This week's most worthwhile clash in the AI circle came from X and a podcast. Anthropic's head of safety Sholto Douglas and Gavin Baker, founding partner of Atreides Management — which manages $4 billion in assets — are sparring. On the surface they're discussing AI governance; underneath, it's a hard reckoning on compute, power, and chips.
What this is
Douglas threw out an "offense-defense balance" framework: in cybersecurity, defense can keep pace with offense, so distributing AI is good. In biosynthesis, building a pathogen costs $10,000 while developing a vaccine costs billions — an offense advantage that will persist at least into the 2030s. He paraphrased CEO Dario Amodei: "We may be the only company left in the world." Baker pushed back: concentration doesn't solve the problem, it just swaps "capability diffusion" for "accountability for power," and creates new single points of failure (one node failing collapses the whole system). His proposal: treat AI like electricity, with costs approaching the cost of energy.
Both agree: physical bottlenecks determine the endgame. Right now, only four hyperscalers — Google, Meta, Amazon, and Microsoft — plus Coreweave, Crusoe, and SpaceX, totaling seven companies, can deploy more than 500 megawatts of power internally in a year. TSMC has refused to double its advanced chip capacity. These aren't market choices — they're physics.
Industry view
Those backing Anthropic's framing argue that before superintelligence arrives, letting a small number of vetted entities hold frontier models is safer than capability diffusion. But Baker's counterquestion cuts: "Who is this company accountable to? Shareholders? Governments? International bodies?" Politically, centralization is nearly impossible.
Baker backs his case with capital: he's overweight on power, cooling, and optics companies — priced for an AI demand explosion — while Nvidia and memory makers trade at lower valuation multiples. Capital votes with its feet, calculating who can squeeze whose neck.
A risk worth flagging: neither side resolves one contradiction — technology diffuses far faster than governance updates. Cybersecurity has roughly a two-year hardening window; in bio, offense advantages persist into the 2030s. If frontier players hold both models and compute, market mispricing may only be the start.
Impact on regular people
For enterprise IT: in the next few years, your company's cloud and AI capabilities will almost certainly come from no more than eight vendors. Bargaining room will narrow.
For individual careers: the ceiling of the AI tools you use (ChatGPT, Claude, Gemini, Wenxin, etc.) is effectively set by five to six labs that control compute and frontier models. What they open up, what they charge — that directly shapes your work productivity.
For consumer markets: if AI can truly become as cheap and ubiquitous as electricity, the surplus eventually trickles down to every user. But in the short term, concentration of power is more likely to push AI service prices up than down.