AI 연구원의 89%가 미국에 오는 것을 중단했습니다.

Key Takeaways

• 89% fewer AI researchers are moving to the US compared to 2017, with an 80% collapse in the most recent year alone, the same year the $100K fee took effect
• The US outspends China 23-to-1 on AI ($285.9 billion vs. $12.4 billion in 2025), yet the performance gap between their best models has closed to just 2.7 percentage points
• A $100,000 H-1B “visa integrity fee” imposed in September 2025 triggered a 34-50% drop in filings from Amazon, Meta, and Google, while FAANG companies added 33,000 jobs in India instead
• For every rejected H-1B visa, US firms hire 0.4 to 0.9 workers abroad, accelerating the offshoring of the exact R&D capability the policy claims to protect

The $100,000 Welcome Mat

On September 19, 2025, the Trump administration signed a presidential proclamation imposing a $100,000 “visa integrity fee” on every new H-1B visa petition filed from abroad. The previous cost was $2,000 to $5,000 per petition, a 20 to 50-fold increase overnight. The fee took effect two days later, on September 21.

Commerce Secretary Howard Lutnick framed the logic plainly: “Do we need to have a person valuable $100,000 a year to the government, or should they return home and hire an American?”

Seven months later, Stanford’s Human-Centered AI Institute released its 2026 AI Index Report and answered his question with data. The number of AI researchers and developers moving to the United States has dropped 89% since 2017, with 80% of that collapse concentrated in the most recent year alone. The math implies the decline was slow for years: an erosion of roughly 45% between 2017 and early 2025, driven by tightening visa rules, COVID disruptions, and maturing AI ecosystems abroad. Then something accelerated. The year the $100,000 fee took effect was the year the pipeline broke.

The US didn’t just lose a policy debate. It lit $285.9 billion in AI infrastructure spending on fire by driving away the people who know how to use it.

The Paradox in Three Numbers

The Stanford AI Index contains a paradox that should alarm anyone who cares about American technological competitiveness.

$$\text{US AI Investment} = 23 \times \text{China AI Investment}$$

In 2025, US private AI investment reached $285.9 billion, more than 23 times China’s $12.4 billion. The US produced 50 notable AI models in 2025, compared to China’s 30. It still has 1,953 newly funded AI companies, more than ten times the next closest country.

But the gap between the best American AI model and the best Chinese AI model? 2.7 percentage points.

That gap was 9.26% in January 2024. On some benchmarks, it collapsed to virtually zero. On MMLU (Massive Multitask Language Understanding), the margin fell from 17.5% to 0.3% between 2023 and 2024. Chinese and American models have traded the top position multiple times since early 2025, with DeepSeek-R1 briefly matching the top US model.

Three numbers define the paradox:

Metric	Value	Direction
US AI spending advantage	23x over China	↑ Growing
US-China model performance gap	2.7%	↓ Shrinking
AI talent moving to the US	-89% since 2017	↓ Collapsing

America is spending 23 times more money to maintain a lead that has shrunk to 2.7%, while the talent pipeline that built the lead is down 89%. You don’t need a PhD in economics to see that this equation doesn’t balance.

The Hiring Cliff

The $100,000 fee didn’t hit the industry uniformly. It hit the companies that build America’s AI infrastructure the hardest.

Department of Labor data on certified H-1B applications for Q1 of fiscal year 2026 (October through December 2025, the first full quarter after the fee took effect) shows steep declines across nearly every major technology employer:

Company	H-1B Filing Change (Q1 FY2026 vs. FY2025)
Amazon	-34% (4,647 → 3,057 certified applications)
Meta	~-50% year-over-year
Google	~-50% year-over-year
Walmart	>-50% year-over-year
Nvidia	+18% (369 → 434 certified applications)

One company bucked the trend. Nvidia, whose GPUs underpin the entire AI buildout, increased its H-1B filings by 18%. At $100,000 per petition, Nvidia can afford it. A 15-person AI startup cannot. The fee doesn’t filter for quality. It filters for treasury size.

Meanwhile, the Labor Department ramped up H-1B enforcement, with probes jumping 48%.

The message to the world’s top AI researchers was unambiguous: America will charge $100,000 to walk through the door, audit the employer after arrival, and then watch the job go to Bangalore instead.

Where the Talent Went

The researchers didn’t disappear. They went somewhere else.

In 2025, Meta, Amazon, Apple, Microsoft, Netflix, and Google collectively added 33,000 workers in India, an 18% increase over the previous year. As of early 2026, these companies had roughly 4,200 open positions in India, with nearly half in AI, machine learning, cloud computing, and cybersecurity. Google alone listed 365 open roles, two-thirds of them in Bengaluru.

The pattern is not accidental. It is a direct, measurable consequence of visa restrictions.

A February 2026 report by the Center for Strategic and International Studies (CSIS) found that for every rejected H-1B visa, multinational corporations hire 0.4 to 0.9 employees abroad. Among the most globally integrated firms, the ratio approaches 0.9. The underlying research, published in Management Science and the National Bureau of Economic Research, is blunt: “H-1B restrictions do not protect American jobs. They result in more hiring abroad.”

The $100,000 fee was designed to “protect American workers.” The CSIS data says it is doing the opposite: shipping the jobs overseas faster than any trade deal ever could.

The Founders America Won’t Get

The talent drain isn’t just about engineers filling seats at Google. It is about the next generation of companies that will never be founded in the United States.

An analysis by the Institute for Progress (IFP) of the 2025 Forbes “AI 50” list, the most promising AI startups in America, found that 60% of the US-based companies were founded or co-founded by immigrants. Twenty-five of the 42 US-based companies on the list had at least one immigrant founder. The National Foundation for American Policy found the same pattern in 2023, with 65% immigrant-founded companies. Seventy percent of those founders first came to the US on student visas.

These aren’t people filling American jobs. They are people creating American jobs. OpenAI was co-founded by Ilya Sutskever, who came from Russia via Israel and Canada. Google was co-founded by Sergey Brin, a refugee from the Soviet Union. Tesla’s CEO came from South Africa via Canada. The $100,000 fee doesn’t distinguish between a routine outsourcing application and the next Sutskever.

The Bulletin of the Atomic Scientists put it starkly: the fee will “hurt Silicon Valley and AI startups” most, because early-stage companies cannot absorb $100,000 per visa on top of legal costs, relocation, and salary. The large incumbents absorb the cost or offshore. The startups that would challenge them never form.

The Historical Rhyme No One Wants to Hear

The United States has seen this movie before. It was just on the other side of the screen.

In April 1933, Nazi Germany issued a law requiring the dismissal of anyone in government positions, including universities, who had at least one Jewish grandparent or who was a political opponent. Twenty-five percent of German physicists lost their jobs, including eleven past or future Nobel Prize winners. Roughly 2,400 academics fled Germany in the first wave alone.

The scientists who reached America and Britain read like a roster of the 20th century’s most consequential minds: Albert Einstein, Hans Bethe, John von Neumann, Leo Szilard, Edward Teller, Rudolf Peierls from Germany, and Enrico Fermi from Italy, who fled Mussolini’s racial laws in 1938. They didn’t retire. They built the atomic bomb — for the countries that took them in.

Stanford economist Petra Moser found that US patents increased by 31% in fields common among the émigré scientists after 1933. The innovation effect rippled for generations: the refugees attracted new researchers, who trained more researchers, who built Silicon Valley.

Germany was the world’s leading scientific power in 1932. By 1945, it was rubble. The relationship between those two facts is not coincidental.

The parallel is imperfect. The US isn’t persecuting scientists on racial grounds. It is pricing them out on bureaucratic ones. But the mechanism is identical: a dominant power drives away its intellectual base through policy, and the beneficiaries are the countries smart enough to welcome them. In the 1930s, that was America. In the 2020s, it is Canada, India, and Europe.

The Capital-for-Talent Substitution

The implicit bet behind America’s AI strategy is that capital can substitute for talent. Build enough data centers. Buy enough GPUs. Spend enough money. And the 2.7% gap won’t matter because the infrastructure will generate its own advantages.

History is not kind to this theory.

The Soviet Union fielded larger conventional military forces than the United States for decades during the Cold War. It had more tanks, more artillery, more soldiers. But the United States had better technology and better-trained people. When the gulf between quantity and quality became unsustainable, the Soviet system collapsed, not from a battle lost, but from an economy that couldn’t convert spending into capability.

The AI version of this dynamic is already visible. DeepSeek, a Chinese lab, trained a model that briefly matched the top American model for a reported $5.6 million in direct GPU costs, a fraction of what US frontier models spent on training alone. The gap didn’t close because China outspent America. It closed because China used its talent more efficiently.

The irony cuts deeper. According to Andreessen Horowitz partner Martin Casado, roughly 80% of US startups that build on open-source models are now building on Chinese ones, primarily DeepSeek and Alibaba’s Qwen. Not OpenAI. Not Anthropic. The American AI ecosystem is increasingly running on Chinese foundations, while American immigration policy drives away the researchers who could build domestic alternatives.

Meanwhile, the US continues to tighten the policies that created the very talent advantage being eroded. Chip export bans were supposed to cripple China’s AI hardware supply; instead, they accelerated domestic Chinese chipmaking. Now a $100,000 visa fee is supposed to protect American AI workers; instead, it is accelerating the offshoring of AI research.

The pattern is consistent: every attempt to weaponize American advantages is turning them into liabilities.

The Steelman: Why This Might Not Matter

The strongest counterargument deserves honest engagement.

The US still has the world’s deepest AI ecosystem. It produced 50 notable AI models in 2025, compared to China’s 30. Its private AI investment is 23 times larger than China’s. It hosts more AI researchers than any other country. It has the best universities, the most sophisticated venture capital networks, and the largest installed base of AI infrastructure on Earth.

Some of the talent decline may reflect natural maturation of AI ecosystems abroad, with more opportunity at home rather than less opportunity in America. And the $100,000 fee does prioritize higher-wage applicants, which could theoretically shift the composition of H-1B hires toward more senior, higher-value talent rather than bulk outsourcing positions.

The fee also expires on September 21, 2026, five months from now. It could be renewed, but it could also lapse.

All true. But the data implies the first 45% of the decline happened gradually over eight years. The remaining collapse compressed into one. The fee may not be the sole cause, but it arrived in the same year the cliff appeared. And the CSIS data demonstrates that the offshoring it triggers is not easily reversed. Once a Google establishes an AI research lab in Bengaluru and hires 33,000 people across India, those jobs do not come back to Mountain View when the fee expires. The infrastructure, the talent networks, and the institutional knowledge have already relocated.

The Self-Defeating Cycle

Here is the feedback loop that should terrify American policymakers:

1. Restrict immigration → AI talent stops coming (89% decline)
2. Companies can’t fill domestic roles → They offshore to India, Canada, Germany (0.4-0.9 jobs per rejected visa)
3. Foreign AI ecosystems mature → Chinese, Indian, Canadian labs produce competitive models (2.7% gap)
4. US responds by spending more → $285.9 billion in 2025 alone
5. But capital without talent is just expensive hardware → Spending goes up, lead keeps shrinking
6. Repeat → The cycle accelerates

This isn’t a skills gap that a training program can fix in two years. It is a structural withdrawal from the global talent market at the exact moment when AI development depends on the deepest possible talent pool.

The roughly $660 billion in AI data center capital expenditure projected for America in 2026 assumes that the engineers, researchers, and scientists needed to develop the AI models running inside those centers will be available. The Stanford data suggests they increasingly won’t be, at least not in the United States.

What Happens Next

The AI race is not a sprint. It is a relay. And America just told 89% of its incoming runners to find another team.

The immediate consequences are already measurable: FAANG hiring surges in India, collapsing visa applications at home, and a performance gap with China that shrinks with every model release. The medium-term consequences are structural: competing AI ecosystems in Toronto, Bengaluru, London, and Berlin are absorbing the talent America is rejecting, building the institutional knowledge that will compound for decades.

The long-term consequence is the one that should haunt American strategists. In 1933, Germany had every advantage: the best universities, the most Nobel laureates, the largest scientific establishment in Europe. Within twelve years, the scientists it drove out had built the weapon that ended the war. The country that won wasn’t the one that spent the most. It was the one that welcomed the talent.

The US spent $285.9 billion on AI in 2025. The visa to bring in the people who would use it costs $100,000. The math isn’t hard. The question is whether anyone in Washington can do it.