Anthropic sells investors an AI future it has not demonstrated.
The company is meeting potential investors ahead of a planned initial public offering that could launch in September or early October, according to The Wall Street Journal. Those investors are asking the ordinary questions that appear less ordinary when the company is being described as a blockbuster: What happens when cheaper Chinese systems arrive? What happens when the Trump administration turns hostile? What happens when communities oppose the data centres required to run the business?
Anthropic’s answer, in part, is that it plans to push further into healthcare and biology.
That is not a business plan. It is an expansion of the promise surface.
To be fair, the investors are asking sensible questions. A company facing cheaper competitors, political conflict, infrastructure resistance, and the physical cost of computation should be asked how it intends to survive. The trouble is that “healthcare and biology” is being offered as a future market before the public has been shown the technical, financial, or institutional machinery that would make that market real.
The same pattern appeared in the earlier comparison of Anthropic’s and OpenAI’s federal problems: the companies are presented as separate corporate stories, but their investor narratives rely on the same underlying move. Present limitations are treated as temporary obstacles. Future applications are treated as evidence of present capability. The valuation arrives somewhere in the middle, carrying the confidence of a result that has not yet been produced.
This is the capability-gain bait-and-switch. A genuine improvement in a narrow task is used to imply a much broader competence. The narrow gain may be real. The inference is not.
Large language models generate outputs by predicting continuations from patterns learned during training. That can be useful. It can assist with bounded forms of drafting, classification, retrieval, coding, and analysis. But “useful in a bounded task” is not the same claim as “ready to serve as a dependable instrument across healthcare and biology.” The missing work is the specification: What exactly is the system supposed to do? Under what conditions? Against which baseline? With what error rate? Who checks the result? What happens when the model is wrong?
A system without a precise, externally defined specification cannot be formally tested against that specification. In cryptographic protocol verification, this is not a philosophical objection. It is the first line of the proof. You cannot establish that a system satisfies a security property if the property has not been written down precisely enough to test. Healthcare does not become less demanding because the product is marketed as intelligent.
Biology is not a single task. It includes image analysis, sequence analysis, molecular structure prediction, clinical documentation, trial recruitment, diagnostic support, laboratory automation, and a great deal of work that depends on instruments, samples, regulatory controls, and human judgment. A model that performs well in one of those settings has not thereby crossed into the others. Even a valid result must survive the stages between computational suggestion and clinical use.
The phrase “push further into healthcare and biology” compresses those stages into a market category. It turns a long chain of technical and institutional obligations into a slide in an investor presentation.
Kate Crawford’s account of AI is useful here because it puts the machine back into the material world: fuel, infrastructure, data, and human labour. The model is not an airborne intelligence waiting to be pointed at the next profitable sector. It is a system built from hardware, electricity, training material, evaluation labour, software maintenance, and contracts. Its performance depends on the conditions around it. Its costs do not disappear because the company has found a more prestigious noun for the next market.
The Chinese competition question exposes the same structure from another direction. If cheaper systems can perform adequately for many customers, then Anthropic’s problem is not simply that it needs a more impressive demonstration. It needs a durable reason for customers to pay more. That reason might be better performance, better security, better privacy, better reliability, better integration, or stronger accountability. “Healthcare and biology” is not itself any of those things.
Whatever the company’s messaging, opposition to data centres is not merely a communications problem waiting for a sufficiently reassuring announcement. Data centres require land, electricity, transmission capacity, water in some cooling systems, construction, and public approval. The opposition is attached to the physical infrastructure, not to a misunderstanding that can be corrected by calling the product transformative. The model runs somewhere. The bill arrives somewhere. The political conflict follows the wiring.
This is where the planned IPO matters. A private company can speak in the language of possibility. A public offering subjects the company’s claims and risks to more formal scrutiny than a private fundraising pitch, including the risks that promotional stories prefer to leave at the edge of the frame. Investors are not wrong to ask about rivals, regulators, and infrastructure. They are asking whether the company can convert technical promise into durable returns without transferring the costs to workers, users, communities, and public systems.
The answer cannot be that the company will enter every large market eventually. That is not diversification. It is a refusal to identify the bottleneck.
Anthropic’s stated future also raises the question of who will own the resulting infrastructure. If a model becomes useful in healthcare, will hospitals and researchers be able to move their data and workflows between competing systems? Will independent auditors be able to inspect the system? Will customers have a meaningful exit, or will they be locked into proprietary interfaces and long contracts? Will the public institutions that supplied the research environment receive access, accountability, or only another invoice?
Open standards and interoperability are not decorative policy preferences here. They are the difference between a useful tool and a new chokepoint. Antitrust enforcement can address concentration, but interoperability lowers switching costs before a merger case finishes moving through the courts. Public procurement can require portability, auditability, and documented performance. Privacy rules can give patients and individuals a remedy that does not depend entirely on a regulator choosing to act.
Those measures would force the business to compete on what its systems do rather than on what its investor narrative suggests they might someday do.
There is an honest version of the Anthropic pitch. The company has a product, it faces real competitive and political constraints, and it wants to find more valuable uses for the systems it has built. Fine. The product should be evaluated at the level of the task, the evidence, the cost, and the exit route.
The dishonest version asks investors to treat a list of future sectors as if it were proof of present capacity. A public offering is one way to raise money. It is not a substitute for a specification.