Firmus Pulled Its IPO at 3x Its August Price. The Neocloud Window Narrowed

Firmus asked public investors to value it at US$30.6 billion. Two months earlier, private investors had paid in at US$10.5 billion. The book did not hold, and the Firmus IPO was withdrawn on 9 October 2026, two weeks before its planned ASX debut. The lesson is not that AI infrastructure is unfundable. It is that public buyers price risk line by line, and private rounds mostly do not.

The short answer. The Firmus IPO failed because public investors would not pay US$30.6 billion for a company with roughly 46 MW running, about US$30 billion of expected debt and a customer list led by two names. The same company had raised at US$10.5 billion in August. Price, leverage and concentration all hit the book in the same week.

What happened to the Firmus IPO

Firmus is a Singapore-incorporated AI data centre developer with Australian roots and Nvidia as a shareholder. It started as a cooling supplier to bitcoin miners and pivoted to building what it calls AI factories. The float would have been one of the largest in ASX history.

The timetable, day by day

The institutional book opened on Tuesday 6 October at A$11 a share, with the aim of closing by Friday. According to Reuters, indicative orders had initially covered the deal. Investors began pulling those orders on Wednesday 7 October as confidence faded.

Bookbuilding shut on Thursday 8 October. On Friday 9 October, the board withdrew the listing application and said it would pursue capital from the private markets instead. Trading had been pencilled in for 23 October.

The board's statement was blunt. It concluded that proceeding was not in the best interests of the company and its shareholders, because the available terms would not reflect the strength of the business. In plain terms, the price that would have cleared was lower than the board would accept.

The size of the ask

Reuters described the offer as a US$5 billion IPO. Pre-withdrawal coverage put the full size at up to US$5.5 billion including the over-allotment option, or about A$7.9 billion. Coverage also reported that advisers looked at cutting the price toward A$8.25, but Reuters says the company chose a private round over a cut.

The spillover was immediate. Maas Group, an ASX-listed construction business that holds about 3.2% of Firmus and builds part of its infrastructure, lost more than 20% of its market value on 8 October. That was about A$517 million gone in a day, according to Investor Strategy News.

Why the Firmus valuation tripled in private and stalled in public

The private marks rose fast. The Next Web put Firmus at about US$1.9 billion after a late 2025 raise. SiliconANGLE reported US$5.5 billion in April 2026. Dealroom reported more than US$10.5 billion after the August round, led by returning backers Coatue and Nvidia with Blackstone and Jane Street joining.

Then the IPO price asked public investors for nearly three times the August number. Nothing on the operating side had tripled in nine weeks. Contracted capacity had passed 900 MW, which Firmus announced on 8 September, but the running capacity was still a few dozen megawatts.

This is the core of the story, and I think it gets misread. The private rounds were not proof that US$30 billion was fair. They were negotiated by a small group of strategic holders, several of whom also lend to, supply, or buy from the company. A public book has no such relationships to protect.

Jun Bei Liu of Ten Cap told Forbes Australia that confidence "evaporated" and that the company needed to be more mature. Josh Gilbert of eToro said investors were not prepared to pay a high price up front for capacity largely still on the drawing board. Both are describing the same gap: a private markup that had not been earned in megawatts.

The debt public investors had to price into the Firmus IPO

A data centre developer is a capital project before it is a business. The build is paid for mostly with borrowed money, and the shares sit behind that debt. So the right number for public buyers is enterprise value, not the equity headline.

US$10 billion signed, about US$30 billion estimated

In February 2026, Firmus closed a US$10 billion debt facility led by Blackstone funds, with support from Coatue. The Next Web described it as long-dated infrastructure debt for the national rollout of Project Southgate. Interest rate, maturity and drawdown schedule were not public.

Analysts working for the joint lead managers, Bank of America, JPMorgan, Morgan Stanley and Morgans, put total debt at about US$30 billion. That takes enterprise value to roughly US$60 billion. The equity was half the capital structure, and the half that gets paid last.

What six times earnings looks like with no earnings yet

Reuters reported that the draft prospectus pointed to about US$5 billion of annual earnings from the data centres within five years. Debt of US$30 billion against that figure is about six times a profit line that does not exist yet. On a running business, six times is heavy but serviceable. On a plan, it is a bet that every site energises on time.

The near term pointed the other way. A Reuters report on the draft prospectus, relayed by RuntimeWire, said Firmus expected a US$77 million after-tax loss for the first half of its 2027 financial year. Forbes Australia put revenue for the last financial year at about $50 million. Even the most generous reading of that figure puts the ask at several hundred times trailing revenue.

The listed comparison most readers reach for is CoreWeave. Its debt structure and the covenants that bind it are covered in the breakdown of where CoreWeave's debt stack is under stress. The relevant contrast here is simpler: CoreWeave listed with revenue in the billions. Firmus asked for a comparable equity value with revenue in the tens of millions.

Leased capacity and the CDC split behind the Firmus listing

Reuters reported that Firmus had two online data centres, in Melbourne and Singapore, and that both were leased. Five more sites across Asia-Pacific were planned. The difference between a leased hall and an owned campus matters more than it sounds.

Two leased sites online

A leased site means someone else owns the building, the grid connection and often the cooling plant. The tenant installs GPUs and pays rent. That makes an early start possible, but it also means the most defensible part of a data centre business, a powered and permitted site, belongs to the landlord.

The obligations do not disappear because they are rent rather than debt. A signed lease is a long-dated fixed payment that ranks ahead of shareholders in practice. The way those commitments can sit outside a reported debt figure is set out in the analysis of data centre leases that never reach the balance sheet.

Southgate ended at about 42 MW

Project Southgate was announced in October 2025 by Firmus, CDC Data Centres and Nvidia. It was billed at up to A$73.3 billion of investment and 1.6 GW of capacity across Australia by 2028. On 6 October 2026, the day the book opened, W.Media reported that the partnership had ended with about 42 MW deployed.

Firmus co-chief executive Oliver Curtis said the two companies had mutually agreed earlier in the year not to proceed. He said it did not affect contracted customer capacity. CDC chief executive Greg Boorer said Firmus had made other choices about doing its own data centre developments. Per the AFR, as reported by W.Media, Firmus stays in a CDC Melbourne facility but will not expand to the 150 MW once planned there.

The timing did the damage, more than the substance. Public investors learned on day one of the bookbuild that the flagship partnership had been dead for months. Whatever the commercial logic, a disclosure landing in the press during the offer reads as a disclosure the company did not lead with.

Partner concentration: who sits on how many sides

Firmus's customer story is genuinely strong on paper. OpenAI is the anchor customer for dedicated capacity at two Malaysian sites. Meta takes capacity in Melbourne and has agreed to more in Southeast Asia. Firmus has said it expects US$25 billion to US$30 billion of offtake over its first six years.

The problem is not the quality of the names. It is how few of them there are, and how many roles each plays. The pattern of suppliers financing their own customers is mapped in the breakdown of how the AI circular deals actually work, and Firmus is a compact example of it.

Nvidia held 7.2% of Firmus and is also the source of every accelerator in the plan. Firmus has said it secured access to 170,000 Nvidia chips for its Indonesian site. Blackstone held 6.7% and led the debt. Coatue held 8.4% and supported both the equity and the facility.

None of this is improper, and all of it was disclosed. But it changes what the August valuation means. When your shareholders also lend to you and sell you chips, the price they pay for equity is one term in a larger commercial relationship. A public fund manager is buying only the equity, and prices it alone. The same supplier-concentration question, seen from the other side of the trade, appears in the analysis of Nvidia's customer concentration.

Who the Firmus float was really for

An IPO is usually sold as fresh capital for growth. The structure of this one told a more complicated story, and public investors read it closely.

According to a Reuters report carried by iTnews on 6 October, Firmus planned to allocate about half the offer to selected existing strategic and financial investors. Reuters later reported that the escrow terms would have let existing holders sell more than half the stock from day one.

Put those two facts together and you get a book where insiders take a large share of new stock while most existing stock is free to trade. That is not unusual in a hot market. In a cold week it reads as an exit window, whatever the intent. Forbes Australia reported that major superannuation funds largely stayed out.

I would treat the escrow term as the single most fixable mistake in the deal. Debt and construction risk are inherent to the business model. A lock-up covering most insider stock for 12 months costs the company almost nothing and removes the most corrosive question a buyer can ask. How preferences and ratchets reshape a listing price is covered in the piece on liquidation preferences and IPO ratchets in AI.

Where this argument is weakest

The post so far reads the withdrawal as a verdict on Firmus. There are two serious reasons to resist that reading.

The book was covered before it was not

Reuters reported that indicative offers had initially been enough to cover the transaction. Demand did not fail to appear. It appeared and then left over two days, while the CDC news was circulating and global AI stocks were volatile. That is a timing and sentiment failure as much as a valuation failure.

If the CDC story had landed a month earlier, or the market had been calmer that week, the same price might have cleared. Wilson Asset Management's Oscar Oberg said simply that the market had spoken, and he also said he backed management to deliver. Both can be true.

The ASX is not Nasdaq

Investor Strategy News framed the withdrawal as the ASX refusing to pay for the AI trade on US terms. Curtis himself told The Australian that Australia generally takes a dim view of people taking a risk. Tectonic's Ben Haan said a Nasdaq listing was now the most likely alternative, according to Sharecafe.

If Firmus lists in New York within 12 months at a price near its ask, the lesson here becomes about venue, not valuation. I think that outcome is less likely than a lower price somewhere. I cannot rule it out, and nobody else can either. A single failed bookbuild is one data point, not a regime change.

There is also a source limit to state plainly. I could not open the prospectus itself. Every figure in this post comes from reporting on the offer documents, mostly by Reuters and Forbes Australia, and some figures differ across outlets. Operating capacity is reported as both 42 MW and 46 MW, for example.

Tests for the next neocloud IPO

CoreWeave is the precedent that already told this story once. In March 2025 it priced at US$40 a share, below an earlier range, and sold 37.5 million shares for US$1.5 billion, according to Seeking Alpha. It accepted a cut and listed. Firmus refused the cut and stayed private. Those are both legitimate choices, and they lead to different capital costs.

Several neoclouds are expected to test public markets over the next year. My view is that every one of them will face the questions Firmus did, in this order. The table is the checklist I would use as a buyer.

The order matters. A buyer who hears a good answer on the first two tests will tolerate a weak answer on the third. Firmus gave weak answers on all five in the same week. For the wider context on how fast the funding cycle can turn, see the analysis of private credit exposure to AI infrastructure.

Frequently asked questions

Where to start

If you hold or are weighing exposure to an unlisted AI infrastructure company, write down three numbers from its last update this week: megawatts running, megawatts contracted, and total debt. Divide the first by the second. If the company will not give you all three, that is the answer to how a public book would treat it.

If you sell into, build for, or partner with a neocloud, ask which of your counterparty's sites are owned and which are leased. Maas Group's shareholders learned on 8 October how quickly a partner's financing event becomes your share price.

Related reading