Stanford's 423-page AI report is out! The US-China gap is only 2.7%, Tsinghua's DeepSeek breaks into the top ten globally

Written by: New Intelligence Institute

Edited by: Very Tired Peach

【New Intelligence Institute Guide】Stanford’s “2026 AI Index Report” is finally released! This 432-page document is highly valuable: the peak competition between China and the U.S. in AI, the gap has almost disappeared, reduced to just 2.7%. The world’s top AI models produce about 95 annually, mostly from big tech companies. The harshest fact is that employment for developers aged 22-25 has already dropped by 20%.

Today, Stanford HAI proudly releases the “2026 AI Index Report”!

This 423-page annual report comprehensively reveals the latest power landscape of the global AI industry.

It offers a core conclusion: AI’s capabilities are growing rapidly; but humans’ ability to measure and control it has not kept pace.

The most shocking conclusion is—

The performance gap between Chinese and American AI models has essentially vanished, with both sides frequently changing dominance in peak competitions, currently with Anthropic leading by only 2.7%.

The U.S. invests more money in AI than anyone else, but recruiting top talent is becoming increasingly difficult.

The report also points out that AI’s evolution has not encountered a so-called “bottleneck”; instead, it is accelerating at an unprecedented speed.

Over the past year, more than 90% of top models worldwide have matched or surpassed human performance in doctoral-level scientific problems, multimodal reasoning, and competitive mathematics.

Especially in coding ability, SWE-bench scores soared from 60% to nearly 100% within a year.

However, AI’s “specialization” phenomenon is extremely severe, presenting a distorted state:

LLMs can win IMO gold medals but fail to read analog clocks correctly, with an accuracy of only 50.1%.

Meanwhile, AI’s threat to jobs has shifted from prediction to reality, and the first to suffer are the young “workers” of today.

Now, let’s get straight to the essentials—here are the 12 most noteworthy hardcore trends from the “2026 AI Index Report.”

Other highlights at a glance:

Global AI compute power increased 30-fold over three years, with Nvidia dominating 60%, nearly all chips produced by TSMC.

By 2025, global corporate AI investment reached $581.7 billion, doubling year-over-year, with the U.S. taking nearly half.

The number of AI researchers entering the U.S. has fallen 89% over 7 years, with an 80% drop just in the past year.

Employment for software developers aged 22-25 has declined 20% since 2024, with entry-level positions being precisely cut.

China has built 85 public AI supercomputers, more than twice North America’s total, ranking first globally.

AI usage rate in Chinese workplaces exceeds 80%, far above the global average of 58%.

The most powerful models are becoming increasingly black-box; out of 95 representative models, 80 do not disclose training code.

The gap between China and the U.S. has shrunk to just 2.7%.

Stanford mapped the U.S. No.1 and China No.1 rankings on the Arena leaderboard since May 2023 onto the same coordinate system.

In May 2023, GPT-4-0314 scored 1320 points, leading; China’s ChatGLM-6B lagged by over 300 points.

By February 2025, DeepSeek-R1 briefly tied with top U.S. models.

In March 2026, U.S. Claude Opus 4.6 scored 1503, China’s dola-seed-2.0-preview scored 1464.

Now, the gap between China and the U.S. in AI is only 39 points, or 2.7% in percentage terms.

Even more impressive is the frequency of shifts over the past year. Since early 2025, top models from both countries have exchanged positions on Arena multiple times.

The numbers are also roughly evenly split.

In 2025, the U.S. released 50 “notable models,” China followed with 30 top-tier models.

In the top tier, OpenAI, Google, Alibaba, Anthropic, and xAI share the stage, with the top 5 split evenly worldwide.

Further down, among the top 10, Chinese institutions and companies occupy four seats—Alibaba, DeepSeek, Tsinghua, ByteDance.

The open-source ecosystem’s focus has also shifted eastward this year.

DeepSeek, Qwen, GLM, MiniMax, Kimi have all advanced the open-source capability curve.

Adding in paper publications, citations, patents, and industrial robot installations, China ranks first globally in all these metrics.

Price-wise, another front.

Overseas developers have calculated on X that the output cost of Seed 2.0 Pro is roughly one-tenth that of Claude Opus 4.6.

Performance comparable, but at one-tenth the price. The chain reaction of this is just beginning.

90% of cutting-edge models come from industry, with speed of “god-like” breakthroughs unprecedented.

Of the 95 most representative models released last year, over 90% are from industry—not academia or government labs.

Academic research can no longer keep up with the frontier.

Release speed is also accelerating insanely.

In just February 2026 alone, eight or nine flagship models—Gemini 3.1 Pro, Claude Opus 4.6, GPT-5.3 Codex, Grok 4.20, Qwen 3.5, Seed 2.0 Pro, MiniMax M2.5, GLM-5—entered the scene in the same month.

The “god cycle” has shifted from “yearly” to “monthly.”

The upper limit of AI in one year shows no bottleneck.

The most impressive is programming.

SWE-bench Verified, a benchmark for real bug fixing, increased from 60% to nearly 100% in one year.

Not just a few points up, but basically capped.

Terminal-Bench, testing an agent’s ability to handle real terminal tasks, rose from 20% to 77.3%.

Cybersecurity agent success rate increased from 15% to 93%.

Gemini Deep Think won gold at the International Mathematical Olympiad.

PhD-level scientific Q&A (GPQA Diamond), competitive math (AIME), multimodal reasoning (MMMU)—these traditionally considered “impossible for AI to surpass”—have all been cracked by frontier models.

The best illustration is Humanity’s Last Exam.

This test, designed to “fool AI and favor human experts,” is provided by top specialists in various fields.

Last year, OpenAI’s o1 scored 8.8%; frontier models pushed the score up by 30 points in a year, with Claude Opus 4.6 and Gemini 3.1 Pro both surpassing 50%.

Jagged Frontier

Able to win IMO gold medals but unable to read clocks properly

But the same index reveals another set of numbers.

The top models’ accuracy in “reading analog clocks” is only 50.1%.

Robots in lab simulations (RLBench) have an operation success rate of 89.4%. But in real household scenarios—doing dishes, folding clothes—the success rate drops immediately to 12%.

A gap of 77 percentage points between lab and kitchen.

Researchers call this phenomenon “Jagged Frontier.” AI capabilities are uneven—able to win math olympics gold but unable to reliably tell the time.

AI can win gold in math competitions, but only half the time can it understand an analog clock. AI is accelerating, but not in the same directions.

Additionally, in agent tasks, in OSWorld tests, frontier AI (66.3%) is approaching human baseline.

However, in the specialized PaperArena scientific reasoning test, the strongest AI-powered agent scores only 39%, half that of a PhD student.

But this unevenness does not prevent companies from deploying AI on production lines.

Another figure from the AI Index: global enterprise AI adoption rate has reached 88%. Nine out of ten companies have integrated AI into some workflow.

Costs are rising in tandem. AI-related incidents increased from 233 in 2024 to 362.

Money is accelerating—$581.7 billion poured into AI in 2025.

In 2025, global enterprise AI investment hit $581.7 billion, up 130% year-over-year. Private equity investments totaled $344.7 billion, also up 127.5%.

Both curves nearly doubled.

Country-wise, the U.S. dominates. In 2025, U.S. private AI investment was $285.9 billion, with 1,953 new AI startups—more than ten times the second place.

Money is flowing faster into the U.S., but another core resource—people—is flowing out.

People are leaving—AI researchers and developers entering the U.S. have fallen 89%.

A startling figure: since 2017, the number of AI researchers and developers entering the U.S. has decreased by 89%.

And the decline is accelerating—just in the past year, down 80%.

The U.S. remains the country with the highest density of AI researchers, but the inflow is tightening.

The curves of money and talent are reversing—something unseen in the past decade.

Three-year surge in compute power, all in one company’s hands

AI capability curves are accelerating, and behind them, the compute power curve is moving even faster.

Since 2021, global AI compute power has increased 30-fold. Over the past three years, it has grown more than threefold each year.

This curve is supported by a handful of companies.

Nvidia’s GPUs account for over 60% of global AI compute power. Amazon and Google rely on self-developed chips, ranking second and third, but together they still fall far behind Nvidia.

Almost all these chips are produced by TSMC. The steeper the compute curve, the narrower the “choke point.”

Meanwhile, costs are rising.

The total power consumption of global AI data centers has reached 29.6 GW—equivalent to the peak electricity demand of New York State. The estimated carbon emissions from training one xAI Grok 4 model are 72,816 tons of CO₂ equivalent—comparable to the emissions of 17,000 cars over a year.

Where data centers are located, where the electricity comes from, and where chips are manufactured—these three questions are now the most headache-inducing for AI CEOs.

Generative AI penetrates 53% in three years; workplace adoption in China exceeds 80%

Generative AI achieved a 53% penetration rate among the global population in just three years.

This speed surpasses that of personal computers and the internet.

But the penetration rate varies greatly by country. Singapore at 61%, UAE at 54%, both ahead of the U.S. at 28.3%, which ranks 24th among surveyed countries.

If shifting from consumers to workplaces, the contrast is even starker.

Another set of data shows that by 2025, 58% of global employees are regularly using AI at work. But in China, India, Nigeria, UAE, and Saudi Arabia, this proportion exceeds 80%.

China’s workplace AI penetration already surpasses the global average by more than 20 percentage points.

More interesting is the value to consumers.

AI Index estimates that by early 2026, generative AI tools will generate $172 billion annually for U.S. consumers. From 2025 to 2026, the median value per user has tripled.

Most users are still using free versions.

Ordinary people are willing to pay far less for AI than the value AI creates for them. This gap is what all AI companies are now trying to bridge.

Entry-level jobs sharply decline; employment for 22-25-year-old developers drops 20%

Perhaps the most silent yet impactful part of the AI Index for Chinese readers is the employment of young people.

The number of software developers aged 22-25 has decreased by about 20% since 2024.

Meanwhile, older colleagues are actually increasing.

Not only developers—other AI-exposed industries like customer service are experiencing similar patterns.

Even more worrying are the results from corporate surveys. Senior executives generally expect layoffs to be even larger in the coming months.

This is not about macro unemployment; it’s about entry-level positions being precisely cut.

Losing the first job, the entire career ladder is broken. The long-term impact of this is still unpredictable.

AI is rewriting the way scientific discoveries are made

If employment is cold, science is hot.

In natural sciences, physics, and life sciences, AI-related papers increased by 26% to 28% in 2025 compared to the previous year.

Specifically, this year, AI successfully ran end-to-end weather forecasting processes, directly producing temperature, wind speed, and humidity forecasts from raw meteorological data, without traditional numerical models.

AI is shifting from “helping write papers” and “helping with calculations” to “making discoveries on its own.”

Hospitals are doing the same. In 2025, many hospitals deployed AI tools that automatically generate clinical records from consultation dialogues. Doctors reported up to 83% reduction in time spent writing medical records, significantly reducing burnout.

But the same index casts a cold shower on medical AI. A review of over 500 clinical AI studies found nearly half relied on exam-like datasets, with only 5% using real clinical data.

AI can reduce doctors’ keyboard time—this is certain. But its clinical value on real patients remains uncertain.

Self-learning wave explodes globally; formal education falls behind

Formal education can no longer keep up with AI.

In the U.S., 4 out of 5 high school and college students now use AI to complete assignments. But only half of high schools have AI policies, and only 6% of teachers find these policies clear.

Students are ahead, teachers are still in place, and rules have yet to be established.

Meanwhile, the self-learning wave is exploding worldwide. The top three countries for AI engineering skill growth are the UAE, Chile, and South Africa.

Not the U.S., not Europe.

The steepest part of the skill curve is growing where no one is watching.

The most powerful models are becoming the least transparent—experts and the public are increasingly divided.

The Foundation Model Transparency Index’s average score this year dropped from 58 to 40. AI Index explicitly states that Google, Anthropic, and OpenAI have already given up on disclosing the training data scale and duration of their latest models.

Of the 95 most representative models released last year, 80 do not disclose training code.

Public sentiment is also becoming more complex.

Globally, the proportion believing AI does more good than harm increased from 52% to 59%. But at the same time, those feeling anxious about AI rose from 50% to 52%.

Both trends are growing simultaneously.

The most divided is the U.S.—only 33% believe AI will improve their jobs, compared to a global average of 40%. Trust in the U.S. government’s regulation of AI is the lowest among surveyed countries, at 31%.

In Singapore, trust in government regulation is 81%.

After the recent attack on Sam Altman’s home, Silicon Valley insiders “were surprised to find” that ordinary people on Instagram comments didn’t sympathize; some even thought “it should be more intense.”

They didn’t realize how bad things had become.

According to Pew and Ipsos data cited in the report, the perception gap between experts and the public on AI’s impact on employment, healthcare, and the economy often exceeds 30 percentage points, with the largest gap reaching 50 points.

On one side, the curve in labs is soaring; on the other, public anxiety is mounting.

There is no bridge between them.

Final words

The 423-page report contains hundreds of charts, but in fact, only one graph.

The horizontal axis is time; the vertical axis is capability.

Model capability curves are soaring, compute power is soaring, investment is soaring, adoption is soaring. Everything else remains stagnant or declining.

This is the entire content of the 2026 AI Index.

AI is accelerating. Everything else is out of sync.

If you are in this industry, the question is no longer “what will the future be,” but “where do I stand on this curve.”