Stanford's 2026 AI Index: The US-China Gap
TL;DR
The US-China frontier race is now basically neck-and-neck — Stanford’s 2026 AI Index says the performance gap has effectively closed, with DeepSeek R1 briefly matching the best US model in February 2025 and Anthropic leading top Chinese models by just 2.7% as of March 2026.
Benchmarks are breaking because models are improving too fast — frontier systems now hit or exceed human baselines on PhD-level science, multimodal reasoning, and competition math, while SWE-bench Verified jumped from roughly 60% to near 100% in a single year.
Top model capability is converging, so the real fight is shifting elsewhere — Anthropic, xAI, Google, and OpenAI are clustered within just 25 Elo points on the Arena leaderboard, which is why cost, reliability, and domain-specific performance matter more than raw benchmark wins.
AI adoption is exploding faster than the internet did, but so is the footprint — generative AI reached 53% global population adoption within three years, even as Grok 4’s training emissions hit about 72,000 tons of CO2e and AI data center power capacity approached 30 gigawatts, roughly New York State’s peak demand.
The labor shock is showing up first in junior roles and education is nowhere near ready — software developer employment for ages 22 to 25 fell nearly 20% since 2024, four out of five US high school and college students now use generative AI for schoolwork, and only 6% of teachers say school AI policies are clear.
There’s a huge trust gap between experts and the public — 73% of AI experts expect AI to improve how people do their jobs versus just 23% of the US public, while only 31% of Americans trust their government to regulate AI responsibly.
The Breakdown
A 400-page report that tries to map all of AI
The hosts open on Stanford HAI’s 2026 AI Index as one of the biggest macro-level snapshots of AI out there: over 400 pages spanning research, performance, investment, jobs, energy, policy, and public opinion. Mike’s take is simple: even if nobody reads the whole thing, it’s the kind of report worth skimming or tossing into NotebookLM because it gives you the big picture fast.
The US-China gap is no longer a comfortable lead
One of the headline findings is that the US-China AI performance gap has effectively closed. Chinese and US models have traded the top spot since early 2025, DeepSeek R1 briefly matched the best US model, and by March 2026 Anthropic’s top model was ahead by just 2.7%; meanwhile China is already ahead on publication volume, citations, patents, and industrial robot installations.
Models are brilliant and weirdly broken at the same time
The report captures the classic “jagged intelligence” story the hosts often talk about: frontier models can ace PhD-level science questions, multimodal reasoning, and math competitions, yet still miss surprisingly basic tasks. Mike’s favorite example is that they can win Olympiad-level math but still fail to read the time correctly about half the time, which is exactly the kind of uneven progress that makes AI both impressive and slippery.
Frontier models are bunching together, so evals matter more than ever
Paul zeroes in on a big shift: the top four models from Anthropic, xAI, Google, and OpenAI now sit within 25 Elo points of each other, meaning raw capability alone is no longer a clean differentiator. That’s why he argues companies need their own internal evals for real workflows, especially when models keep changing underneath users — he jokes about seeing someone complain that “Claude 4.7” suddenly got sassier and became unusable for their needs.
Adoption is absurdly fast, and the infrastructure bill is coming due
Generative AI hit 53% global adoption in just three years, faster than the PC or internet, while global corporate AI investment more than doubled in 2025 to $581.7 billion and private investment hit $344.7 billion. But the hosts pair that excitement with the physical cost: Grok 4 training emissions were about 72,000 tons of CO2e, and AI data center capacity is nearing 30 gigawatts — enough to make the energy side of the story impossible to ignore.
The labor market warning signs are showing up at the bottom rung
The hosts linger on one especially consequential stat: employment for software developers aged 22 to 25 is down nearly 20% since 2024, with similar pressure in customer service and other AI-exposed roles. Paul notes that one-third of organizations expect AI to reduce headcount in the coming year, even if the broader jobs data hasn’t fully caught up yet.
Students are already all in, while schools are still improvising
Four out of five US high school and college students now use generative AI for schoolwork, but the institutional response is lagging badly. Only half of middle and high schools even have AI policies, and just 6% of teachers say those policies are clear — a mismatch that makes the current education moment feel less like a transition and more like a scramble.
Public trust is low, and Paul thinks government spending is about to spike
The report shows a striking split: experts are much more optimistic than the public, 64% of Americans expect fewer jobs because of AI, and the US has the lowest trust in government to regulate AI responsibly at 31%. Paul’s prediction is that this won’t stay a mostly private-sector story for long — he thinks the government may soon subsidize energy, chips, data centers, and workforce training at massive scale, potentially even reaching over $1 trillion a year within one to two years.