The U.S. still leads in AI model releases, but China just installed nine times more industrial robots than America did last year.
The Summary
- Stanford's 2026 AI Index report shows U.S. companies released 50 notable AI models in 2025 versus China's growing output, but China deployed 295,000 industrial robots compared to America's 34,200.
- Industry now produces over 90% of notable AI models, up from 50% in 2015. Academia is effectively out of the game.
- Local U.S. governments are blocking data center construction while Chinese factories are installing the physical infrastructure for the agent economy at scale.
The Signal
The divergence is stark. America writes the code. China builds the machines that run on it.
Stanford's Human-Centered AI center just dropped their 400-page AI Index report, and the most telling split isn't between models and capabilities. It's between software ambition and hardware deployment. The U.S. released 50 notable AI models in 2025 according to Epoch AI's tracking. China's catching up on model releases, but that's not the story. The story is 295,000 industrial robots installed in Chinese factories in 2024. Japan installed 44,500. The U.S. installed 34,200.
That's not a gap. That's a different game being played.
"China installed nine times more industrial robots than America in 2024."
Here's what those numbers mean in practice:
- Every Chinese robot represents a production line being prepped for AI-driven optimization
- Every installation is infrastructure for agents to control physical processes
- Every factory floor becomes a testbed for multimodal models that need to see, move, decide
Meanwhile, U.S. local governments are blocking new data centers. The report notes resentment toward AI "boiling over" in America, with restrictions and outright bans on development. You can't run the agent economy without compute infrastructure. You definitely can't run it if your political climate treats new data centers like toxic waste sites.
The industry consolidation trend tells another story. Over 90% of notable AI models now come from companies, not universities. In 2015, it was 50-50. In 2003, academia owned the field completely. This isn't just about funding. It's about the death of open research in favor of proprietary advantage.
Key shifts in the AI development landscape:
- Industry releases outnumbered academic/government releases 87 to 7 in 2025
- Notable models are now defined by commercial viability, not research novelty
- The gap between "can build a model" and "can deploy at scale" is widening
OpenAI and Anthropic are racing toward IPOs this year. That will accelerate the trend. Public markets demand moats, not papers. They reward deployment speed, not reproducibility. The incentive structure now punishes sharing and rewards secrecy.
The Implication
Watch what gets built, not what gets announced. Model releases make headlines. Robot installations make products. If you're betting on the agent economy, bet on companies with physical infrastructure partnerships in manufacturing regions. The software will commoditize faster than the hardware deployment will scale.
For individuals: the robots going into Chinese factories aren't taking American jobs. They're creating the manufacturing capacity that will define which products exist in 2028. If you're building agent-based tools, think about what happens when your software has access to production lines that can iterate in days, not months.