10) and United Arab Emirates (8). A model built by several organizations counts once for each of their countries.

The most came from Alibaba (36), Meta AI (24) and Nvidia (20). Groups are named as in the data, so Google's are counted apart: Google, Google DeepMind, Google Research, DeepMind and Google Brain.

Companies, not universities

Of the 109 notable models published in 2025, 92% came from companies alone, 5% from companies working with universities and 2% from universities alone. In 2010–2014, companies alone made 19% of them.

The money: Big Tech's capital spending

Training and running AI models takes data centers full of chips, and the companies that build the largest ones report the bill as capital expenditure: cash paid for property and equipment. Microsoft, Alphabet, Amazon, Meta and Oracle together spent $691B in the four quarters to September 2026, 4.5 times as much as the $154B of the four quarters to September 2022 — the last quarter before ChatGPT was released on November 30, 2022. That is $1.89B a day, or $21,900 every second.

Measured against their size, the spending took 36% of the five companies' revenue in the latest four quarters, against 13% before, and 95% of the cash their operations brought in, against 54%. The largest spender was Amazon, with $215B.

The filings do not say how much of this went to AI. Capital expenditure also pays for offices and other buildings — at Amazon, warehouses and a delivery network too — and each company reports it under its own definitions.

NVIDIA: selling the shovels

Epoch AI records the training hardware of 278 notable models published since 2020: 204 of them (73%) were trained on NVIDIA chips and 68 (24%) on Google's TPUs.

NVIDIA's fiscal year ends in late January: fiscal 2026 ran to January 25, 2026. Its revenue that year was $216B, 8.01 times as much as in fiscal 2023 ($27B), the year that ended 2 months after ChatGPT's release. Net income went from $4.37B (16% of revenue) to $120B (56%).

The latest quarter in the filings, Q2 FY2027 (to July 26, 2026), brought revenue of $96.2B and net income of $59.7B; 75% of revenue was left after the cost of what was sold (gross margin). The four quarters to July 26, 2026 add up to $303B of revenue and $193B of net income — at that pace, net income equal to NVIDIA's whole revenue of fiscal 2016 ($5.01B) came in every 9.5 days.

The SEC data used here give NVIDIA's total revenue, not its split by market, so the chart shows the total. The market value of NVIDIA shares held by the public — the public float on the cover of its annual report, not the market capitalization — was $434B at the end of July 2022 and $4.0T at the end of July 2025.