Last Friday, Firmus Technologies withdrew its application to list on the Australian Securities Exchange. It was due to be the largest Australian float in roughly three decades: about A$7.1 billion raised at A$11 a share, implying an equity value of around A$43.7 billion. Trading was scheduled to begin on 23 October. Instead, the board concluded that proceeding was "not in the best interests of the company" and announced it would pursue capital from private markets, with a Nasdaq listing flagged for next year.
The mechanics of the failure were simple. The institutional book closed with weak demand. Advisers reportedly considered cutting the price to A$8.25 and shrinking the deal, which would have valued the company somewhere near A$30 billion. The board would not accept those terms. Australia's large superannuation funds and fund managers, the natural anchor buyers for a deal of this size, did not come to the table at the price asked. UniSuper's chief investment officer put it plainly: it "just doesn't have a compelling valuation".
This is not a story about a weak company or a weak idea. Firmus has two of the most important names in artificial intelligence as customers, the world's most valuable chipmaker on its share register, and one of the world's largest alternative asset managers as both shareholder and lender. It is a story about what happens when the capital requirements of the AI build-out meet a public market that is asked to fund them at full price, with high leverage, before the assets exist.
The Float That Didn't
The valuation path explains much of the reluctance. In September 2025, Firmus was valued at around A$1.85 billion. In August this year, a fully committed US$2 billion strategic round valued it at more than US$10.5 billion post-money, roughly A$15 billion. Eight weeks later, the IPO asked public investors to pay almost three times that. Against this, Firmus reported revenue of about A$50 million last year. Contracted revenue is real, but it only begins to flow as each facility is energised.
The physical position tells the same story. Firmus describes seven "AI factories" across Australia, Singapore, Malaysia and Indonesia, of which two are operating. Of roughly 912 megawatts of scheduled capacity, about 46 megawatts is running today — around 5%. Its stated pipeline, including growth rights, is 3.2 gigawatts. For perspective, the entire Australian data centre market is about 1.5 gigawatts.
Who Backed Firmus, and Why It Mattered
The share register reads like a map of the AI capital stack. Nvidia holds about 7.2%, having invested in 2025 and again in April and August this year. Coatue holds about 8.4%. Funds managed by Blackstone hold about 6.7%. Jane Street joined the August round. In February, Firmus secured a US$10 billion facility led by Blackstone Credit and Coatue to fund the Project Southgate rollout. Analysts have put Southgate's total capital requirement at about A$73 billion. In other words, the equity raised, the debt facility and the IPO proceeds combined would have funded only a part of the plan.
The customer list is equally strong. OpenAI is the anchor customer for two facilities in Malaysia. Meta has committed to Nvidia GB300 systems at the Melbourne site and to capacity in Southeast Asia. Look closely at Nvidia's role: it is Firmus's supplier, a shareholder, a provider of credit support and a revenue-share partner. This is not unique to Firmus. It is now a common feature of the neocloud sector, and it is the clearest signal of what the AI industry is really asking for.
What a Neocloud Actually Does
A neocloud is an independent provider of GPU computing as a service. Unlike Amazon, Microsoft or Google, which run broad cloud platforms, a neocloud does one thing: it buys large volumes of accelerated computing hardware, installs it in powered and cooled data halls, and rents that capacity to the companies that train and run AI models. The neocloud owns the factory; the AI company owns the product that comes out of it.
The economics are thinner than the headlines suggest. McKinsey estimates bare-metal gross margins at 55% to 65% before depreciation, falling to an operating margin of around 8% once depreciation is included. A GPU hour can lose half its value or more over a five-year depreciation horizon as new chip generations arrive. The sector's bellwether, CoreWeave, shows what this looks like at scale. At 30 June 2026 it carried about US$35 billion of debt. Second-quarter adjusted EBITDA doubled to US$1.5 billion, but net interest expense rose 2.4 times to US$640 million — roughly equal to its net loss of US$626 million. The business is growing very quickly. So is its interest bill.
One of More Than a Hundred: Why Firmus?
McKinsey counts more than 100 neoclouds worldwide. So why did Nvidia, OpenAI, Meta, Blackstone and Coatue all line up behind one Australian provider? Three reasons. First, Nvidia needs independent buyers — its largest customers are also designing their own chips, so every well-capitalised neocloud that commits to Nvidia's roadmap secures its allocation and extends its platform. Second, the AI developers do not want to own the assets. OpenAI and Meta both secured Firmus capacity through contracts, not ownership. Meta's Hyperion campus in Louisiana was financed through a joint venture in which Blue Owl owns 80% and Meta 20%, with about US$27 billion of debt raised outside Meta's balance sheet. Third, geography — ABI Research counts only about 35 neocloud facilities across Asia-Pacific, and Firmus offered land, power and permitting progress in a region where the large AI companies have little capacity of their own.
The AI industry is not backing Firmus because it is the best of a hundred similar firms. It is backing Firmus because it is a balance sheet in the right place, willing to own the assets that produce the compute. Big Tech has guided to somewhere between US$720 billion and US$745 billion of capital expenditure this year. Even those balance sheets are not large enough to carry the whole build-out alone.
The Second Balance Sheet
Neoclouds have become the industry's chosen way to close the capital gap. They are, in effect, secondary balance sheets: vehicles created to hold assets that the primary players want access to but do not want to own. The question is not whether the AI industry should use partner balance sheets — it is what kind of balance sheet should hold which asset, and at what leverage.
Here, the neocloud model has a structural weakness. A single AI factory contains assets with very different lives. GPUs and servers have an economic life of perhaps four to six years. The data hall and cooling may last fifteen to twenty years. Power generation, storage and water infrastructure can last twenty-five to forty years. In most neocloud structures, all of these assets sit together in one vehicle, funded by one stack of equity and debt. When long-life infrastructure and short-life compute share a balance sheet, the short-life risk sets the cost of capital for everything.
Where Leverage Turns Into Fragility
A highly geared neocloud depends on four things holding at once: anchor customers continue to pay, utilisation stays above about 80%, rental prices do not fall faster than the depreciation schedule, and debt markets stay open for the next raise. Customer concentration means one renegotiation can alter the economics of an entire facility. A new chip generation can reprice the rental market within a year. And when interest costs already equal the net loss, as at CoreWeave, there is very little room for any of these to move the wrong way.
Firmus illustrates the point before it had even begun trading. Its plan required a US$10 billion debt facility, an A$7 billion equity raise, and further capital still to complete a programme analysts put at more than A$70 billion. As UniSuper's John Pearce observed, the concern was that Firmus would "keep coming back to the market". That is the defining feature of a highly levered secondary balance sheet. The wider point is that the AI industry, by design, is pushing capital risk outward. If that chain is built on high leverage against short-life assets, the failure of a single large neocloud would not stay contained.
So, What's Around the Corner?
Expect Firmus to complete a private round, most likely with existing backers, and to return to public markets in New York. Expect more neoclouds to follow the same path, and more chipmaker and AI-developer support to flow to whichever balance sheets are willing to own the assets. Expect leverage across the sector to keep rising, and with it the cost of capital for any vehicle that holds both compute and infrastructure.
And expect the market to begin pricing the layers separately. The investors who stepped away from Firmus were not stepping away from AI. They were asking a reasonable question: whose balance sheet should carry which risk? The answer, we believe, is that the power and water beneath the AI build-out belong on low-leverage, securitised, utility balance sheets — funded once, for the long term, by a partner that is not dependent on the next equity round. Let the neocloud own the chips; let a utility balance sheet own the power and water.
This document is prepared by Monard Infrastructure for discussion purposes only. It does not constitute financial, investment, legal or tax advice, or an offer or solicitation to buy or sell any security. Figures are drawn from public sources believed to be reliable but have not been independently verified. Views expressed are those of Monard Infrastructure at the date of publication and may change without notice.
Sources: ABC News (October 2026); Forbes Australia (October 2026); Bloomberg via Yahoo Finance (October 2026); Capital Brief; DatacenterDynamics; McKinsey & Company, "Neoclouds' challenges and next moves"; ABI Research, Neocloud Data Centers Market Data Overview 2Q 2026; CoreWeave Q2 2026 results (Long Yield); Bisnow and Fortune on Meta's Hyperion joint venture with Blue Owl; MLQ.ai on Big Tech capex.