Artificial intelligence: how big is the boom?

1. How much is being invested

The boom shows first in the accounts of the four largest cloud operators. Their purchases of equipment (data centres, servers, chips) reached $202bn in Q3 2026, more than double a year earlier. In 2025 they had already spent $376bn.

Private investors follow the same slope: $345bn invested in AI companies in 2025, against $151bn in 2024, of which $171bn in generative AI. The United States takes $286bn, Europe $20.9bn, China $12.4bn.

2. Chips: the bottleneck

AI is measured in computing power. Epoch AI converts every chip sold into « H100 equivalents », NVIDIA’s reference chip. Since 2022, 31.6M H100 equivalents have been delivered, 6.3M of them in Q2 2026 alone (up 116% on a year earlier).

NVIDIA supplies 62%; in-house chips from Google (TPU), Amazon and AMD make up the rest. Chinese designers (Huawei, Cambricon) account for only 3% of deliveries, held back by US export controls. On the revenue side, NVIDIA’s data-centre business reached $89bn in Q2 2026.

3. Data centres: where construction stands

Three measures complement each other. The number of large hyperscale data centres counted by Synergy rose from 1,136 at end-2024 to 1,360 at end-2025; they hold 48% of world capacity, and nearly 800 more are in the pipeline. Their capacity grew 36% in a year.

The installed power of the largest AI-dedicated sites, tracked one by one by Epoch AI, reaches 14 GW in Q3 2026, against 4.6 GW a year earlier; their projects announce 38.2 GW by end-2028. The United States hosts 86% of it.

Finally, the only official statistic: the US Census Bureau measures data-centre construction spending. It runs at $96.5bn a year in Sep. 2026, 12.2% of all private non-residential construction, against $1.6bn in early 2014. Building one MW now costs $10.7M per MW (excluding equipment), against $7.7M per MW in 2020 according to JLL.

• 1,360large hyperscale data centres at end-20251,136 at end-2024

  Source:Synergy Research GroupSource type : Private

• 67%of world capacity held by hyperscalers in 2031 (forecast)48% at end-2025

  Source:Synergy Research GroupSource type : Private

• 17%of hyperscale capacity in Northern Virginia and around Beijing

  Source:Synergy Research GroupSource type : Private

• 14%of Americans support a data centre in their communityJLL survey, August 2026

  Source:JLLSource type : Private

4. Electricity

According to the International Energy Agency, data centres used 485 TWh in 2025, slightly more than 1.5% of world electricity, after 415 TWh in 2024. The IEA expects a doubling by 2030, to 950 TWh (about 3% of world demand), driven by AI-focused data centres, whose consumption would triple.

The geography is highly concentrated: the United States accounts for about half of installed capacity worldwide and of announced additions. Four hyperscalers alone used 110 TWh in 2024. In Europe, the project pipeline is 2.3 times current capacity.

• 465 TWhused by AI-focused data centres in 2030 (IEA projection)

  Source:International Energy Agency (IEA)

• 110 TWhused by four hyperscalers in 2024

  Source:International Energy Agency (IEA)

• ×2.3European Union: project pipeline relative to installed capacity

  Source:International Energy Agency (IEA)

5. Models: a US-China contest

Each year Epoch AI lists « notable » models (state-of-the-art results, wide adoption, historical significance). In 2025, 107 were released: 64 by US organisations, 35 by Chinese ones. Europe and the rest of the world share the remainder.

On usage, Chinese models, often open and cheap, have taken the lead on OpenRouter, a platform that routes developers’ requests to hundreds of models: DeepSeek, Z.ai and Tencent account for 56% of tokens in the week of Nov. 2, 2026, against 7.8% a year earlier. OpenRouter reflects neither companies’ use under their own contracts nor consumer apps.

6. The cost of tokens

The price of a given level of performance is collapsing. Reaching GPT-4’s level on PhD-level science questions cost $37.5 per million tokens in March 2023 (GPT-4); $0.175 in February 2025 with Gemini 2.0 Flash, 214 times less.

According to Epoch AI, this price falls about 40-fold a year, and 9- to 900-fold depending on the task (median: 50-fold, 200-fold for declines measured since 2024). But each task uses more and more tokens (reasoning, agents): the total bill rises, hence the investment of section 1.

• ÷40 a yearyearly price decline for GPT-4’s level (PhD-level questions)

  Source:Epoch AISource type : Research

• ÷50 a yearmedian yearly decline, all tasks (9- to 900-fold)

  Source:Epoch AISource type : Research

• ÷200 a yearmedian yearly decline measured since 2024

  Source:Epoch AISource type : Research

7. Usage

There is no worldwide meter of AI use; companies publish their own indicators. Google’s models process 32bn tokens a minute through their APIs in Q3 2026, against 7bn a year earlier. The Gemini app claims 1bn users, against over 650M monthly users a year earlier.

Official statistics show slower diffusion across the economy: in the United States, 24% of businesses reported using AI in the previous two weeks (Census Bureau BTOS survey, period ending Oct. 18, 2026).

8. Revenue: who is making money?

For now, AI money goes first to infrastructure suppliers. NVIDIA sold $278bn of data-centre hardware over four quarters. The clouds follow: Google Cloud revenue grew 103% year on year in Q3 2026.

Model labs are only starting to publish accounts. Anthropic’s IPO prospectus, filed with the SEC, shows revenue rising from $41M in 2023 to $4.6bn in 2025, then $17bn in Q3 2026 alone. Its compute and infrastructure expenses reached $9.9bn that quarter. OpenAI publishes no accounts: its revenue figures could not be checked against a primary source.