Back to AI Index 2026 overview

Chapter 01

Research and Development

Stanford frames AI R&D as an industrial system, not just a research race: frontier labs, hyperscalers, chip supply chains and environmental limits now shape what can be built.

Why it matters

This chapter explains why frontier AI progress is increasingly gated by disclosure, capital intensity, hardware concentration and energy demand.

How to apply it

Read it as a map of bottlenecks: who can train, who can disclose, who controls compute and who absorbs the environmental cost.

Core signals

Core signals

Two signals shape this chapter: frontier model production has become an industrial operation, and the hardware base sustaining it is concentrated in a small number of actors.

The frontier is industrialized

More than 90% of notable AI models in 2025 came from industry, confirming that cutting-edge development is now concentrated inside a small number of heavily capitalized actors.

Infrastructure concentration is strategic risk

The report ties model progress to a narrow hardware base: U.S. data-center dominance, Nvidia-heavy compute share and TSMC's central role in advanced chip fabrication.

Selected data points

Selected data points

Condensed numbers and comparisons pulled from the official Stanford chapter materials.

SignalValueContext
Industry share of notable models>90%Industry produced over 90% of notable AI models in 2025.
Notable models, U.S. vs China59 vs 35The U.S. produced 59 notable models in 2025 while China produced 35.
Global AI compute capacity17.1M H100-eqGlobal AI compute capacity has grown 3.3x per year since 2022.
U.S. data centers5,427The United States hosts more than ten times as many data centers as any other country.
Training emissions signal72,816 tCO2eStanford estimates Grok 4 training emissions at 72,816 tons of CO2 equivalent in 2025.

Implications

  • Model capability discussions increasingly need infrastructure and energy context.
  • Disclosure quality is becoming part of competitive positioning, not just research etiquette.
  • Open-source scale matters, but it does not erase concentration in compute and fabrication.