Firmus Just Killed Its A$5.5B IPO. The AI Data-Center Class Just Lost Its Public-Market Free Pass.

Firmus had been running institutional bookbuilding on October 7 and 8, with shares marketed at A$11 each for an equity valuation near A$30.6 billion — nearly triple the A$10.5 billion post-money mark it had set just ten weeks earlier in an August 2026 round led by [Nvidia], Coatue, Blackstone and Jane Street. By Thursday evening Sydney time, the company filed to withdraw its ASX listing application; brokers and cornerstone investors learned the deal was off, with several major accounts publicly accusing the underwriting syndicate of providing 'misleading information' during the marketing process.

What killed the deal was the gap between the valuation step-up and the operating numbers behind it. Firmus had guided to roughly US$230M of revenue for the first half of FY27 alongside a US$77M after-tax loss — meaning the A$30.6B equity ask priced the business at well over 100x forward-half revenue. Co-founders Oliver Curtis and Tim Rosenfield, who together control roughly a quarter of the equity, also disclosed they had no escrow restrictions on their stakes — an unusual structure that gave public-market holders no protection against an early secondary wall.

The 900+ megawatts of contracted capacity Firmus was leaning on also broke down under scrutiny. Roughly two-thirds of that book sits behind facilities not yet under construction, including the two Malaysian data centres that anchor OpenAI's September 2026 multi-year deal. And the Australian rollout itself contracted in the days before the IPO: on October 6, the AFR reported that Firmus and CDC data centres had terminated the Project Southgate partnership, ending an expansion plan that had been valued at A$1.3B and scrapping the bulk of a 150MW Melbourne buildout. With debt-funded capacity contracting and forward revenue still mostly contractual paper, the A$30.6B price was the math investors refused to fund.

Eight days before Firmus pulled its deal, Accelevation Holdings — the Ohio-based manufacturer of power-distribution, cooling and modular kit for AI and hyperscale data centers — priced its Nasdaq IPO at $18.00 per share, below the marketed $20–$24 range and valuing the company at roughly $3.9B versus a target as high as $5.37B. The stock opened at $17.55 and traded below issue on day one. Accelevation's largest customer represented 35.2% of revenue in H1 2026, and the company generates real, recurring sales — yet it still had to take a price cut and a smaller valuation to clear.

Accelevation's discount was the dress rehearsal for Firmus's withdrawal. The supply chain gets a lower multiple, but it still gets capital — the neocloud/AI factory class, by contrast, is now being asked to underwrite 'capacity will monetize' against the CoreWeave tape.

Three large private AI-infrastructure names remain in the 2026–27 IPO window, and each carries the same fragility Firmus exposed. Nscale — the British neocloud that filed its S-1 on September 18 for a NYSE listing under the ticker NSCL — is targeting up to $3B raised at a $30–35B valuation. Its prospectus discloses $103.4B of total contract value but only $140.6M of H1 2026 revenue and a $1.02B H1 net loss. Nscale backed that filing with a $3.36B convertible round on September 25 (initial $2.36B at closing, plus $1B from [Nvidia]), with conversion pricing that ratchets up if the IPO clears above a $30B mark.

The AI-cloud proxy CoreWeave closed at $81.58 on October 8, 2026, down 7.8% on the session and within touching distance of its 52-week low of $60.55 — about 47% below the $153.20 high. The selloff comes despite Q2 2026 revenue of $2.58B (+112% year-on-year), a $104B contracted backlog as of June 30 and a fresh $14.2B Meta compute deal plus a multi-year Anthropic agreement announced in April. The thesis is no longer in dispute; the financing is. CoreWeave's net debt-to-EBITDA stands at 11.5x on TTM metrics through June 2026, capex ran at 2.7x revenue, and the company is still GAAP-unprofitable (-25.4% net margin TTM).

CoreWeave also raised $3.7B in convertibles in September at $97.85 per share — the company has to clear that conversion hurdle while building the data-center capacity to recognize even a fraction of its $104B backlog. Public-market investors just watched CoreWeave trade roughly to its conversion price within weeks; that is now the discount rate applied to every AI-factory IPO candidate, and it is materially below what Firmus was asking.

The withdrawal does not break the AI capex trade — it bifurcates it. Nvidia reported Q2 FY2027 revenue growth of 105.9% and earnings growth of 127.8%, with TTM revenue of $302.97B and TTM EBITDA of $201.27B against a net debt-to-EBITDA ratio of 0.07. The chip vendor gets paid for every data center Firmus, Nscale, [Crusoe] or Lambda build — at A$11 or A$8.25, listed or private, Australian or American. Upstream physical-infrastructure vendors Vertiv, Eaton and Schneider Electric face the same economics: data-center capex is forecast to approach $800B in 2026 across the top five hyperscalers, and roughly a third of that spend is debt-financed per Goldman Sachs research. The customers buying racks, cooling and power distribution keep building whether the IPO window is open or not.

The repricing hits the AI-factory layer. With [OpenAI] signed as the anchor for [Firmus's] Malaysia facilities and Meta signing a parallel Asia-Pacific capacity deal on September 29, hyperscaler demand is still signed. The risk is on the equity-issuance side: the equity that was supposed to fund the next wave of capacity now costs more, and the runway to convert contracted MW into recognized revenue — and revenue into FCF — is longer than the IPO class's marketing documents suggested. Until the CoreWeave tape stabilizes, expect every AI-factory IPO candidate to either price lower, price late, or stay private.

Firmus's withdrawal is the most consequential AI-infra capital-markets event since CoreWeave's March 2025 listing. Three things change because of it. First, the 'AI capex narrative clears public investors' trade is dead: A$30.6B with a $77M loss and a contracting Australian rollout was never going to work, but the speed of the rejection — bookbuild on day one, withdrawal by day two — tells you the bid was thinner than the underwriters thought. Second, the discount Accelevation absorbed ($18 vs $20–24 range) and the gap CoreWeave is trading through ($81.58 against a $153.20 high) jointly set the discount frame for Nscale, [Crusoe], Lambda and Humain. Third, the Nasdaq pivot changes who gets to underwrite these deals — US sell-side desks and US-domiciled funds will own the 2027 AI-infra calendar.