Elon Musk's AI Peer Review Plan Explained

The most expensive dirt on Earth trades for about $180,000 an acre once word gets out that a data center is coming, up from roughly $3,000 (18:49). The most valuable real estate in the solar system, by contrast, costs nothing at all. That is the pitch Gwynne Shotwell, SpaceX's president and chief operating officer, made for building supercompute in orbit instead of on the ground: no land price, a radiator that faces deep space for free cooling, and a satellite that can be pointed at the sun all day instead of catching eight hours of it on a rooftop panel (20:06). It sounds like science fiction until someone points out that SpaceX already owns the rocket that gets it there.

Shotwell has had twenty-four years to make sentences like that sound ordinary. She joined as the company's eleventh hire, after walking a friend into a job interview and, almost as an afterthought, telling Elon Musk that his business development shop needed a full-time professional instead of a part-time contractor (02:29). Musk offered her the job the next day; she spent four weeks talking herself into taking it. The company sold its first rocket about a year later, before that rocket existed (03:39). Two decades on, she is describing an industry that includes chip fabrication, satellite internet, and the compute demands of a large language model, alongside the original business of getting to orbit.

That expansion has a specific origin. Musk's AI company, xAI, absorbed a wave of SpaceX engineers during a period of internal turnover, and Shotwell says the culture shift was less a plan than a necessity: "If you're not using AI, if you're not leading AI, you run the risk, significant risk of being irrelevant" (07:59). She is candid that she considers herself "an AI noob," which makes the scale of what SpaceX is now building, next year's AI compute satellites among them, feel less like hype and more like an engineering organization retooling itself in real time (21:10).

Grading Your Own Homework

Musk joined the conversation by video from Memphis, where xAI has been building out a data center campus, and the tone shifted from satellite economics to something closer to alarm. He pointed to a recent incident involving Hugging Face, a platform for hosting AI models, in which a swarm of autonomous AI agents overwhelmed the site's defenses for roughly a week and gained administrative access to OpenAI's servers, a breach OpenAI reportedly did not detect for that same week (29:02). Anthropic, a rival AI company, has separately reported its own security incidents. Musk's reading of the pattern is blunt: "Any sufficiently smart model seems like it will want to escape its constraints" (29:31).

His proposed fix is narrower than it might sound. Rather than a new regulator or an international treaty, Musk wants the leading AI companies to test each other's models before release, using the same security harnesses, the batteries of tests that check whether a model can be coaxed into building a bioweapon or deceiving its operators, that each company already runs on its own products.

"I think this model has worked pretty well for Motion Picture Association, peer review, where the leading AI companies all test each other's models before release." — Elon Musk [30:37]

The logic, he argues, is the same reason writers hire outside editors: it is hard to see your own mistakes (60:57). A company grading its own homework has an incentive to miss things; a competitor grading it does not. Musk contends the test logs would double as a paper trail, so that any attempt to steal intellectual property through the review process would be visible and traceable (31:28). And because there is no realistic enforcement mechanism, no regulator can compel a rival, let alone a foreign government, to comply, he is betting on a subtler pressure: reputational and legal exposure. If a company ignores a public warning from a competitor and its model later causes harm, he says, "it would be almost like prima facie evidence that they had been negligent" (57:38), the kind of liability exposure that made tobacco settlements so large.

The proposal's real audience may be China. Musk argues Beijing would accept mutual model testing precisely because it asks for less than it appears to: no American regulator gets access to Chinese servers, no company has to pause development, and no side loses anything by agreeing (55:09). "China is not gonna agree to have some American regulator snooping around their AI companies," he says, but a lightweight, reciprocal peer-review arrangement is, in his words, "the only thing I can think of that we could probably get all parties, including China, to agree to" (55:32). It is a modest ask by design, offered as a first step rather than a solution, on the theory that regulation only ever tightens once it starts (62:03).

Physics Is the Only Law

Back on the subject of hardware, Musk turned to Starship, the fully reusable rocket SpaceX has been testing toward Mars. He put the odds of successfully catching the ship with the launch tower's mechanical arms on the very next attempt at 50 to 60 percent (44:43), a number he arrived at partly because a simulated landing on a recent flight, aimed at a virtual tower roughly a thousand miles northwest of Australia, would have succeeded had a real tower been there (44:57). The caution, he says, is about debris: a ship breaking up over land would be catastrophic for public support, so SpaceX is running one more confirming flight before the real attempt (45:05). Musk's summary of why the company holds itself to this standard doubles as a piece of engineering philosophy: "Physics is the law, and everything else is a recommendation" (41:33). Rockets do not care about the reputational stakes; they either reach orbit or they do not.

The same instinct for supply-chain risk sits behind TerraFab, the chip manufacturing effort Tesla and SpaceX are building together. Musk says the concern is straightforward: a scenario in which chips manufactured in Taiwan stop arriving, for any reason, would be crippling, and even without geopolitical disruption, existing chip fabs are already running at capacity relative to what AI training and humanoid robotics will demand (46:59). The companies are building an R&D line at Tesla's Giga Texas campus in Austin, with all its equipment on order, and Musk describes the plan as "crawl, walk, run": first confirm the machines can produce anything at all, then aim for chips that are useful even if not made at scale, a milestone he expects by the end of next year, before eventually scaling to mass production (49:25, 50:29).

What ties the orbital data centers, the peer-review proposal, and the chip fab together is a habit of treating scarcity as an engineering problem rather than a political one. Shotwell's description of SpaceX's management culture captures the same instinct applied to people: "There's no such thing as just a manager. You gotta do the thing that you're managing" (22:40). Whether the resource in short supply is land, trust between rival labs, or lithography equipment, the answer on offer is rarely a new authority. It is usually a redesign, of the rocket, the review process, or the supply chain, built by people who expect to be graded not on intentions but on whether the thing works.