Everything About AI
NextFin News - Firmus Technologies, the Nvidia-backed builder of liquid-cooled "AI factories," is set to begin investor meetings for an initial public offering on the Australian Securities Exchange, setting the stage for what could be the country's largest technology listing — and the most scrutinized founder story in years. The meetings come after a US$2 billion private raise in July that valued the company at US$15.5 billion on a post-money basis, nearly tripling the US$5.5 billion valuation it carried only three months earlier.
The float, targeted for the final quarter of 2026, will put a price on one of the most polarizing bets in the global AI infrastructure rush: a company with US$10 billion of Blackstone-led debt, a US$30 billion offtake pipeline from a dedicated 360-megawatt campus in Indonesia, and a co-founder whose path to the C-suite ran through a 2016 insider-trading conviction and a year in prison.
The Deal and the Timeline
Firmus and its advisers are seeking to put the public offering in front of investors late in the September quarter, with a listing in the final three months of the year, once the company has secured financing and customer commitments for its Batam, Indonesia data centre development and firmed up plans for a third Australian site. The sequencing matters: the company has already lined up the equity. In July, it called an extraordinary general meeting for July 31 asking existing shareholders to vote on its biggest equity raising to date, offering preference shares at US$230 apiece and lifting its equity value to US$15.5 billion. That followed a US$505 million strategic round in April led by Coatue Management with continued participation from Nvidia, and a US$10 billion debt facility in February led by Blackstone Tactical Opportunities and Blackstone Credit & Insurance.
The capital stack now resembles a small utility's balance sheet more than a typical technology start-up's. Roughly US$15.5 billion of equity value sits on top of US$10 billion of long-dated infrastructure debt, structured on the premise that contracted AI data-centre assets are bankable the way toll roads and power plants are. If the public offering raises the US$2 billion the company has been working toward, Firmus would rank among the largest technology listings in Australian history — and the ASX has been angling to win the mandate away from Nasdaq.
But the timetable has already slipped once. As recently as April, the company was targeting a June or July debut. The push into the December quarter reflects the conditions the company set for itself: financing and customers for Batam, plus a firmer plan for the next Australian site. In other words, the IPO is no longer a bet on a vision — it is a bet on contracts that have to be signed, and signed soon.
The Asset: AI Factories, Not Data Centres
Firmus's central pitch is that it is not building data centres at all. It is building what it calls "AI factories" — modular, fully liquid-cooled campuses purpose-built for Nvidia's densest GPU clusters. The distinction is not cosmetic. Liquid cooling, the company says, cuts energy consumption by up to 60 percent compared with conventional air-cooled facilities and roughly halves construction costs. If those metrics hold at scale, they change the unit economics of large AI training runs, which are among the most power-intensive industrial processes in operation.
The flagship is Project Southgate, a US$4.5 billion construction programme anchored in Launceston, Tasmania, where a hydroelectric grid gives Firmus a low-carbon power profile most rivals cannot match. The first stage, delivering 90 megawatts of AI infrastructure, is targeted for completion in 2026, and the site is designed to eventually house 36,000 Nvidia GB300 Grace Blackwell chips. The plan then expands to Melbourne, Sydney, Canberra and Perth, taking total capacity to 1.6 gigawatts across five Australian locations by 2028, with a projected total build cost of US$73.3 billion developed alongside Nvidia and listed data-centre operator CDC Data Centres.
Beyond Australia, the company announced in June a strategic compute partnership with Nvidia running through 2034, anchored by a dedicated 360-megawatt Nvidia DSX AI Factory campus in Batam, Indonesia, developed with Singapore-based DayOne. The agreement covers up to 170,000 Nvidia AI accelerators across Grace-Blackwell, Vera-Rubin and Vera platforms through 2027 and 2028. Based on customer commitments, Firmus expects to receive between US$25 billion and US$30 billion from committed offtake agreements during the first six years of the partnership, with operations expected to start in 2027.
AI-Native companies need access to scalable, energy and cost-efficient compute infrastructure to compete globally. This partnership with NVIDIA provides AI-Natives with unprecedented access to the most advanced AI accelerators in the world, with the certainty, scale, and flexibility that best fits their high-growth trajectory.
Tim Rosenfield, Firmus's co-founder and co-CEO, said in the June announcement. The company has also framed the Australian build-out as a springboard for regional expansion. "This investment reinforces Australia's role in global AI infrastructure while accelerating our Asia-Pacific growth strategy," Oliver Curtis, Firmus's co-founder and co-CEO, said in the April capital raise. "Project Southgate provides a strong foundation for exporting efficient AI compute globally, and our experience in Singapore gives us a proven blueprint for scaling across Southeast Asia."
The customer side is not entirely undisclosed. Firmus's biggest customer is US social-media giant Meta Platforms, which holds the right to acquire the group's compute capacity at any or all of the Australian sites under a deal agreed earlier this year. That concentration is both the strongest validation of the technology and the single largest customer-risk line on the balance sheet.
The Founder Problem
No Australian IPO in years has carried a founder overhang like this one. Oliver Curtis, Firmus's co-founder and co-CEO, was convicted of insider-trading conspiracy in 2016 and sentenced to two years' imprisonment, serving about a year at Cooma Correctional Centre after profiting roughly A$1 million from the offence, according to the Australian securities regulator. He is also the husband of Sydney socialite Roxy Jacenko, which has kept the story in the public eye well beyond the financial pages.
The ASX applies a "good fame and character" test to people in control of listed entities, and Firmus declined to comment when asked whether Curtis satisfied it. In March, Curtis told reporters he would step aside if the board determined he was not the right person to lead the company through a public float. The question for institutional investors is not whether Curtis can raise money — the US$2 billion July round, fully subscribed at US$230 a share, shows he can. It is whether a public-market register of retail and institutional holders will tolerate a founder whose redemption narrative is inseparable from a criminal conviction.
There is a precedent for the market separating the executive from the enterprise: other Australian companies have listed with controversial founders and traded on the numbers. But those were not floatations arriving into a market suddenly attentive to governance after a decade of AI hype. The discount, if there is one, will show up in the offer pricing.
What the Market Is Pricing — and What It Is Not
The valuation trajectory tells the story of how fast the AI infrastructure trade has been repriced. Firmus was valued at US$5.5 billion after the April raise. By July, that was US$15.5 billion. That is not a company compounding cash flows; it is a company re-rating on the scarcity value of power-connected, liquid-cooled GPU capacity in a market where hyperscalers cannot build fast enough.
The bullish case is straightforward and largely priced: AI compute demand exceeds supply, power is the binding constraint, and Firmus has secured power, debt and a chip supply line from Nvidia. The US$25 billion to US$30 billion offtake range over six years, if realized, would validate the US$15.5 billion equity value many times over. Wilson Asset Management's Shaun Weick called the Nvidia arrangement for the Indonesian campus "materially positive" for the IPO valuation.
The bear case is less discussed and less priced. First, offtake commitments are not revenue until capacity is built and powered — and the Batam campus does not start operating until 2027. Second, the US$73.3 billion total build programme requires far more capital than the company has raised; every tranche of new equity or debt dilutes or leverages the existing stack. Third, the Meta concentration means a single customer's capex cycle could swing the Australian business. Fourth, the technology bet is real but unproven at the scale promised: liquid-cooled density advantages must survive the transition from pilot campus to a 1.6-gigawatt network.
There is also the market-timing question. The AI infrastructure trade has been bid up on the assumption that the hyperscaler capex cycle has years to run. If AI spending slows, or if Nvidia's own guidance disappoints, private valuations marked up in July can look very different by December. Firmus is not just selling compute; it is selling a duration asset into a market that has not yet been tested by a sustained AI capex slowdown.
Cyclical or Structural: The Real Question
The central judgment for investors is whether Firmus is riding a cyclical capex wave or capturing a structural shift in how compute is built. The evidence points to both, and confusing them is the most common way to misprice the stock.
The cyclical leg is the hyperscaler capex cycle. It has been unusually strong, it is front-loaded, and it will mean-revert — every infrastructure build-out in technology history has paused to digest capacity. When it does, utilization rates fall, pricing power weakens, and the equity value of promised offtake compresses. That is the cycle, and it is not different this time.
The structural leg is different. Power is now a genuine, physical constraint on AI compute, not a line item. A company that can deliver liquid-cooled, low-carbon, grid-connected capacity at lower cost per token has a durable advantage that survives the cycle — but only if the engineering claims hold at scale and only if the customer base diversifies beyond Meta. The structural claim rests on three conditions: the 60 percent energy saving must be verified in operation, the 1.6-gigawatt network must actually be financed and built, and the customer roster must broaden. Fail any of those, and the structural story collapses back into a cyclical trade.
The second-order implication is the one the market is not fully asking. Even if Firmus executes perfectly, a successful IPO at US$15.5 billion transfers the AI infrastructure risk from private backers — Coatue, Nvidia, Blackstone — to public holders at the top of the cycle. The early investors have already captured the re-rating. Public-market buyers at the offer price are buying the execution risk, not the multiple expansion.
What to Watch
The falsifying signal for the bullish case is specific: if Firmus lists without signed offtake contracts covering a material share of the Batam campus, or if the offer prices below the US$230 preference-share level at which private investors bought in July, the scarcity-value thesis is impaired. Conversely, a listing above US$230 with disclosed long-term contracts would confirm that public markets are willing to underwrite the AI-factory model at private-market valuations.
Over the short term, watch the offer size and pricing, the escrow terms for early investors — past reports suggested early holders could sell on day one, which would add supply pressure — and any disclosure on Curtis's role post-listing. Over the medium term, the 90-megawatt first stage of Project Southgate due in 2026 is the first real-world test of the unit economics. Over the long term, the question is whether Firmus becomes a power-and-compute utility with recurring revenue, or a capital-intensive builder whose returns are perpetually diluted by the next raise.
The base case is a listed float in the December quarter at a valuation between the July US$15.5 billion private mark and a discounted public price, with heavy two-way trading as early investors test liquidity. The upside case is a fully subscribed offer above US$230 with contracted offtake disclosed, which would make Firmus the bellwether for private AI infrastructure valuations entering public markets. The downside case is a delayed or cut-back offering if Batam financing or customer commitments are not locked in, which would force a repricing of the entire Australian AI infrastructure trade.
The bottom line: Firmus is not really selling GPU capacity to public investors. It is selling a claim on the AI capex cycle before that cycle has been stress-tested — and the price at which it lists will tell the market whether the AI-factory model is a new asset class or just the latest way to package a boom.
Explore more exclusive insights at nextfin.ai.
