Compute, Capital, and the Concentration of AI Power
Frontier AI requires vast capital, specialized hardware, and rare talent. Each input concentrates capability in a small set of firms and states — with major implications for democratic accountability.
Key facts
- Fewer than a dozen organizations have trained frontier-scale models as of 2025.
- NVIDIA supplies the majority of frontier AI training accelerators.
- OpenAI, Anthropic, and Google DeepMind dominate cumulative frontier AI investment.
- US chip export controls are the most consequential AI governance instrument enacted to date.
Compute Concentration
Training frontier models requires tens of thousands of advanced GPUs and hundreds of millions of dollars in compute. As of 2025, fewer than a dozen organizations globally have run frontier-scale training jobs.
Cloud providers (AWS, Microsoft Azure, Google Cloud) and chip designers (NVIDIA in particular) sit at the center of this bottleneck.
Capital Concentration
Frontier AI development is increasingly capital-gated. OpenAI, Anthropic, and Google DeepMind have raised tens of billions cumulatively. Smaller labs and open-source efforts face widening capability gaps.
Geopolitics
US and China dominate frontier AI investment. EU, UK, and Gulf states have launched sovereign AI initiatives partly to avoid full dependence. Export controls on advanced chips have become a defining instrument of techno-economic strategy.
Democratic Questions
Concentration raises classic political-economy questions: who decides which systems are built and deployed, who captures the gains, and what oversight is meaningful at this scale.
Proposals range from antitrust enforcement to public AI infrastructure to mandatory pre-deployment evaluations.
Frequently asked
Will open-source close the gap?
+
Partly. Open-source models trail frontier closed models by 6–18 months on most benchmarks, though the gap fluctuates.
Is AI concentration like past tech concentration?
+
It is more concentrated than the early internet, more capital-intensive than mobile, and more strategically consequential than either.
Sources & further reading
Continue in this series
Risk Overview
A Taxonomy of AI Risks
Fairness
Bias and Fairness in AI Systems
Privacy
Privacy in the Age of AI
Information Integrity
Deepfakes, Synthetic Media, and Trust
Surveillance
AI-Powered Surveillance
Security
AI in Warfare and Autonomous Weapons