On Monday, Stanford's Human-Centered AI institute dropped its 2026 AI Index — the annual 400-page document that the entire industry pretends not to read but actually memorizes line by line. This year's version has a finding that should ruin a few quarterly earnings calls.

The US spent $109 billion on private AI investment in 2024. China spent $9.3 billion. That is a 12x gap in capital. The performance gap between the best American model and the best Chinese model? 2.7 percentage points.

Let that sit for a second.

What the numbers actually say

On MMLU — one of the benchmarks that venture capital slide decks have been waving around since 2023 — the gap between top US and Chinese models collapsed from 17.5% in 2023 to 0.3% in 2024. On HumanEval it's a similar story. In February 2025, DeepSeek-R1 briefly matched the top US model outright. As of March 2026, Anthropic's flagship leads China's best by 2.7 points. That's it. That's the lead the United States bought for a hundred billion dollars.

US institutions released 40 notable AI models in 2024 versus China's 15. So the US is still out-shipping its rival on volume. But China now publishes 23.2% of global AI research papers and holds 69.7% of AI patent grants worldwide. The pipeline of future capability is not in America.

And while everyone was busy benchmarking, the industry quietly got darker. The Foundation Model Transparency Index dropped from 58 to 40 in a single year. Training code, dataset sizes, parameter counts — the stuff that used to be standard disclosure is now labeled "competitive advantage" and locked in a vault. Stanford's word, not mine: the largest labs are "increasingly keeping" this information to themselves.

The environmental receipt nobody wants

The report pegs Grok 4's training emissions at 72,816 tons of CO2 equivalent. That's the annual output of 17,000 cars. GPT-4 was 5,184 tons. Llama 3.1 405B was 8,930 tons. Training runs are now ~14x more carbon-intensive than they were eighteen months ago, and we're still calling it efficiency gains.

AI data center power capacity hit 29.6 GW — enough to run New York State at peak demand. The cumulative electricity draw of AI systems now rivals Switzerland or Austria. GPT-4o's annual inference water use alone may exceed the drinking water needs of 12 million people. These are Stanford's numbers, citing public reporting and lab statements.

Adoption is eating the economy faster than anything ever has

Generative AI hit 53% population adoption in three years. The PC took longer. The internet took longer. Stanford estimates US consumer value at $172 billion annually — and the median value per user tripled between 2025 and 2026. This isn't hype exhaustion. This is a generational behavior shift compressed into thirty-six months.

The workforce bill is also coming due. Stanford says AI's labor disruption has "moved from prediction to reality" and is hitting young workers first. Meanwhile India leads global AI hiring growth at ~33% annually, positioning itself as the talent supplier while the West argues about safety frameworks.

My Opinion

Here's what bugs me. American AI policy is still behaving like it's 2021 — like the only question is whether to slow down the labs or unleash them. Stanford's data says that argument is already obsolete. The US can't slow down because China didn't wait. The US can't unleash further because it already unleashed $109 billion and got a 2.7-point lead.

The honest read is that the money didn't buy a moat. It bought a parity race with a country that spent one twelfth of the budget. The only remaining American advantages are compute infrastructure, chip supply chains, and two or three individual researchers whose names you'd recognize. Everything else — benchmarks, publications, patents, even talent pipelines — is either tied or tilting the other way.

And the transparency collapse is the part I'll be watching. When labs go from 58 to 40 on disclosure in one year, that's not competitive strategy. That's a preemptive defense against accountability they expect to arrive. Watch that number next year. If it drops again, we're not in an AI race anymore. We're in an AI cartel, and the only question will be how long governments let it run.

Author: Yahor Kamarou (Mark) / www.humai.blog / 14 Apr 2026