Frontier AI is built by international talent. The engineers and researchers who design model architecture, build inference pipelines, and ship the products generating the sector's revenue are not cost-reduction hires. Regulatory Intelligence shows that dependency is not gradual. The curve is steep, the wave is recent, and every firm in the sector is on it.1
The AI Talent Arc
"Artificial Intelligence" first appeared in a sponsored job title on November 29, 2007. After that, the terminology did not return in meaningful volume until 2015, when 123 workers emerged under Machine Learning titles.
The first wave of industrial AI sits between 2017 and 2019. Worker positions certified under AI-related job titles grew from 326 in 2017 to 2,220 in 2019, a six-fold increase in two years. This was the first generation of dedicated ML engineering teams: tech companies scaling their first ML organizations and in volume for roles that had not existed at scale before.
2020 through 2023 was institutional buildup. Worker counts grew from 3,878 to 5,822. Median wages climbed from $140,000 to $161,700. The number of distinct job title variants expanded from 409 to 700, a signal that companies were creating specialized roles (ML Safety, ML Infrastructure, Applied Research) rather than using a single catch-all title.
2024 and 2025 marked the generative AI era. Worker positions for AI-titled roles reached 3,971 in 2024 and 6,183 in 2025. The median wage crossed $175,000. The vocabulary expanded to include "AI Engineer," "Generative AI," and "LLM," with 1,189 distinct title variants in 2025 alone, the broadest occupational spread in the dataset's history.
Across the United States, sponsored workers in AI-titled roles since 2015 total 35,434. 2026 alone, with only the first quarter complete, already accounts for 3,136 of them.
The Frontier Cohort
The first frontier-AI-native position was certified in 2016. By the end of 2025, the same company had certified 681 workers. A second-generation lab, founded in 2021, reached 638 worker positions in five years, the same scale the first-generation lab took ten years to reach. A third-generation company, incorporated in March 2023, certified 244 worker positions in less than three years.
Each generation of AI-native firm reaches sponsorship scale faster than the last.
The aggregate trajectory across the frontier cohort is steeper than any individual line. From 80 worker positions in 2020 to 1,220 in 2025, with 630 already certified in just the first quarter of 2026.
The numbers above are only representative. The actual sponsored workforce is larger.
The most senior tier of the AI workforce frequently arrives in the United States through pathways that produce no visible record. Individuals of extraordinary ability, outstanding researcher classifications, multinational transfers with specialized knowledge.
The Structural Dependency
The dependency is structural at the entry level. More than half of PhD-level Computer and Mathematical Scientists in the U.S. are foreign-born. Half of all S&E PhD recipients on temporary visas come from China and India, the two countries with the longest employment-based green card backlogs. The talent pool that AI companies hire from is, by demographic composition, an immigration-dependent pool. There is no version of frontier AI staffed primarily from the domestic doctorate market because that market does not exist at the scale required.2
Moreover, the industry-wide median wage for AI-titled roles in 2017 was $120,845. By 2021 it had reached $145,800. In 2025 it crossed $175,000. The first quarter of 2026 already shows $187,574. That is a 55% increase in the median over nine years, well above U.S. wage inflation across the same period.
At the frontier cohort the curve is steeper. Median wages at the leading research labs now sit between $245,000 and $310,000. The newer entrants cross $300,000 at the median for senior research roles. These are not import-substitution wages. They are scarcity wages, paid because the companies competing for the talent know what the alternative is, and the alternative is not a domestic candidate at the same level. The alternative is no hire.
The Steepening
The geography compounds the dependency. The two countries supplying the majority of the foreign-born S&E doctorate pool are the same two countries facing the longest employment-based green card backlogs. A worker certified today on an H-1B from a Chinese or Indian university faces a path to permanence measured in decades. That dynamic is embedded in the foundation of the U.S. AI industry, and embedded most heavily in the firms hiring at the steepest curve.
What follows is unambiguous. Any policy that constrains inbound international talent pipelines constrains AI capacity directly. The engineers and researchers in this dataset cannot be substituted with domestic equivalents because, by the demographic composition of the talent pool itself, those equivalents do not exist in sufficient numbers, and will not exist in the foreseeable future. The pipeline producing them runs through universities and research labs whose graduates have been arriving on temporary visas for thirty years.
The companies acting on this fact are absorbing one in five sponsored AI workers in the country. The companies still calibrating their position are losing the candidates they wanted, to the companies that already decided.