Dario Amodei’s Anthropic IPO Faces a Test Beyond the A.I. Market
Observer ·
Anthropic is betting that A.I. adoption will accelerate fast enough to justify its enormous infrastructure commitments. But many businesses, argues McMillanAI’s Jeff McMillan, are still building the systems needed to deploy it responsibly.
Anthropic’s leaked IPO prospectus offers an unusually detailed look at the economics behind the A.I. race. The numbers are striking: roughly $518 billion in computing commitments over the next decade, with about 80 percent of that amount reportedly owed whether customers use the capacity or not. Anthropic says the commitments are necessary because compute, rather than demand, is the constraint on its growth. A second question interests me more.
While Anthropic CEO Dario Amodei and his competitors race toward a public listing—and to secure the compute needed to build increasingly capable models—the customers they are all selling to are, for the most part, ill-prepared to deploy this technology at scale. Anthropic has laid out, in detail, what its models can do wrong. When one of them does make a mistake, the company that deployed it is the one that answers for the consequences.
Managing that responsibility will require changes to how companies build and run software, who tests it and what they expect from their vendors. The filing is therefore both a document about Anthropic’s economics and a warning about the operational infrastructure its customers will need to build if the company’s growth assumptions are to hold.
Reuters puts Anthropic’s computing commitments at $518 billion over a decade , with about 80 percent owed whether customers use the capacity or not. That averages out to roughly $50 billion a year, against annualized revenue of about $46 billion at the latest quarter’s pace . The company is reportedly seeking a valuation above $2 trillion, based in part on its own forecast of $190 billion to $200 billion in revenue in 2028, about four times its current annualized pace.
The valuation is in line with Palantir , SpaceX and Cloudflare , which trade at roughly 42 to 53 times this year’s expected revenue. Whether the revenue arrives depends on demand for compute and on how well Anthropic competes beyond model access.
The first risk is timing, and it has two sides. Anthropic’s assumptions have to match two things it does not control: when customers want the capacity, and when the capacity itself arrives. The prospectus says the business is limited today by the availability of compute rather than by buyers, which helps explain why Amodei has committed so aggressively to capacity. If the chips and data centers arrive late, Anthropic cannot serve demand and revenue falls short. If they arrive on time and the demand does not materialize, Anthropic is left paying for computers that sit idle.
The ultimate revenue number is probably realistic. My concern is how quickly adoption spreads beyond those early adopters and into the operational core of large enterprises. In many of the businesses I work with, A.I. deployment is still at an earlier stage than the spending by A.I. companies might suggest: basic use cases that older models already handle, a handful of agents and little or no infrastructure for running more consequential systems. The commitments Anthropic has made assume that those two curves—compute capacity and enterprise adoption—line up. A.I. use will almost certainly grow in the shape of a hockey stick. The question is where we are on the stick today.
The second, and even more important risk, is competition. High-quality, cheaper models are arriving every month. For most enterprise work, several models can do the job equally well, so the cheaper one wins. Anthropic’s answer is to sell more than model access: applications, agents, controls for managing them and forward-deployed engineers to deliver all of it. The other labs, the cloud providers and the large software vendors are building the same thing. Only a few will win the corporate and retail markets, and whether Anthropic is among them is an open question. Both risks ultimately depend on what Anthropic’s business customers do next.
Roughly 80 of the filing’s 261 pages describe what can go wrong with Anthropic’s technology. The company is effectively telling investors that capable A.I. will create risks that customers must be equipped to manage. The responsibility for a failure sits with the company that deployed the model. In 2024, Air Canada argued that its website chatbot was a separate legal entity responsible for its own statements, and the British Columbia Civil Resolution Tribunal rejected that and held the airline liable. The implication for businesses is straightforward: buying an A.I. system does not outsource accountability .
In my client work, the firms with agents already in production often have operational controls that are much less mature than the technology. Teams may be watching three to five agents by hand, with few have a written evaluation sets—real business tasks with agreed criteria for an acceptable answer—independent testing or monitoring built for the purpose. Anthropic’s disclosures raise two practical questions for those firms: how will they know when an agent is getting something wrong, and who has the authority to stop it? The prospectus assumes its customers are prepared to do both. In my direct experience, few firms are.
Answering those questions requires a different way to build, test and monitor this technology. Every A.I. project needs an evaluation set built by the people who understand the underlying work, an acceptance threshold set by the business and a written risk assessment with safeguards before anything reaches production. It then needs validation by a group independent of the team that built it, and that is where the work often stalls.
The people qualified to judge whether an A.I. system is producing acceptable work already have demanding jobs inside the business. Asking them to take on validation means deciding what other work they will stop doing. Most clients tell me they lack the resources or the skills. That is usually true, but that does not eliminate the responsibility.
My recommendation—and a more practical model—is to build the capability into the organization through in-depth training, rotations and part-time validation panels drawn from the business rather than permanent transfers into a risk function.
This also requires a new technical architecture for managing agents. Over time, companies will run thousands of agents making decisions and processing transactions, and no number of employees can watch that much activity by hand. Governance and controls will have to be built into the platforms themselves: software that checks outputs against business requirements, looks for anomalies, requires a person’s approval before any action that cannot be undone and escalates problems to a named person accountable for the agent. The concern is that companies will add agents much faster than they build the means to oversee them.
For investors, the prospectus comes down to two questions. The first is timing. Anthropic is betting that demand for advanced A.I. will grow quickly enough to justify infrastructure commitments that are largely fixed regardless of usage. The near-term risk is that enterprises are not ready to use that much capacity on Anthropic’s schedule.
The second is whether Anthropic can capture enough of the value created above the model layer. The company needs to sell Claude as well as the agents, applications and implementation capabilities that turn model intelligence into business processes.
Anthropic’s reported commitments include more than $111 billion with Google , $110 billion with Amazon and $31.4 billion with Microsoft , while Broadcom is also a major infrastructure partner. Amodei is therefore betting that A.I demand will be large enough to consume the compute he has secured, and that Anthropic will capture enough of the resulting enterprise value to justify the cost.
For a CEO driving an A.I. transformation, the question is different. Whatever Anthropic writes in its prospectus, the company deploying the technology assumes the operational risk. The question is whether the company has the processes, infrastructure and people to manage that risk responsibly: evaluation sets and independent validation, platforms with the controls built in and domain experts with the time and training to judge the work.
Vendor contracts need the same attention. Anthropic’s published model-deprecation policy gives customers at least 60 days’ notice before it retires a model. For many regulated or highly integrated businesses, that may be less time than the need to test a replacement, validate outputs and obtain internal approval. Contracts should give customers sufficient notice and technical information to manage those transitions. As models become embedded in workflows, agents gain access to internal systems and automated decisions begin to affect customers and employees, switching models becomes a governance problem as much as a procurement decision. That makes model portability, evaluation standards and exit plans part of the A.I. strategy.
Investors are being asked to fund the capacity. Business leaders will decide how quickly that capacity gets used, and whether their organizations can safely absorb it. Anthropic’s filing makes that first decision visible. The second will be made inside thousands of companies, as executives decide whether their organizations are ready for the systems they are being sold. The future of Anthropic’s economics may depend on the answer.