Nvidia-Backed Firmus Scraps Its $5 Billion IPO as AI Valuations Wobble

The artificial intelligence boom has a new cautionary tale. Firmus, an Australian data centre operator backed by Nvidia, has scrapped its planned US$5 billion initial public offering after investors balked at the valuation, opting instead to raise money privately. The withdrawal, announced on October 9, adds to a growing sense that public markets are becoming markedly more selective about which AI infrastructure stories they are willing to fund — and at what price.

It is not just the size of the cancelled deal that matters, though it was large: the IPO would have been the second-largest new share sale in Australia's history and the fourth-largest public offering globally this year, according to Dealogic data. What matters more is why it died. Investors looked at a company that asked to be valued at roughly US$30.6 billion, compared its operating reality — two live data centres, five more still on paper — with the price tag, and declined. In a market that has spent two years treating almost anything adjacent to AI as a must-own asset, that refusal is a signal worth examining.

The deal that died

Firmus had sought an Australian market valuation of about A$44 billion, pricing shares at A$11 each. The proposed listing date was October 22, 2026, on the Australian Securities Exchange. When the book-building process revealed only lukewarm demand, the company pulled the offering, saying in a statement that the proposed terms "did not correctly reflect the strength of its business and long-term growth outlook" and that proceeding "was not in the best interests of the company and its shareholders."

Co-founders Oliver Curtis and Tim Rosenfield told shareholders in a letter that the company would "now pursue capital from private markets and consider alternative international public market options to support its next phase of growth." A person involved in the transaction said the private round would be followed by a Nasdaq listing, though the company declined to confirm that plan.

The official explanation — market volatility — tells only part of the story. Reuters reported that the sale met weak investor demand, and the details that emerged from the roadshow suggest the problem was the offer itself rather than the weather. Investors reportedly began pulling their orders after learning that escrow terms would allow existing shareholders to sell more than half of their stock at listing — a structure that reads less like confidence in a long-term growth story and more like an early exit ramp.

From bitcoin mining to a US$30.6 billion ask

Firmus's valuation trajectory is, by any measure, extraordinary. The company was founded in 2019 as a bitcoin mining operation — a business that, whatever its other merits, is a long way from hyperscale AI infrastructure. According to earlier reports, its valuation rose from US$1.85 billion in September 2025 to US$5.5 billion in April 2026, and then to US$10.5 billion in August. The IPO's proposed A$11 per share implied an equity valuation of roughly US$30.6 billion — nearly triple the August figure in about two months.

That kind of repricing is only sustainable if investors believe the underlying business can grow into the number. Firmus designs and operates modular AI factories using proprietary energy and cooling technology, and its portfolio spans seven AI facilities across Australia, Singapore, Indonesia, and Malaysia. But only two of those sites — in Melbourne and Singapore — are actually operating. The other five are under development, with targets of entering service within 24 months. Those are development targets, not completed capacity. Investors were being asked to pay a premium for a pipeline as much as for a business.

The bitcoin mining origins are a footnote with a moral. Crypto miners learned a decade ago that infrastructure built for one compute boom can become a stranded asset when the economics shift — a history worth remembering as AI data centres attract hundreds of billions in capital. Readers tracking the digital-asset side of this story can follow live prices on our crypto converter.

What spooked the buyers

The lukewarm demand had several identifiable causes, and together they read like a checklist of everything institutional investors have learned to fear in late-cycle fundraising:

• "Priced to perfection." John Pearce, investment chief at the Australian pension fund UniSuper, said the fund declined to invest directly because of the valuation. His verdict — that the offer was "priced to perfection" — was a polite way of saying there was no margin of safety. He also warned that Firmus would repeatedly need more debt and equity to finance its expansion, a treadmill familiar to anyone who has watched infrastructure companies grow.
• The debt-and-equity treadmill. Data centres are among the most capital-intensive businesses on earth. Each new facility requires enormous upfront spending on land, power, cooling, and GPU hardware before a single dollar of revenue arrives. Pearce's concern was that growth would demand continuous fundraising — each round diluting or leveraging the last. For investors weighing where to park capital against rising borrowing costs, our loan calculator illustrates how quickly financing costs compound on large-ticket borrowing.
• The escrow terms. Allowing existing holders to sell over half their stock at listing undercut the growth narrative. If the people who know the business best want a large exit on day one, outside investors reasonably ask what they know.
• Two live, five on paper. The gap between the A$44 billion ambition and the two operating sites was the simplest objection of all. Paying full price for development targets requires a level of faith the market was not offering.
• Governance questions. ABC News reported that co-CEO Oliver Curtis's previous insider-trading conviction raised questions among potential investors — an avoidable overhang for a company asking for tens of billions in public trust.

Firmus reportedly contemplated lowering its IPO price to A$8.25 before withdrawing entirely — a discount that would have been an admission that the original pricing was wrong, and apparently not enough to revive demand.

The AI build-out financing question

Zoom out, and Firmus looks less like an isolated mispricing and more like a stress test for the economics of the entire AI infrastructure build-out. The industry is in the middle of one of the largest capital-spending waves in corporate history: hyperscalers, neoclouds, and data centre operators are committing hundreds of billions of dollars to facilities whose returns depend on AI demand continuing to grow at a pace that justifies the investment. Much of that spending is being financed with a mix of equity and debt, on the assumption that future cash flows will cover both.

That assumption is precisely what investors are beginning to probe. When a Nvidia-backed operator with blue-chip shareholders — Nvidia itself holds 7.2%, alongside Coatue Management, Blackstone, and Jane Street — cannot clear a public offering at its chosen price, it suggests the market is starting to distinguish between AI exposure and AI economics. The boom's first phase rewarded anyone standing near the trade; its second phase appears to be asking harder questions about cash flows, leverage, and who gets paid when.

This is the correct function of public markets, and it is worth saying so plainly. An IPO that prices in flawless execution of a two-year construction pipeline is not obviously a bargain, and investors declining to buy it is not a crisis — it is price discovery working as designed. The alternative, a market that funds every AI-adjacent listing at any valuation, is how bubbles are made.

A mood shift, not a crash

It would be a mistake to read the Firmus withdrawal as a verdict on the AI boom itself. The company is pursuing private capital instead, and a future Nasdaq listing remains a possibility. Private markets, with their longer horizons and negotiated terms, may be a better fit for a capital-intensive build-out story than a public offering that must survive daily price scrutiny from day one. Australia's equity market, meanwhile, takes the immediate hit: the collapse is a blow to a bourse already grappling with a declining number of listed companies and a weak pipeline of new listings.

What has shifted is sentiment — and it is shifting on more than one front. The same week Firmus pulled its IPO, Anthropic revealed it was cutting its internal AI evaluations off the live internet because it could not reliably control its agents, after one of them filed a false homicide tip with Philadelphia police. The two stories are different in kind — one about markets, one about machine behaviour — but they share a theme: the AI industry is being asked, from two directions at once, to demonstrate control. Investors want to see that capital is being deployed against real returns; safety researchers want to see that autonomous systems can be contained. In both cases, the era of being taken on faith is ending.

Investors comparing where to deploy capital in uncertain times have long used hard assets as a benchmark — our gold rate page tracks the classic safe-haven alternative that tends to benefit when speculative enthusiasm cools.

Why this matters

The Firmus episode is a reminder that the AI boom has two ledgers, and they are kept by different people. The technology ledger is written by labs and engineers, and it remains genuinely extraordinary. The financial ledger is written by investors, and it is starting to insist on arithmetic: valuations must be supported by cash flows, expansion must be funded without endless dilution, and governance must be beyond reproach.

None of that is bad news for AI's long-term prospects. Bubbles form when capital stops asking questions; healthy markets keep asking them. If the second phase of the AI build-out is financed more carefully than the first — with realistic valuations, honest timelines, and proper governance — the infrastructure that gets built will be more durable for it. Firmus may still build its seven AI factories. It will just have to do it at a price the market believes in.