Stanford’s Institute for Human-Centered AI just dropped the ninth edition of its AI Index Report — 400+ pages of data on where AI actually stands. Not where the hype says it is. Not where the doomers think it’s headed. Where it measurably is.

The short version: AI is more capable, more adopted, and more expensive than ever. It’s also less transparent, more environmentally destructive, and outrunning every guardrail we’ve built. Here are the numbers that matter.

The US-China Gap Has Evaporated

For years, America held a comfortable lead in AI performance. That’s over.

US and Chinese models have been trading the top spot on benchmarks throughout the past year. The Arena leaderboard (March 2026) shows Anthropic leading, followed by xAI, Google, and OpenAI — with DeepSeek and Alibaba trailing by margins thin enough to be noise.

The US still dominates infrastructure: 5,400+ data centers, more than ten times any other country. But China leads in AI patents, research publications, and — critically — robotics deployment. China installed 295,000 industrial robots in 2024. The US installed 34,200.

That ratio should worry people more than benchmark scores. Software models are impressive. Robots that do things in the physical world are a different kind of advantage entirely.

And it’s not a two-player game anymore. South Korea leads the world in AI patents per capita. Forty-four countries now run state-backed supercomputing clusters. The sprint became a marathon, and the field is crowded.

Performance Keeps Climbing — We’re Running Out of Tests

Despite the “AI has plateaued” takes, the models keep improving. On SWE-bench Verified — a benchmark for real-world software engineering — top scores jumped from ~60% to near-perfect in a single year.

AI now matches or beats human experts on PhD-level science, math, and language understanding. An AI system independently produced a weather forecast in 2025. Adoption has hit 88% of organizations. Four in five university students use generative AI regularly.

“I am stunned that this technology continues to improve, and it’s just not plateauing in any way,” says USC computer scientist Yolanda Gil, a report co-author.

But there’s a catch. AI can ace a doctoral physics exam and then fail at common-sense reasoning a kindergartener handles instinctively. Robots succeed at just 12% of household tasks. The gap between benchmark performance and real-world reliability remains enormous — and we’re bad at predicting where the holes are.

The Environmental Bill Is Arriving

This is the section of the report that should make everyone uncomfortable.

AI data centers worldwide now draw 29.6 gigawatts of power — enough to run New York State at peak demand. Running just GPT-4o consumes more water annually than 12 million people need for drinking.

Training costs are exploding. Training xAI’s Grok 4 generated an estimated 72,000+ tons of CO₂-equivalent emissions. GPT-4 was estimated at 5,184 tons. That’s a 14x increase in a few model generations. An independent Epoch AI estimate puts Grok 4 even higher, at roughly 140,000 tons.

AI compute capacity has grown 3.3x per year since 2022 — a 30-fold increase since 2021. Nvidia GPUs account for over 60% of global AI compute.

Yes, efficiency is improving. DeepSeek’s V3 models use significantly less power during inference. But efficiency gains are being drowned by the scale of deployment. The treadmill is accelerating faster than we’re running.

Every ChatGPT query has a real cost in watts, water, and carbon. At current trajectory, AI’s environmental footprint becomes a political issue — not just an engineering one.

Transparency Is Collapsing

Google, Anthropic, and OpenAI have all stopped disclosing training dataset sizes for their latest models. Eighty of 95 notable models launched last year shipped without training code. Over 90% of notable AI models now come from private companies, up from under 50% in 2015.

This matters because, as Gil puts it: “We don’t know a lot of things about predicting model behaviors.” When you can’t see inside the model, you can’t study why it fails, what biases it carries, or how it behaves in situations its creators didn’t anticipate.

Meanwhile, AI industry lobbyists have tripled their presence at congressional hearings since 2017 while academic voices have plummeted. The companies building the most powerful systems are also the ones shaping the rules — and they’re telling us less about what those systems actually do.

Everyone’s Using It, Nobody Trusts It

Generative AI has hit 53% global adoption — outpacing PCs, the internet, and smartphones. Corporate AI investment has grown 40-fold since 2013. US consumer surplus from generative AI reached $172 billion this year.

But trust is cratering. Only 31% of Americans trust their government to regulate AI properly. 52% say AI makes them nervous, even as 59% acknowledge it delivers more benefits than drawbacks.

There’s a striking geographic split. The US ranks just 24th in adoption — only 28.3% of Americans use generative AI regularly. In Southeast Asia, over 80% expect AI to profoundly impact their lives within five years.

America is building the tools but not using them. The rest of the world is doing the opposite. That gap has implications for workforce competitiveness, cultural influence, and which values get baked into the next generation of AI systems.

The Big Picture

If this report had a thesis statement, it would be: AI is a teenager with a sports car — growing faster than anyone expected and operating without enough adult supervision.

Coding benchmarks approaching 100%. Scientific reasoning matching PhD experts. Adoption faster than the smartphone revolution. These are generational leaps happening in months.

But the support structures haven’t kept up. Transparency declining. Environmental costs surging. Hardware supply chains dependent on a single fab in Taiwan. Public trust eroding even as usage climbs. Geopolitical stakes ratcheting higher.

AI isn’t a tech story anymore. It’s an energy story, a geopolitics story, a labor story, and an environmental story — simultaneously. The models will keep getting more powerful. Whether everything around them can mature fast enough is the question this report can’t answer.

The data says probably not. But data has been wrong before.

Sources: