Firmus did not fail an equity test this week. It failed a credit test that happened to be wearing an equity ticker. Morningstar estimates the Australian data center developer's debt at about $30 billion. At the A$8.25 price its bankers are now defending, the equity would be worth roughly $23 billion on an unchanged share count. The IPO is therefore no longer selling a growth story. It is selling a subordinated cushion under a debt stack, and the bookbuild has priced that cushion like one.
That framing explains what the "AI window slams shut" headline misses. Demand for AI infrastructure exposure has not disappeared. What has disappeared is the willingness of new money to take the bottom slice of a capital structure that early backers built from the top down.
From A$11 to A$8: The Cushion Disappears
The arithmetic is short. Firmus marketed up to $5.5 billion of stock at A$11 a share, a valuation of A$43.7 billion, or about $30.4 billion. At that level, equity and debt were roughly one-for-one. Each dollar lent to the company had about a dollar of shareholder capital beneath it.
The book closed without a clear price. Per the Australian Financial Review, as relayed by Reuters Breakingviews, the price was cut to as low as A$8 and the banks were working to hold A$8.25. Scaling the original valuation gives about A$32.8 billion at A$8.25, or roughly $22.8 billion, and about $22.1 billion at A$8. Against an estimated $30 billion of debt, the shareholder layer now sits around $7 billion short of the obligations above it, before a single share trades.
One caveat matters here. The Morningstar figure is an estimate, and Firmus' disclosed financing is a $10 billion debt package led by Blackstone and Coatue in February. The gap between $10 billion signed and $30 billion estimated most likely reflects debt the build-out plan still needs to draw or raise. That reading would make the problem worse, because the debt the equity has to support is largely debt that does not exist yet.
The debt-to-equity ratio is not the most telling number in the deal. Even after the cut, the IPO price is more than double the $10.5 billion valuation of the August round that brought in Nvidia, Coatue, Blackstone and Jane Street. Public investors were still being asked to pay a large markup on a six-week-old private mark. The haircut reduced the markup without removing it.
EV+1/EBIT+2 as a Disguised Credit Model
The deal's signature metric, "EV+1/EBIT+2," deserves more attention than it has received. As Breakingviews' Antony Currie describes it, it divides enterprise value using net debt 12 months out by EBIT expected two years out. The rationale is that developers fund up to 90% of construction with debt and start earning within a year, so trailing figures understate the business.
Read as a credit analyst would, the metric is a construction-loan model presented as an equity multiple. It front-loads the liabilities and back-loads the cash flow, and it carries a hidden assumption that the schedule holds. Currie names the failure mode directly. Delays in permits, equipment, power or water turn EV+1 into +2 or +3, and EBIT slides to +3 or +4. Each year of slippage adds interest without adding EBIT.
The cost of that slippage depends on rates, and here the macro channel carries real weight. The 10-year Treasury stood at 5.27% on October 6, with the 2s10s spread at +0.48%. A Goldman Sydney desk note put the 10-year at 5.36% on Wednesday, a 24-year high. A positively sloped curve with long rates at multi-decade highs means term financing for multi-year builds is getting more expensive at the long end, which is exactly where data center debt sits. A metric that assumes EBIT arrives on time will look worse in this rate environment than it would have a year ago.
The schedule risk is not hypothetical. The current record is poor. Project Jupiter has declared force majeure, with its loans reported below 90. Oracle's 1.3GW Wisconsin "Lighthouse" campus slipped this week. Firmus has built 42 megawatts against a pipeline of about 1 gigawatt, just over 4% of its stated capacity. Its draft prospectus reportedly forecast a $77 million loss in the first half of FY27 and gave no forecasts beyond June 2027. The "+2" in the denominator falls in a period the company itself declined to project.
Who Is Selling, and to Whom
The allocation structure points the same way. About half the IPO was earmarked for "selected existing strategic and financial investors," a list that includes Nvidia, Coatue, Blackstone and Jane Street. Pre-IPO holders could reportedly sell up to 20% of their stakes from day one. Bloomberg identified that overhang, together with aggressive pricing, as the reason prospective buyers turned cautious.
Look at what those backers hold. Blackstone and Coatue led the debt and the equity rounds. Nvidia is an equity holder, an eight-year partner on the Batam, Indonesia project with DayOne, and the supplier of the GPUs the proceeds would buy. When the same institutions sit at the senior and junior ends of the capital structure and on the supply contract, a public equity raise works partly as a deleveraging event for them. New outside equity strengthens the collateral behind their loans, cuts their concentration, and funds chip orders, all at once. Whether the debt documents make the equity raise a condition of further drawdowns is not public, so that specific link stays unverified. The economic direction is clear even without it.
Outside buyers appear to have reached the same conclusion. Plato, an Australian manager running about $6 billion, called the stock a "screaming short" and cited 30 red flags. Shares of backer Maas Group fell as much as 30% in Sydney, a record drop. Ten Cap's Jun Bei Liu described international interest that faded "when they were asked to put up the capital that's required." Three days earlier, the bookbuild had been pulled forward because of "early indications of demand in excess of the offer size." Indications cost nothing. Allocations require cash, and the gap between the two is where this deal broke.
The Countercase: AI Capital Is Still Flowing
The strongest objection is that none of this signals scarce capital. Goldman Research estimates global AI investment will exceed $1 trillion this year, with the US and Asia taking more than 80% of net data center additions. Hedge fund net exposure to the Mag 7 on Goldman's prime book is about 22% of total US exposure, the highest since early 2022. London's Nscale just raised $3.36 billion. DayOne, Firmus' own partner, is still aiming for up to $5 billion at a $20 billion valuation in a US listing as soon as November. Switch has filed confidentially, and Vantage and CyrusOne are exploring listings.
That evidence holds up. It also supports the credit read once you look at the form the money takes. Nscale's $3.36 billion came as pre-IPO convertibles, an instrument with debt-like downside protection and equity upside. The hedge fund flow is going to companies whose balance sheets can self-fund capex. Capital is not leaving AI infrastructure. It is moving up the capital structure and toward issuers with existing cash flow. Investors want convertibility, seniority, or a hyperscaler's balance sheet, and plain first-loss common stock in a 42-megawatt former Bitcoin miner offers none of those.
Shenton Research's Ke Yan drew the dividing line: whether demand is "contracted and already energised, or only planned." That distinction separates two kinds of credit. Energised, contracted capacity behaves like infrastructure debt. Planned capacity behaves like project finance in construction, and in that phase lenders depend on equity holders absorbing delays. Firmus sits clearly on the planned side, which is why the equity cushion was the variable that mattered.
The countercase would win if the deal priced at A$8.25, traded steadily, and DayOne then listed near its target. That would show that a lower entry point is enough to clear developer equity and that the issue was price, not structure.
The Tells in October and November
The credit markets have already registered the stress that equity buyers are now pricing. Oracle bond yields and CDS sit at record levels. The bond tied to Meta's Hyperion data center, Project Beignet, has hit a record low. Project Jupiter loans trade below par. SB Energy postponed its roughly $60 billion IPO in late September amid SEC questions and concern about its reliance on OpenAI as a single customer. Nvidia had agreed to guarantee up to $105 billion to help OpenAI lease SB Energy's Ohio site. Two mega-IPOs have stalled in three weeks, and both carried heavy debt, thin contracted revenue, and early backers looking for liquidity.
The dates are fixed. Firmus is due to lodge its prospectus on October 12 and list on October 23. Three observations will settle the read:
- Disclosed debt. If the prospectus confirms that debt drawn or committed in the build plan approaches the $30 billion estimate, the equity-below-debt framing becomes the stated structure rather than an analyst's napkin math.
- Day-one float. If the listing holds A$8.25 with insiders absorbing half the book and the 20% sell-down window open, it shows real outside demand. If it trades below, the cushion shrinks further in public.
- DayOne in November. A clean $20 billion US listing for Firmus' Batam partner would weaken the read. A delay, a cut, or a convertible structure would confirm that developer equity has moved to the back of the line.
The thesis breaks if construction-phase data center equity clears at issue price without insiders supporting the book. Until that happens, the cleaner reading is that the AI build-out still has buyers for its senior paper and its hyperscaler balance sheets, and has lost its buyers for the risk underneath. When an IPO is priced so the equity cannot cover the debt, a stall in construction leaves the shareholders without value and leaves the loss with the lenders.
