Stanford HAI released its 2026 AI Index on 13 April, and the headline could not be starker. The performance gap between the top US and Chinese frontier models has collapsed to 2.7%. Three years ago, that gap ranged from 17.5 to 31.6 percentage points across major benchmarks. American and Chinese labs have traded the lead multiple times since early 2025, and DeepSeek-R1 briefly held the global crown in February.
The report finds that the number of AI researchers and developers relocating to the United States has dropped 89% since 2017, with an 80% fall in the last year alone. US private AI investment reached $285.9B in 2025, more than 23 times China’s $12.4B, yet that capital advantage is not translating into a durable capability lead. 84% of students now use AI tools, while jobs, infrastructure, and policy are visibly struggling to keep pace. The Foundation Model Transparency Index also dropped to 40 from 58 a year ago, the first material decline since it was established.
For the last five years, the western AI story has rested on two assumptions: American labs would stay multiple benchmark points ahead, and global AI talent would keep flowing into the US. Both assumptions now look fragile. Chinese labs are shipping open-weight models that match proprietary US systems on most tasks at a fraction of the compute, while US immigration friction hollows out the country’s historic talent pipeline. The AI race is no longer a two-horse sprint with one horse four lengths ahead. It is a photo finish, and one of the horses is running on Huawei chips.
The US AI lead is not just narrowing, it is being redefined. For the first time, the Stanford Index frames the competition not as a benchmark gap but as a systems race, where compute, talent flows, open-weight release velocity, and regulatory posture matter as much as raw model quality. The 2.7% headline is attention-grabbing, but the 89% collapse in inbound AI talent migration is the number that should keep US policymakers up at night. Capability gaps close. Talent pipelines, once reversed, take a decade to rebuild.
For enterprises, the practical implication is that open-weight Chinese models are becoming a credible strategic option rather than a fringe experiment, particularly for organisations operating outside the US. DeepSeek V4’s imminent launch on Huawei silicon, under an Apache 2.0 licence, turns that shift into a quarterly procurement decision. For policymakers, the window to turn capital advantage into policy advantage is narrower than last year’s Index implied.
The bottom line: the AI race has moved on from who has the biggest model to who can retain the people, ship the deployments, and write the rules. On each of those, the Index shows the US lead is softer than the conventional wisdom assumes.
From ‘Does the Model Work?’ to ‘Whose Workflow Is It Wired Into?’ This week’s stories share a single undercurrent: AI is no longer a story about which lab builds the smartest model. It is a story about which organisations, countries, and sectors can turn raw capability into compounding economic value. Stanford’s Index shows frontier-model performance gaps collapsing while US dominance softens. PwC shows 20% of companies racing ahead with AI-led revenue growth while the other 80% fall behind. Healthcare systems are building their own AI agents because patients are already using generic chatbots.
The competitive frontier has moved from ‘does the model work?’ to ‘whose workflow is the model wired into?’ and that is a very different race. The winners of the next two years will be the organisations that stop experimenting and start deploying AI inside the systems that generate their revenue.
Anthropic launched Claude Design in research preview on 17 April, a new Opus 4.7-powered workflow for generating design systems, website prototypes, slide decks, one-pagers and interactive mockups from plain-language prompts. Figma shares slid on the news.
What makes Claude Design different is that it can read a team’s existing codebase and design files to apply in-house visual styling to every artifact it produces. Exports include Canva, PDF, PPTX and standalone HTML. Available on Pro, Max, Team and Enterprise plans today. Canva and Anthropic have partnered before, so the bigger question is whether this is a new front in the AI-design wars or a tide that lifts all boats. Expect Figma to respond before the end of Q2.
Read the coverage →