Firmus Pulled Its $5 Billion IPO. AI Infrastructure Just Got a Price Check

On the morning of October 9, 2026, bankers in Sydney closed the books on what was supposed to be Australia’s second-largest share sale in history. There was no champagne. Firmus, the Nvidia-backed operator of AI data centers that had spent weeks marketing a $5 billion initial public offering, announced it was walking away. Citing market volatility and conditions, the company said it would raise money privately instead. In one press statement, roughly $30 billion of paper valuation met the market, and lost.

The numbers explain why investors flinched. Firmus initially planned to sell shares at A$11 each, implying an equity valuation of about $30.6 billion. That price was nearly three times the $10.5 billion valuation the company reached in August, when Nvidia, Coatue Management, Blackstone and Jane Street put fresh money in. In about eight weeks, the same company asked public buyers to pay almost triple what its most sophisticated private backers had just paid. The public declined the invitation.

Behind the scenes, the pricing had already been slipping. By the time the books closed, the A$11 target had been cut to as low as A$8, with bankers scrambling to hold the line at A$8.25, according to the Australian Financial Review. At that level the raise would have shrunk from a headline A$7.9 billion toward A$5.9 billion, and the equity value from nearly A$44 billion to about A$33 billion. Bloomberg reported that the books closed without a clear indication of price or deal structure, the classic tell of a listing losing its anchor investors.

Two data centers, a $30 billion price tag

What Firmus actually owns today is much smaller than what it asked to be valued as. The company designs and operates what it calls modular AI factories, data centers built around proprietary energy and cooling technology and Nvidia hardware, and it currently has just two of them online: one in Melbourne, one in Singapore. Five more across the Asia-Pacific region are in early stages of development. The business plan is essentially a construction program financed by the capital raise, and the raise was priced as if the construction were already finished.

That gap between operating reality and asking price is what turned the listing into a referendum on AI infrastructure valuations. Firmus is not a mature business with predictable cash flow. It is a developer: it raises money, builds capacity, and signs long-term contracts with hyperscalers and AI labs to fill it. The model can work. It also requires enormous capital delivered in the right sequence, against deadlines involving power connections, cooling systems and chip deliveries. Miss any of those and the economics invert.

Jun Bei Liu, the co-founder of fund manager Ten Cap and a public critic of the deal, put the diagnosis in perspective: Firmus was an important reality check for the AI investment boom, but not the beginning of the end of the AI trade. The problems, she said, were specific to this deal: the speed of the valuation increase, the enormous capital requirements, and the execution risks attached to the expansion plans. That verdict matters, because it separates one failed listing from the health of the sector.

Firmus asked public investors to pay triple the August price for a company with two data centers running. The market’s answer was the whole story.

The IPO window narrows, selectively

Firmus did not fall alone, and that context is what makes the story bigger than one Australian listing. 2026 has been littered with offerings that never made it out the door. Smart-ring maker Oura postponed its US listing in September, shelving a deal that could have raised up to $2.2 billion at a $15 billion valuation. Nuclear technology company Holtec withdrew its offering the same month. Wall Street brokerage Clear Street scrapped its IPO in February, and homeowners insurer Bamboo postponed in late September. Investors are not on strike. They are demanding valuation discipline, and they are getting it.

At the same time, AI infrastructure paper is still moving. DayOne Data Centers filed for a US listing on October 5. Base Electron, the power infrastructure company formed by Applied Digital to supply its AI campuses, confidentially filed for an American IPO on October 9, the very day Firmus pulled its own. The window is not shut. It has a new rule: arrive with revenue, defensible contracts, or a price that leaves something on the table for the buyer.

What the pricing tells builders

For the founders and CFOs watching from the sidelines, the Firmus episode is a pricing memo. Private rounds can move at the speed of optimism, especially when strategic investors like Nvidia are in the syndicate and the narrative is about compute scarcity. Public markets, however, price the construction risk. A valuation that triples between a private round and a listing asks investors to pay for the upside of a buildout before a single new megawatt is delivered. (Public markets are not the only place AI money is moving inside the 2026 AI funding frenzy.) This week, they declined.

There is also a quieter signal in Firmus’s choice of next step. Rather than cut the price further and list anyway, the board said the offer terms did not correctly reflect the strength of the business, and concluded that proceeding was not in the best interests of the company and its shareholders. Firmus will now pursue capital from private markets and weigh alternative public and private options, a path that reportedly could include a future Nasdaq listing. Private capital is still willing to fund AI infrastructure. It just wants to do it at a price closer to August than to October’s first asking.

The broader AI buildout continues at full speed: the demand for compute is real, the chips are still scarce, and the power problem is still unsolved. None of that changed on October 9. What changed is the terms on which new money enters. Firmus discovered the difference between a valuation that investors admire and a valuation they will actually pay. Every data center developer filing after it will have read the memo.