Stanford HAI released its 2026 AI Index Report on April 16, putting AI’s energy bill into terms familiar to crypto markets: global AI system electricity use is now close to the national power consumption of Switzerland or Austria, or about half of global Bitcoin mining power demand. The report also says the lead held by top US frontier models has narrowed sharply, with Anthropic’s best model ahead by only 2.7%.
According to the report, global data center electricity demand has reached about 47,000 MW, excluding crypto mining. Within that total, AI data centers had a combined power capacity of 29.6 GW by the end of 2025, a level comparable to peak electricity demand in New York State. The implication is plain: training and serving large models is becoming a major power load.
Data center buildout is pushing AI energy demand higher
The US still leads on infrastructure. Stanford’s report says the country has 5,427 data centers, more than 10 times the count of the second-ranked country, while AI hardware continues to take a larger share of overall data center investment. Using the report’s end-2025 reference point, computing expansion has not yet shown signs of easing.
That comparison matters because Bitcoin mining has long been used as the standard example in debates over electricity consumption. AI data centers are now moving into a similar range, and the report makes that contrast explicit.
Top Chinese models are closing in on US leaders
On model capability, the report presents a much tighter race. As of March 2026, Anthropic’s top model held only a 2.7% lead. The gap between Claude Opus 4.6 and ByteDance’s Dola-Seed-2.0 Preview had narrowed to just 39 Elo points. Stanford also notes that DeepSeek-R1 briefly reached parity with leading US models when it launched in February 2025, and this level of competition is no longer treated as a one-off result.
Anthropic, xAI, Google, OpenAI, Alibaba and DeepSeek all placed their flagship models in the same top performance tier under the report’s evaluation system. In practical terms, the US-China difference is no longer described as a generation-level gap but as a marginal one.
On output, the US produced 50 representative frontier models in 2025, while China produced 30. Among the top 10 representative models, Alibaba, DeepSeek, Tsinghua University and ByteDance all appeared on the list.
Capital spending still shows a huge split
Investment tells a different story. The report says private AI investment in the US reached $285.9 billion in 2025, versus $12.4 billion in China, a gap of roughly 23 times. The contrast is stark: model performance is converging even as funding levels remain far apart.
Patent and academic citation data point in another direction. China accounted for 74.2% of global AI patents, while the US held 12.1%. In 2024 AI paper citations, China contributed 20.6%, Europe 19.5%, and the US 12.6%, placing the US third on that metric.
Adoption is rising fast, and incident counts are rising too
Use of AI continues to spread quickly. The report says organizational AI adoption has reached 88%, and generative AI hit a 53% population adoption rate within three years, faster than previous technology waves tracked in the study.
At the same time, recorded AI incidents climbed from 233 in 2024 to 362, including hallucinations, bias, and security flaws. Talent flows are also shifting: AI talent moving into the US has fallen 89% since 2017. On the hardware side, TSMC still manufactures nearly all leading-edge AI chips, leaving supply risk heavily concentrated around the Taiwan Strait.