America Is Winning the AI Spending Race While Losing the Scientists Who Make It Possible

Here is the number that changes everything in the Stanford AI Index 2026: 89%. That is the collapse in the flow of AI researchers and developers choosing to move to the United States since 2017. And of that 89% total decline, 80% happened in the last year alone. Not a gradual erosion. A cliff. The United States is spending $285.9 billion annually on AI, more than 23 times what China spends, and the people who actually build frontier AI systems are choosing to go somewhere else. That is the story the headlines missed.

What the 2026 AI Index Actually Found

Stanford's Institute for Human-Centered Artificial Intelligence published the 2026 AI Index in April, and the headline numbers are the kind that get shared widely: AI reached 53% consumer adoption within three years of mainstream emergence, faster than the personal computer or the internet. Models now meet or exceed human performance on PhD-level science questions. On SWE-bench Verified, the key benchmark for software engineering capability, performance jumped from 60% to near 100% in a single year. Enterprise organizational adoption reached 88%. Four in five university students now use generative AI. The estimated annual value of generative AI tools to US consumers reached $172 billion. These are the numbers AI optimists share, and they are genuinely extraordinary.

But the report contains a second set of findings that received far less attention. The Foundation Model Transparency Index, which tracks how openly AI companies document their training data, safety processes, and model architecture, dropped from 58 points to 40 points in a single year. As models became more powerful, companies shared less about how they work. AI researcher migration into the United States fell 89% since 2017, with 80% of that decline occurring in the past year. Switzerland has surpassed the United States to rank first globally for AI researchers and developers per capita. And despite US private AI investment of $285.9 billion in 2025, China closed the AI model performance gap on key benchmarks to just 2.7% by March 2026, having started the year significantly further behind. These are not footnotes. They are the structural story underneath the capability headlines.

Why This Matters More Than People Think

The dominant theory of American AI leadership is a capital theory: the US outspends everyone else, and spending converts into compute, which converts into capability. The Stanford data challenges this theory at its foundation. China spent approximately $12.4 billion on AI in 2025, less than 5% of the US figure, yet closed the performance gap to 2.7% by March 2026. The gap narrowed while China's investment was 23 times smaller. This is not measurement error; it is a measurement of the efficiency differential between the two approaches. The US is buying compute at scale. China is optimizing algorithms and deployment efficiency. Four Chinese labs released competitive open-weights coding models in a 12-day window in early 2026, GLM-5.1, MiniMax M2.7, Kimi K2.6, and DeepSeek V4, all performing at roughly the frontier on agentic engineering benchmarks at less than a third of the inference cost of comparable Western models. Infrastructure investment is necessary but not sufficient for sustained AI leadership.

The $172 billion consumer value figure is the most underappreciated number in the entire report. Most public discussion of AI's economic impact focuses on productivity gains and job displacement, AI as an input to other economic activity. But $172 billion in direct consumer value means AI has crossed the threshold into being an economic sector of its own, not just a productivity multiplier. For context: $172 billion annually exceeds the entire GDP of Hungary. It represents only the consumer segment, enterprise value is substantially larger and measured separately. The implication is structural: AI is no longer a feature of the economy. It is becoming a sector of the economy, with its own value chains, supply chains, and competitive dynamics that need to be analyzed as such.

The Competitive Landscape

The talent data is the most structurally significant finding in the report, and it operates differently from other competitive dynamics. Compute advantage can be acquired through capital expenditure. Algorithm improvements can be achieved through research investment. But talent, specifically the researchers who discover new architectures, identify training instabilities, and solve the problems separating frontier from near-frontier models, cannot be purchased in bulk. It accumulates through institutional culture, career trajectory, and geographic ecosystem effects over years. The 89% decline in AI researcher migration into the US is not a one-quarter blip. It represents a structural change in where the most capable people in this field are choosing to build their careers, and that choice compounds.

Switzerland's emergence as the per-capita leader for AI researchers is not coincidental. It combines excellent universities (ETH Zurich, EPFL), political neutrality, strong rule of law, and proximity to European research networks, without the immigration uncertainty that has characterized the US environment since 2017. The implication for long-term AI competition is that Europe may be quietly building a talent reservoir that neither the US nor China has fully recognized. European AI companies, Mistral in France, Wayve in the UK, and a growing number of university spinouts across Germany, Switzerland, and the Netherlands, are positioned to benefit from a talent pool that is neither migrating to Silicon Valley nor absorbed into Chinese national programs. Watch for European AI research publication share as a leading indicator of this shift over the next 18-24 months.

Hidden Insight: The Transparency Collapse Is the Canary Nobody Is Watching

The drop in the Foundation Model Transparency Index, from 58 to 40 points in a single year, is the finding that should most alarm anyone thinking seriously about where the AI industry is headed. The index measures how openly companies document their training data, safety processes, model architecture, and evaluation methodologies. A decline of 18 points in one year, during a period when AI systems became dramatically more capable, means the most powerful AI systems in the world are simultaneously the least understood by anyone outside the companies that built them. This is not primarily a regulatory problem. It is an epistemic problem. We are deploying systems we cannot independently evaluate because the organizations deploying them have reduced transparency as a competitive response to a race where documentation is treated as a liability rather than a responsibility.

The transparency decline compounds the talent data in a way that deserves explicit attention. Researchers who want to understand how models work, publish their findings, and build on each other's discoveries are less well-served by an ecosystem where the most capable models are black boxes. Academic AI research has always depended on access to published findings from industry labs. When industry labs reduce transparency, as the index confirms is happening at accelerating pace, academic and independent research becomes progressively disconnected from the frontier. The researchers leaving the US are not just leaving a geography; they may be leaving an ecosystem that is becoming less hospitable to the kind of open inquiry that produced the breakthroughs being monetized today.

The most uncomfortable conclusion from the 2026 AI Index is that the two metrics we most need to track, capability and safety, are moving in opposite directions. SWE-bench near 100% and PhD-level science performance are extraordinary achievements. Foundation Model Transparency Index at 40 and declining, with AI researcher migration at its lowest point in nearly a decade, are serious sustainability signals. The industry is building faster than it is explaining what it is building, and the people who could provide the most independent checks on that process are choosing to be somewhere else. That is not a trajectory any investor presentation, policy document, or industry conference has yet seriously grappled with. The 2027 AI Index will tell us whether this year was an anomaly or a trend.

What to Watch Next

The most critical leading indicator to track is whether AI researcher migration into the US reverses or continues to decline through 2026. Stanford's data reflects collection through early 2026; the full-year picture will capture the impact of US immigration policy developments, global AI talent market tightening, and whether major non-US AI hubs, Switzerland, UK, UAE, Canada, continue building ecosystem gravity. Watch specifically for foundational AI research publications in Nature, Science, and NeurIPS from non-US institutions: if their share continues growing relative to US institutions, it confirms the talent drain is translating into a research capability gap, not just a geographic preference.

On the China performance gap, the 2.7% figure as of March 2026 is a snapshot, not a steady state. The key question is whether Chinese labs can sustain their improvement rate while the US frontier continues advancing. DeepSeek V4, which reportedly required substantially less compute than comparable Western models, suggests algorithmic innovation is compensating for compute constraints imposed by export controls. If China achieves benchmark parity on multimodal reasoning and complex agentic tasks within the next two quarters, possible given the pace of improvement, the compute theory of US AI dominance will require fundamental revision. Watch MMMU, GPQA, and agentic benchmark scores from Chinese labs in Q2 and Q3 2026 as the key data points for this assessment.

America is spending 23 times more than China to build AI, and the scientists who determine whether that spending becomes capability are choosing, at historic rates, to be somewhere else.

---