Anthropic Is Calling for More Caution – and Locking In $518 Billion of ComputeVIEW IN BROWSER | FOLLOW LUKE ON 𝕏 
A few weeks ago, Anthropic’s Dario Amodei asked the AI industry to slow down. Last week, the company put billions behind the idea. Anthropic chose Accenture to place outside evaluators inside its lab, where they will test models, examine safeguards, assess alignment, and watch how frontier systems are developed. Anthropic and Accenture each expect to invest at least $1 billion over the next five years. That is a remarkable amount of money for a job that barely existed inside frontier AI labs a month ago. And it sharpens the investment question behind Amodei’s so-called “Oppenheimer moment.” Does greater caution shrink the AI buildout – or create another layer of it? The Oppenheimer Comparison
In June 1945, before the first atomic bomb test or the bombing of Hiroshima, J. Robert Oppenheimer joined a panel of scientists that supported using the bomb against Japan. But the panel acknowledged that scientists disagreed. Knowing how to build the bomb, they wrote, did not make them uniquely qualified to decide how it should be used. After the bombings, Oppenheimer and his colleagues warned that having the most advanced weapons would not necessarily keep America safe. In 1946, he helped develop a proposal to put atomic energy under international oversight. That’s where the comparison with Amodei becomes useful: How much say should the people building a powerful technology have over its future? And who else deserves a seat at the table? In We Must Pace the Frontier, Amodei calls for independent reviewers to work inside AI companies, for leading labs to coordinate their efforts, and for governments to get more involved. Under Anthropic’s proposal, outside reviewers would get a close look at the company’s work and could publish their findings, with limits to protect confidential information and security. There is an important difference, though. Oppenheimer helped build a weapon through a government-run program. Amodei runs a private company and is asking for more outside oversight. Calling that “handing off responsibility” assumes more than we know. Amodei is proposing changes to who can examine and influence AI development. That alone doesn’t tell us he is trying to escape responsibility for what his company builds. The Scientists Speaking From InsideAnother parallel involves the people doing the work and their concerns about where it might lead. In July 1945, Leo Szilard and fellow scientists petitioned President Harry Truman, arguing that the United States should not use atomic bombs before giving Japan clear surrender terms and a chance to accept them. They also warned that America’s decision would set an example for how other countries might use these weapons. Last month, Anthropic researcher Jacob Coxon publicly announced his resignation and accused Anthropic and OpenAI of “gambling with our lives.” In an interview with WIRED, he described colleagues’ concerns about how quickly AI was advancing and the pressure to keep up with competitors. In both cases, the warnings came from people who knew the work firsthand. That gives us a reason to listen. It doesn’t mean every danger they foresee will come to pass. The timing does not prove Coxon’s resignation prompted Amodei’s essay. Amodei had already raised similar concerns in his January essay, The Adolescence of Technology, and the latest proposal continues an argument he has been making for months. The biggest difference between Oppenheimer and Amodei is what had already happened when each man spoke out. Oppenheimer’s postwar push for oversight came after atomic bombs had devastated Hiroshima and Nagasaki. Amodei is calling for action to prevent future harm from increasingly powerful AI, while also responding to problems already reported. The comparison helps us think about the responsibilities of people who build powerful technologies. It does not mean the consequences are the same. AI Safety Is Becoming a BusinessHistory will eventually form its own opinion about Amodei’s motives. Investors have something more concrete to examine right now: the market forming around his proposal. Anthropic’s Accenture agreement turns embedded evaluation from an idea into a commercial service. Faculty, Accenture’s specialist AI unit, will work inside Anthropic to red-team models, assess alignment, test safeguards, and follow the decisions that shape how those systems are trained and released. That creates paid work around the model itself: evaluation teams, monitoring tools, security systems, testing environments, audit trails, and the computing capacity required to run all of it. One concern is that expensive safety reviews and complicated rules could help the biggest AI companies hold on to their lead. Those companies have the money and staff to meet new requirements. Smaller rivals may struggle to keep up. The OECD identifies complicated regulations as a potential obstacle for new competitors. It also notes that safety and certification rules can determine which companies are allowed to serve certain markets. But oversight can also help competition. Making AI systems easier to inspect and easier to use together could give customers more confidence and more choices. The details will decide who benefits. Which companies have to comply? How much will compliance cost? And will the new rules open regulated markets to more providers or raise the price of entry? A proposal can address a real safety concern and benefit the company promoting it. Both can be true. For investors, the business effects deserve attention. Guessing what’s on a CEO’s conscience won’t tell us much about future earnings.
Slower Development Doesn’t Mean Demand StopsHere’s where this debate becomes especially useful for investors in AI. AI needs computing power for two main jobs. Training is how developers build and improve a model. Inference is what happens when someone puts that model to work – asking a question, writing code, reviewing a document, or completing another task. Finishing the training doesn’t end the need for computing power. Every time someone uses the model, computers have to do more work. Deloitte’s 2026 outlook projected that running AI models would account for roughly two-thirds of AI computing, up from about half in 2025. That’s a forecast, but it shows how much demand could come from using the technology already built. A company can put an existing AI model to work in more departments while the next version goes through safety testing. Developers can create new products using capabilities already available. All of that still needs servers, memory chips, networking equipment, cooling, and electricity. Google’s new Gemini 4 Argon gives us an early look at what pacing may mean in practice. Google developed a powerful new frontier model and initially released it to a limited group of trusted cyber defenders. The company is collecting feedback, refining guardrails, and participating in a voluntary government pre-release process before opening Argon to developers, businesses, and consumers more broadly. Development continued. Access is widening in stages. That’s the basis of my investment case for AI infrastructure: More people using today’s AI can keep demand growing, even if tomorrow’s AI takes longer to arrive. Training is where a slowdown would bite first. Amodei says the industry should consider limits on the computing power used to train models, the training process itself, and the use of AI to improve AI. His proposal goes beyond making companies wait longer to release a finished product. Limits on training could affect equipment orders. Delayed releases could also hold back applications that need abilities today’s models don’t have. Extra safety testing and monitoring would require some computing power, too. But we shouldn’t assume that work would make up for everything delayed or canceled. The investment question is whether growing everyday use outweighs any slowdown in development. Anthropic’s checkbook tells us where it thinks this is headed. According to a confidential IPO prospectus reviewed by Reuters, the company expects to commit at least $518 billion to computing infrastructure over the next decade. Roughly 80% of that amount is either noncancelable or payable regardless of how much capacity Anthropic ultimately uses. Anthropic told prospective investors that access to compute could become the main constraint on its future growth. In other words, the company making the loudest case for caution is also locking in enormous amounts of compute well into the 2030s. Anthropic expects the release process to become more deliberate. Its infrastructure base is still getting much larger. What Would Change My ViewI’m watching what businesses actually do: how much they plan to spend, whether they keep ordering equipment, how much of their computing capacity they use, and whether more customers are paying for AI. If customers cut spending plans, that matters. If businesses slow their adoption of AI, that matters. If chip orders weaken, we need to understand why. But if companies keep finding useful ways to put existing AI systems to work, demand for the equipment supporting those systems can hold up even as development slows. Government action now belongs on that list, too. The Federal Trade Commission has opened an investigation into OpenAI, Anthropic, and other AI developers over possible risks to consumers. The agency has not disclosed the full scope of the inquiry, but its involvement shows that the debate is already moving beyond voluntary commitments from the labs themselves. A defined testing window would be manageable. An open-ended approval process would be more concerning. I would also reassess the thesis if regulators begin limiting chip purchases or computing capacity directly, if safety reviews materially delay enterprise deployments, or if customers start canceling the infrastructure contracts already supporting this buildout. None of this makes every AI stock a good buy at any price. A business can grow and its stock can fall. If investors paid a price that assumed much faster growth, even solid results can disappoint. Shares can also drop well before a slowdown shows up in reported sales. I still see a strong long-term opportunity in AI infrastructure, with those conditions in mind. The case rests on more customers finding useful, valuable things to do with AI. It doesn’t depend on every lab releasing its next model as quickly as possible. The Final WordPerhaps this is Amodei’s Oppenheimer moment. History will judge that through his decisions and their consequences. For investors today, the more immediate question is, are customers continuing to find valuable work for the machines already running? Elon Musk has spent two decades betting that the answer will be yes. I call that bet XPANSE. Musk has attached a 1,000-fold upside case to the project. That number grabs attention. But I’m more interested in what has to be built before anything close to that scale becomes possible. Every foundational layer has to expand together. And each one runs through vulnerabilities serious enough that a senior government official has warned of an “economic apocalypse.” I’ve spent months tracing those pressure points and the public companies positioned around them. I’ve pulled that work into a new briefing: a map of the bottlenecks, a three-part approach to the opportunity, and one investment whose name and ticker I’m giving away for free. Smarter AI can only go as far as the infrastructure beneath it. XPANSE is the wager on how far that buildout can go. Explore the full investment case right here. Sincerely, |
Kevin Warsh's Federal Reserve recently did something the central bank hadn't done in more than three years...
