The maker of Claude wants to do to healthcare what it did to programming. Its crazy prospectus shows why it needs to.
Anthropic’s prospectus, filed last week ahead of a planned Nasdaq listing under the ticker ANT, contains some of the most extraordinary numbers in corporate history.
Healthcare is in there as one way it hopes it can make those crazy numbers work.
The crazy numbers
Revenue was US$41 million in 2023, US$386 million in 2024 and US$4.6 billion in 2025. In the first half of 2026 it was US$16.3 billion, 14 times the same period a year earlier.
But that growth is coming off a base that barely existed two years ago. It is already slowing and there is credible expert analysis that says the models being built are running out of revenue runway fast.
The losses are wild.
Anthropic booked an US$80.7 billion net loss for the first half, though about US$72 billion of that was a non-cash accounting charge on its convertible notes.
The committed cost
But the real problem is the commitments to spending in the near term.
The amended filing discloses about US$895 billion in future spending, almost all on computing power and data centres, with deals across Broadcom, Google, SpaceX’s AI unit and Amazon.
About US$70 billion falls due in the next 12 months and nearly US$450 billion by the end of 2030.
Annualise the September quarter and Anthropic is earning about US$68 billion a year. Covering roughly US$110 billion a year in commitments to 2030, on top of salaries and everything else, implies revenue more than tripling, to something north of US$200 billion a year, within four years.
The filing itself warns of “significant cash outflows” that could weaken its finances.
Then there are the 92 pages the prospectus spends on risk. Ninety-two pages. It’s like, don’t bother reading it all, you’re either a believer in our “look into our eyes” sales pitch or not. Move on.
The market and data centre problem
Although it’s identified in the 92 pages of risk, there’s a huge emerging problem in Anthropic’s forward commitment to the build of data centres.
Do any search on any major “neo cloud” (meaning the “new cloud” of AI data centres) business and they all have the same problems:
- Massive and building community opposition;
- A switch from net zero power commitment to gas intensive in reality with a huge carbon footprint;
- Workforce shortages;
- Circular financing that looks awfully like it will come tumbling down around their ears pretty quickly;
- A jittery market: the biggest neo-cloud, CoreWeave, has lost about half of its value from its relatively recent peak, our local Firmus play pulled its IPO within days of opening its books and the UK’s biggest player Nscale has lost Microsoft from its flagship UK site and is in serious trouble trying to get its major US project Monarch off the ground with massive community and state government opposition; and,
- Well documented over-promises and under-deliver so far, including huge miscalculation of regulatory and construction costs.
On top of this big unfolding issue, and as a part of it, the market for AI is finally starting to get the jitters big time.
Related
The wary punter will understand that the public markets have started pricing all this risk in, but the private credit markets haven’t and a lot of commentators are very worried about the cost of private capital now and what will happen if the AI financial dream starts to unravel.
Why healthcare
The filing describes extending the agentic architecture behind Claude Code, the tool that turned programming into Anthropic’s biggest growth engine, to knowledge workers generally, through “applications purpose-built for specific professions and industry verticals”.
It names healthcare and life sciences among the domains it plans to reach.
Anthropic says nine of the top 10 big pharma companies use Claude, and it has built its own molecular biology lab and acquired Coefficient Bio.
In January it launched Claude for Healthcare, a HIPAA-ready offering with tools for medical coding, prior authorisation and coverage checks, and with US customers including Banner Health and Stanford.
A crowded space now
OpenAI launched ChatGPT Health in January, saying about 230 million people already ask ChatGPT health questions every week, and in April it made ChatGPT for Clinicians free for verified US doctors.
Epic is building AI directly into the record systems that most large US hospitals already run, and Microsoft, Google and Amazon are all in the market.
Health is also much harder than helping healthcare coding.
As FHIR creator Grahame Grieve told last week’s HSD summit, software has automated tests, so AI output can be checked in some way. Clinical work mostly doesn’t have this luxury.
Regulators are circling too: the TGA has found most AI scribes it reviewed are likely medical devices. And as the Heidi/Best Practice fight shows, access to patient data still runs through incumbents who control the records.
Programming took off for Anthropic because developers could adopt Claude themselves, in days, with little red tape.
Healthcare offers few of those shortcuts. Yet the filing’s numbers suggest Anthropic needs markets this size to pay its compute bill.