The pitch was straightforward: the fastest-growing builder of AI data centers in the Asia-Pacific, backed by the world’s most valuable chip company, raising capital in the middle of a computing boom. Australian institutions looked at it and declined. Firmus Technologies cut its offer price from A$11 a share to A$8.25 and still could not fill the book. Earlier this month it withdrew the listing entirely.
On Wednesday, people familiar with the discussions said Nvidia is in talks to invest hundreds of millions of dollars in the company, now referred to in filings as Firmus Grid. Nvidia declined to comment through a spokesperson.
The sequence says something about how the AI infrastructure trade is being financed. Firmus had planned what would have been the largest listing on the Australian Securities Exchange since Telstra in 1997, targeting a valuation around A$44 billion. Its portfolio consists of seven planned AI factories across Australia, Singapore, Indonesia and Malaysia; two are operating. The rest are designs, permits and land.
In August, before the float, the company raised $2 billion from investors including Nvidia, Coatue and Blackstone at a valuation of $10.5 billion. Nvidia holds about 7.2% of the equity. Nvidia’s willingness to add to that position after public investors walked away is being read by the sector as a statement about intent rather than about the size of the cheque.
The withdrawal itself was handled quietly. Firmus told the exchange that recent market volatility meant the offer would not properly reflect its long-term prospects, and said it would pursue private funding while it considered alternatives. That is the standard language for a book that failed to fill, and it leaves the company in the position of negotiating with exactly the investors it had hoped to bypass.
What the cheque buys is time. Neoclouds, the specialist operators that rent out GPU capacity, fund their buildings with debt secured against contracts with AI developers. The model works while demand exceeds supply and while lenders believe the contracts will be paid. It becomes fragile when either assumption weakens, because the assets are specialized and the depreciation schedule is short.
The market began testing that assumption this month. CoreWeave, the largest of the group, and several smaller listed peers fell about 20% last week. The pressure spread to Oracle, whose cloud business has been signing capacity deals at a pace its balance sheet has to support. Analysts have estimated that Nvidia carries more than $170 billion of exposure to Firmus and other large customers, through equity stakes, backstops and prepayments.
Nvidia’s chief executive, Jensen Huang, said last week that CoreWeave no longer needs additional credit support from its supplier. Michael Intrator, CoreWeave’s chief executive, took a different line internally, blaming smaller operators such as Firmus for damaging the sector’s reputation with investors, according to a person who saw the message.
The debate underneath is about circularity. When the supplier of the chips also funds the companies that buy them, demand measured in orders is partly demand measured in the supplier’s own capital. Ken Griffin, the founder of Citadel, said this week that he believes AI is almost certainly a bubble, joining a list of investors who have made the same point with less specificity. Nvidia’s answer is that capacity is genuinely scarce: SpaceX’s Elon Musk said on an earnings call this week that older chips now rent for more per hour than new ones.
Firmus has its own complications. Its co-founders and co-chief executives, Oliver Curtis and Tim Rosenfield, built the company around access to cheap power in Tasmania, where the first site operates at St Leonards near Launceston. Investors who examined the float also weighed Curtis’s record: he is a former investment banker who served a year in prison about a decade ago for insider trading. Several fund managers cited governance and disclosure among their reasons for staying out, alongside the price.
None of that makes the assets worthless. A data center with power, cooling and a signed tenant has value to a lender even if the equity market will not price it. It does mean the equity is being repriced privately, and that the buyer of last resort is the company that sold the equipment.
For Nvidia, the calculation is defensive. A collapse at a mid-sized operator would raise the cost of capital for every other neocloud, including the ones whose orders underpin its own revenue forecasts. A few hundred million dollars is a small price for stability in a customer base that represents a large share of demand.
Lenders will read the talks that way. A neocloud’s debt is priced off the credit of its tenant and the credibility of its sponsor, and an equity cheque from the chip supplier functions as both. If Nvidia completes the investment, Firmus can take its construction plans back to the banks with a shareholder list that includes the company whose chips the buildings are designed around.
For everyone else, the lesson of the past three weeks is that the IPO window for AI infrastructure has narrowed sharply. Companies that planned to fund construction with public equity will now fund it with private capital, at prices the public market declined to pay. The question is how many of them have a supplier willing to write the cheque.
