The collapse of the initial public offering (IPO) by the Australian AI company Firmus last week was the result of several factors. They included the lack of a clear perspective by its founders, “heroic” assumptions about its rate of growth, the prospect that investors would be taken for a ride as existing shareholders dumped the stock soon after the IPO to realise a quick profit, and concerns over the level of debt that would be needed.

But the demise of the IPO, which at $A44 billion was to be the biggest float since the privatisation of the telecommunications firm Telstra in 1997, has implications which go beyond the shores of Australia. While there have been hasty reassurances that the Firmus debacle does not mean the AI bubble is bursting, there are signs pointing in that direction.

Last week began on a high note for the company’s chief executives, Tim Rosenfield and Oliver Curtis. For Curtis it was going to be the ultimate redemption story—from so-called jailbird to billionaire. Curtis was sentenced to two years jail in 2016, of which he served a year, for insider trading.

Last Monday the firms running the float—JP Morgan, Bank of America and Morgan Stanley, together with the Australian firm Morgans—said that demand at $11 per share was “well in excess” of what was on offer. A Morgans adviser emailed clients to say that the demand had been “enormous.”

By the middle of the week, however, there was discussion about reducing the price to $9, amid concerns that, as it stood, the IPO would allow big shareholders who had bought in at a much lower price to cash in when it went public, leading to losses by investors who had bought in at a higher level.

The result of these and other concerns was that when markets opened last Friday, Firmus announced that the float would not proceed.

Citing “recent market volatility” and prevailing conditions, it said the board had concluded that “proceeding with the offer was not in the best interests of the company and its shareholders.” Instead, it would pursue capital from private markets.

How far that goes remains to be seen because there were many so-called “red flags” in the offer—one analysis listed as many as 30. As one comment on the collapse noted, the company had asked investors to “buy a skyscraper off the plan” but for once the market “asked to see the foundations first.”

Yet the essential point, which goes to the heart of the AI boom, is that the Firmus operation, while it may be considered an extreme case, was not fundamentally different from what has been going on in the rest of the AI universe.

The Firmus business model is based on renting out high-tech Nvidia chips to customers such as Open AI and Meta for AI processing. Its specific pitch was that it had developed a more efficient water cooling system to significantly cut costs. And in another example of the circular arrangements that characterise the AI industry, Nvidia is one of the largest existing shareholders in Firmus.

Firmus was not short of boosters. Before the IPO went south, Shaun Weick, the portfolio manager at Wilson Asset Management, said Firmus was the “BHP of the token generation” at the “bottom of the cost curve globally.”

“AI is the most disruptive technological advance in history,” he said. “We view it as an arms race driving a once-in-a-generation capex boom, and Firmus is at the centre.”

Others were having doubts, however. CDC Data Centres has cancelled its collaboration with Firmus to construct $73 billion worth of data centres in Australia. Firmus said it was going ahead to construct data centres in Tasmania in addition to the one it has established there.

The funds from the IPO were to be used to build a network of data centres in southeast Asia to cash in on a future Asian AI boom. But so far these are just plans on the drawing board.

Jun Ben Liu, the portfolio manager at Ten Cap, which describes itself as a boutique investment management firm, told the Australian Broadcasting Corporation that she was steering clear of the Firmus IPO because of the “unprecedented lack of detail” and enormous structural risks, adding that 97 percent of the data centre capacity Firmus had promised had yet to be built.

“They might deliver cheaper earnings in two years if they build what they promise, but that’s speculative,” she said.

In the run-up to the IPO proposal, the company’s valuation had surged from $A6.9 billion in April to $A15.5 billion. The IPO would have almost tripled that valuation.

Others have been concerned about the level of debt. The model of the so-called neo-clouds like Firmus is to borrow money against the contracts they have with major AI companies to buy chips and then use the rent from those chips to pay back the loans.

In the case of Firmus, however, there were concerns about the debt levels involved. Firmus was expected to carry $US30 billion in debt in 2028, six times its forecast of $US5 billion in operating earnings.

While the Firmus IPO failure has certain specific features, there are indications that it is part of an international trend. Gaurav Gupta, involved in research at the global technology analysis firm Gartner, told the Australian Financial Review that the “sceptical market reaction” wasn’t out of the ordinary. Recent IPO filings from the UK-based neocloud Nscale met with similar polarised sentiment.”

Neoclouds rent out time, using high-powered chips to train AI models.

In a warning of the instability of this area of the AI boom, Gupta predicted that cost pressures and the lack of a differentiating business model would mean that 80 percent of neocloud operators would cease independent operations by 2030.

Bloomberg pointed to the broader implication of the Firmus IPO failure, saying it was “revealing fresh cracks in the AI funding boom” and underscored “the growing concern over how much capital AI infrastructure companies are demanding from public markets at a time when borrowing costs are rising.”

It cited the comments of one investment manager who said that while investors still believed in AI, they were not prepared to “pay any price for companies that spend huge amounts on data centres, depend on a few big customers and promise profits years from now.”

Noting that the warnings about inflated valuations are getting louder, Bloomberg pointed to a projection by Bain & Co that the AI industry would need to generate $6 trillion in annual revenue by 2031 to justify the massive amounts of capital being deployed to build data centres. Placed within this context, the Firmus IPO collapse is a pointer to what is building up within the AI boom.