Who should control powerful AI? Concentration-of-power worries across the labs
September 29, 2026 · 22 min read
Leaders from Anthropic, OpenAI, Perplexity and Meta, plus the DeepSeek analysts, all raise the risk of a few companies or governments controlling compute and AGI, but propose very different safeguards. The shared worry shows this is a mainstream industry concern, not a fringe one.
The rules, numbers and steps worth keeping in front of you.
- Leaders of rival labs and independent analysts all name concentration of power as a serious AI risk.
- Dario Amodei: AI increases the total power in the world, so who holds it matters.
- Amodei: frontier models are heading above $10 billion, with $100 billion clusters targeted by 2027.
- Amodei's if-then scaling policy: stricter safety and security rules apply when a model crosses a capability threshold.
- Amodei: voluntary commitments cannot simply be trusted, but badly designed regulation creates backlash.
- Sam Altman: no single person or company should control AGI; his board legally could fire him but in practice could not.
- Altman: governments should set the rules; open-source some things and not others.
- Aravind Srinivas: regulate application and access to compute, not weights.
- Yann LeCun: proprietary control is the danger; open-source models let communities build their own systems.
- LeCun's question for any rule: whose jobs or market position does it protect?
- Dylan Patel: leading-edge chip research happens in only three places, and almost every product depends on TSMC.
- Patel: the cost of GPT-3-level intelligence has fallen about 1,200 times in roughly three years.
- Nathan Lambert: running open weights on your own computer keeps control of your data.
- Lambert: open models still need expensive compute and expertise to build on.
This is the short form. The full insight has every expert's reasoning, their own words and the sources.