The biggest IPO collapse in Australian corporate history went largely unnoticed in health circles. It shouldn't have.
Last week, Firmus Technologies pulled what would have been the second-largest IPO in ASX history – a $44 billion float that collapsed within 48 hours of its book-build opening.
For most clinicians, this registered somewhere between background noise and irrelevance.
Firmus was an AI infrastructure company – a data centre business pitching itself as the backbone of the coming AI revolution.
It had raised billions from heavyweight investors, attracted enormous bank fees, and was being sold to the market at a $44 billion valuation. It had actual annual revenues of somewhere around $50 million, which means it was asking for a valuation of about 900 times its actual earnings.
The story of how it fell apart is a case study in how AI hype gets packaged and sold, which is handy to understand even if you’re in healthcare trying work out how AI is going to help and how it could go wrong.
Firmus had only two data centres operational, and both were troubled.
A significant partnership with CDC – another major data centre operator, not the Centres for Disease Control – had been quietly killed and was only disclosed at the last minute before the book-build, as was a new deal with Meta, sprung on prospective investors with no time to evaluate it.
US institutional investors, who were meant to anchor the float, looked at a commodity business in a market with millions of competing compute options (ways to access processing power) and decided the numbers simply didn’t add up. More than half the existing register would have been free to sell from day one. The bankers stood to make $200 million in fees. The whole thing was, in the Financial Review‘s understated assessment, not well bid.
Ed Zitron, the technology analyst and writer, has been making a related point about the broader AI industry for some time.
His argument is that the revenue figures being cited by the largest AI companies – the annualised run rates, the projected annual recurring revenue (ARRs), the numbers that move markets and fund valuation rounds – are largely made up in the sense that they mean almost nothing.
Anthropic, OpenAI and others have been citing “annualised run rates” pegged to non-specific 28-day windows, gross rather than net revenue, and projections built on the best possible slice of consumption data.
When you compare the actual quarterly revenues these companies have disclosed – OpenAI made $12.4 billion in the first half of 2026, Anthropic $15.5 billion – to the annualised run rates being shopped around at the same time, the gap is extraordinary.
The numbers exist, Zitron argues bluntly, to make investors think something is bigger and more inevitable than it actually is.
None of this means AI is going away. And it certainly doesn’t mean that the healthcare AI products you’re being shown demonstrations of aren’t genuinely useful.
Related
But the Firmus implosion is a useful reminder that the same atmosphere of hype, speed, and look-into-my-eyes certainty that characterises the big end of AI investment is operating at every level of the market, including healthcare.
When a company tells you it has an “annualised run rate” of $X billion, ask what that actually means and over what period.
When a healthcare AI product promises transformative productivity gains or tells you its pipeline is full of tier-one hospital clients, apply the same scepticism you’d apply to any other vendor pitch.
When a platform raises at a valuation that would make your eyes water and tells you it’s cheap relative to its peers, remember that Firmus was being sold as cheap relative to CoreWeave (a big league AI cloud company) right up until the moment it wasn’t being sold at all.
Big AI is a bubble.
That doesn’t mean it pops tomorrow, or that everything inside it is worthless.
Bubbles can run longer and further than anyone expects, and real value gets created inside them.
But they do eventually correct, and when they do, the companies that survive are the ones that had genuine unit economics underneath the hype.
In healthcare AI, as in every other sector touching this technology, buyer beware is not paranoia.
That’s the early-adopter response Heidi is betting on.
But this is healthcare. Can you actually do this in healthcare?
The answer so far: Heidi has gotten away with it up until now – first with their scribe, then with their evidence knowledge layer.
I have not met a doctor who doesn’t love it when they see it.
Heidi II is an escalation, a big one, into agentic AI, where those same agents that can be set hard to work scraping your incumbent PMS for patient data to run the sort of modern consult Heidi is pitching: no computer work, just talk to the patient and give instructions.
“Give the computer work back to the computers,” Dr Kelly likes to say.
When asked whether the incumbents could build this themselves, his answer is blunt: “There’s just no chance they’re going to be able to build safe agents in healthcare.”
It sounds arrogant but that doesn’t seem to be where he is at. He points out that Heidi’s entire focus is on agentic AI for a doctor, as a path to an interoperable and highly functional new clinical workflow layer.
