Firmus wants $5bn for megawatts it has already contracted

Firmus is asking public investors to price something its peers have mostly kept off the page: contracted AI demand. The company is seeking as much as $5 billion from an Australian listing that could price as soon as the end of October, according to Bloomberg and the Australian Financial Review, both citing people familiar with the matter. Data Center Dynamics, which relayed the two accounts, notes the size and timing remain subject to change—a hedge that matters when the headline number runs roughly two-thirds larger than the $3 billion Nscale is reported to be seeking this month.

The demand side is named: OpenAI signed a capacity agreement with Firmus earlier this month covering its Malaysia data centers, a deal the reports credit with pushing total contracted capacity past 900MW. The company began meeting investors in Asia on Monday and intends to add Europe and the United States before returning to Australia, a route that reads as a global bookbuild rather than a domestic one—the right register for an Australian issuer whose contracted load sits in Malaysia. Founded in 2019 as a crypto and high-performance compute operator with an immersion-cooling specialty, Firmus now describes itself as a pure AI factory builder, a pivot that moves its comparables from mining rigs to GPU clouds and puts its equity in front of investors who have been generous with named tenants and stingier with everything else.

Firmus took $505 million in April and $2 billion in August, the later check roughly four times the earlier one, and the August round valued the company at more than $10.5 billion post-money. The two rounds come to about $2.5 billion, so the IPO target is roughly double everything Firmus has raised in equity this year. The coverage does not put a valuation on the listing, which leaves the market to guess where a public mark lands against the August private one; Firmus has declined to comment to other publications, per DCD.

The test in Lambda's $3 billion round was that a report can stack equity and debt ahead of a listing without ever naming the contracted backlog that would justify the price. Firmus has handed the market the other half of that ledger: a tenant, a geography, and a capacity figure. The 900MW is contracted load, and an issuer that can point a roadshow at OpenAI's Malaysia footprint is selling something closer to an infrastructure cash flow than a growth story—the distinction neocloud equity is now graded on, and one an IPO price will make public.

Nscale is reported to be seeking $3 billion in an IPO this month, and CoreWeave listed at the end of March 2025, which leaves the sector's public-market record, on this coverage, at one listing and a queue of reported ones. Firmus would arrive with an anchor tenant on record; the reports summarized here do not say whether Nscale's filing carries a comparable one. Where a neocloud cannot name a buyer, investors are being asked to underwrite the build and the demand in the same check.

Firmus has also been buying its way to a shorter schedule, entering a binding agreement in late August to acquire the Fabrication, Design and Projects business of Benmax, an HVAC and mechanical systems developer that the report identifies as the operation behind Firmus's HyperCube modular data center product. Read that as a lead-time hedge: modular fabrication compresses the distance between a signed capacity deal and a live rack. The gate on AI capacity moved from capital to grid capacity and interconnection queues some time ago, and control of a fabrication line is one way to buy a position that money alone does not clear.

Firmus comes close to a clean test of the capital hierarchy now sorting AI capacity. Anchored capacity—a named counterparty, a signed capacity deal—gets priced like infrastructure; unanchored capacity has to fund its own grid works, fabrication, and permitting before persuading public investors to take the residual. Firmus's contracted load sits on the anchored side, in Malaysia, while what the three Tasmanian projects, or the sites in Singapore and Indonesia, will sell and to whom the coverage does not state.

Geography does quiet work in that split: the four jurisdictions are separate power markets with separate permitting paths, and the argument that consent is now a priced good rather than a public process bites hardest where demand is contracted and the jurisdiction is the variable. The Malaysia campus has the tenant on record; the Australian pipeline carries development risk the coverage leaves unquantified.

The roadshow arithmetic is blunt: $5 billion sought against a little more than 900MW of contracted capacity and the $10.5 billion mark set in August. If the listing prices anywhere near that, the backlog becomes the reference the next neocloud prospectuses are written against, and the sector's pricing power moves from GPU narratives to signed megawatts. A strong print also makes Nscale's $3 billion target look modest and widens the queue; a soft one tells everyone behind Firmus that the window is narrower than the filing calendar implies. The end of October is the window, and Nscale is the print.