Why Patient Capital Walked Past the Firmus Roadshow

1. Patient capital rejected the Firmus IPO because the valuation relied on aspirations rather than earnings. Despite strong backing from major investors and extensive marketing support, long-term investors could not justify paying for a business with no earnings, no clear profitability forecast and significant future borrowing requirements.

2. The key risk was the gap between project-level targets and company-level economics. Firmus promoted ambitious revenue and EBIT targets, but investors were left without a consolidated forecast that accounted for corporate costs, debt financing, interest expenses and GPU amortisation.

3. The failed float highlights the importance of independent analysis over sales-driven narratives. Unconflicted investors focused on valuation, financing and risk, while the IPO structure rewarded parties paid to place the stock rather than those responsible for long-term investment outcomes.

On 9 October, after 48 hours of bookbuilding, Firmus Technologies withdrew its A$7.1bn IPO. Despite cutting the price from A$11 to about A$9 a share, the book still couldn't be filled without existing shareholders taking more than half. Valued at A$43.7bn just a fortnight ago, the company is now raising funds privately, with a Nasdaq listing pencilled in for next year.

The IPO had everything a roadshow needs: mega-backers such as Nvidia, Blackstone and Coatue on the pre-IPO register, alongside hyperscaler tenants Meta and OpenAI. The ‘sales machine’ had joint lead managers, a co-manager list that read like a directory of Australian broking and wealth management firms, and lead-manager research valuing the business at two to three times the float price, making the IPO look like a discount.

It had no earnings and no forecast for when it would.

Firmus currently has 46 megawatts (MW) running and 42MW in someone else's data hall. A total of 911MW is contracted, with the remaining 865MW to be built in Indonesia, Malaysia and Tasmania within 24 months.

Revenue for the year ending June 2026 was US$50.8m. The draft prospectus forecast a US$77m half-year loss, with no further forecasts. At A$11 per share, that was roughly 600 times trailing sales.

The ‘sales machine’ roadshow answer consisted of project targets: A$20.2bn in revenue, a 91% EBITDA margin and US$5.8bn in EBIT once everything is built. These are site-level figures for different periods, calculated before corporate costs, and Firmus management never consolidated them into a forecast it could be held to. They are an aspiration for 2029, presented as though they were a budget.

Between a project margin and a company margin sit the costs the roadshow left out. Reaching 911MW would require about US$36bn in servers, compared with US$3.4bn of equity raised. The difference, roughly US$30bn, is debt that an unrated company would have to borrow, secured against contracts whose terms prospective investors never saw. At 7% per annum, interest would be US$2.1bn a year, a third of the EBIT target. At 11% per annum, where CoreWeave's Blackstone-led GPU facility is priced, it would exceed half the target. The chips are written off over five years, and nobody yet knows what a five-year-old GPU rents for. CoreWeave, the largest listed operator of this model, posted an adjusted operating margin of about 5% last quarter, compared with the 41% implied by Firmus's EBIT target.

Firmus’s share price rose faster than its build plans. In August, Nvidia, Blackstone, Coatue and Jane Street invested at a valuation of US$10.5bn. By October, it raised US$30.6bn from the public. Meanwhile, it signed a contract for one plant, but that plant was never built, and the debt was never raised. Although 42% of the register was escrowed, the founders could sell a tenth of their holding in the first year.

It’s important to note that none of this is peculiar to Firmus. The credit market has already marked down the entire neocloud debt tier. CoreWeave's 2030 bonds trade about 10% per annum over Treasuries, yielding close to 16%, while Meta, its tenant, borrows at just 1% per annum over Treasuries. Gartner expects four in five neoclouds, the independent AI-compute landlords that include Firmus, to have been bought, merged or broken up by 2030. For a company in this tier, the base case isn't growth into its valuation; it's being absorbed, usually by its lender or biggest customer, on their terms.

Two other factors weigh on the valuation. Token volumes are growing quickly, but token spending is not. The gap will determine what the chips rent for when the first contracts expire. The last comparable cycle is not encouraging. The US telecom sector, the TMT boom's biggest capex spender, peaked in November 1999 and fell 47% before its capex, cash flow, or returns declined. Infrastructure is derated during the build phase, well before operating numbers deteriorate, so construction-site progress reports offer no protection.

An IPO of this magnitude is sold by three parties with aligned interests. First, the issuer, which wants the highest price it can get. Second, the investment banks, who are paid a percentage of the amount raised. Finally, the broking and wealth firms, who earn a selling fee on every allocation placed in a client account. That fee is paid on placement, not on outcome. Add a theme as seductive as AI, and fee-driven selling meets FOMO.

The Firmus IPO came dangerously close to working. At about A$9 a share, with insiders taking half the book, the public would have owned a A$36bn company with 46 megawatts up and running, a six-month forecast, and US$30bn still to borrow. This was a week before CoreWeave in the United States fell 25%. Co-manager clients were spared that outcome, but the channel cannot take credit. They are paid to place the stock, and a selling fee rarely makes them think twice about the price. UniSuper and other long-term managers spared them because the valuation required too much to go right, and they had nothing at stake or riding on the outcome except their investors' long-term wealth security. Conflicted remuneration is alive and well in Australian wealth management. In this case, investors who were not being paid agreed to beat it off.

The ASX is meant to be the gatekeeper. Yet it allowed a company to come to market with a six-month forecast, unconsolidated targets and 42% escrow, despite a 1.6-gigawatt partnership having been quietly abandoned. It remained undisclosed until the other party mentioned it mid-bookbuild. The company was also on course for index inclusion, which would have forced every index-tracking super member to buy its shares. Of course, an exchange competing with Nasdaq for listings cannot be expected to reject applicants voluntarily. That is the purpose of listing rules.

The Firmus float showed the market working, though only just. A well-dressed roadshow ‘dance’ reached a handful of unconflicted investors who wanted to see the financing before paying for the forecast - the price couldn't withstand that scrutiny. Next time the sellers may be luckier, yet the conflicts that nearly carried this one over the line will remain. Private investors don't have UniSuper's analysts, but they do the same arithmetic, and it takes about an hour.