NextFin News - Australia is about to host its biggest initial public offering in years, and the company behind it is a pure bet on the artificial-intelligence infrastructure boom. Firmus Technologies, the Sydney-founded data-center builder backed by Nvidia and Blackstone, is targeting an ASX listing as soon as the end of October with a raise of up to A$7 billion (about US$5 billion), according to people familiar with the matter. The deal would value the company at more than US$10.5 billion, following a US$2 billion equity sale in August 2026 that itself came four months after a US$505 million funding round.
The three developments converging this week - a blockbuster AI data-center IPO, the final stage of the sale of the Whyalla steelworks, and the implementation of President Donald Trump's voluntary AI safety framework - tell a single story about the global economy in 2026: the race to build physical infrastructure for the AI age is colliding with industrial policy and a regulatory framework still trying to catch up.
The Firmus Bet: A$7 Billion on Australia's AI Moment
Firmus was founded in 2019 as a crypto and high-performance-compute operator specializing in immersion cooling. Today it describes itself as a pure "AI factory" builder. That pivot mirrors the broader market: the company is raising capital at a valuation that would make it one of the largest technology listings in Australian history, and it is doing so on the back of contracted demand rather than speculative capacity.
The proceeds are earmarked for Project Southgate, a 1.6GW AI compute buildout across Australia, plus expansion in Indonesia and Malaysia. The Malaysia data centers will serve OpenAI under a capacity agreement signed earlier this month, bringing Firmus's total contracted capacity to more than 900MW. In late August, the company also entered a binding agreement to acquire the "Fabrication, Design, and Projects" business of HVAC developer Benmax, the team behind Firmus's HyperCube modular data-center solution - a move that verticalizes its supply chain ahead of the listing.
The financing stack behind the equity story is substantial. On February 9, 2026, Blackstone-led funds arranged a US$10 billion debt financing facility for Firmus, with participation from Coatue, to support the next phase of Project Southgate. The facility is built on Nvidia's DSX reference architecture and is designed to fund a national rollout of Firmus's AI Factory platform. Put together, the US$10 billion debt facility and the more than US$10.5 billion private equity valuation mean Firmus is attempting to bring roughly US$20 billion of capital-market firepower to bear on Australian data-center construction.
The timing is deliberate. Firmus began meeting investors in Asia this week, with Europe and the United States stops planned before returning to Australia. The company is following a path blazed by fellow neoclouds: CoreWeave went public at the end of March 2025, and Nscale is reportedly looking to raise US$3 billion in an IPO of its own this month. The question for Australian investors is whether Firmus is a once-in-a-generation entry into the AI infrastructure trade - or the top of a cycle that has already run hard.
Whyalla: Industrial Policy Meets the Steel Reality
While Firmus courts growth investors, the Australian government is closing out a very different infrastructure story. The sale of the Whyalla steelworks in South Australia has reached its final stage, with two bidders remaining: M Resources, an Australian-owned Queensland coal miner, and Jindal Steel, the Indian steelmaker. BlueScope Steel, which was among the final five bidders, retains a right of last offer.
The process has been long. More than 70 parties expressed initial interest worldwide, whittled down to five binding bids earlier this year, and now to two. KordaMentha administrator Sebastian Hams has said he is confident the sale will be completed by the end of September. The federal and South Australian governments are investing up to A$1.9 billion in partnership with the incoming owner to transform the site into a modern, low-emissions facility capitalizing on the Upper Spencer Gulf's magnetite resources.
"If we start playing favourites, rather than playing who is the best new owner to invest the capital that is required to give this place a future, we could make a mistake we regret," South Australian Premier Peter Malinauskas said.
The backstory matters. The steelworks' former owner, Sanjeev Gupta's GFG Alliance, was placed into administration in early 2025 after its operating company owed tens of millions of dollars to creditors. GFG remains the largest single creditor at A$536 million. The A$1.9 billion rescue package was assembled in the hope of delivering a future for the business and the city of Whyalla - a sovereign steelmaking capability that successive governments have treated as strategically essential even as the commercial case has been repeatedly questioned.
The physical reality is unforgiving. The iron-producing blast furnace has been offline since April, and Malinauskas has conceded there is a risk of job losses if it cannot be restarted before the site transitions to a direct-reduced iron plant and electric arc furnace. "The $1.9 billion package is what's on the table and there is no suggestion from the Commonwealth, which provides the lion's share of that, to change that," he said. The buyer inherits not just an asset but a race against time.
Trump's AI Order: Voluntary in Form, Mandatory in Effect
In Washington, the regulatory counterweight to the AI buildout is taking shape. On June 2, President Trump signed the executive order "Promoting Advanced Artificial Intelligence Innovation and Security," which establishes a voluntary framework for government review of the most powerful AI models before they are released. Under the order, developers of designated "covered frontier models" can provide the government with access to their models for up to 30 days before a wider release, subject to confidentiality and intellectual-property protections.
The order is explicitly not a licensing regime. It states that nothing in the relevant section authorizes "a mandatory governmental licensing, preclearance, or permitting requirement" for AI models. But the institutional architecture is real: a consortium of agencies led by the National Security Agency must, within 60 days (by August 1, 2026), develop a classified benchmarking process to assess the advanced cyber capabilities of AI models and determine which cross the "covered frontier model" threshold. The order also directs the Attorney General to prioritize enforcement of existing federal criminal statutes - including computer-fraud and wire-fraud provisions - against anyone who uses AI to illegally access or damage a computer.
The final text is a compromise born of internal conflict. An earlier draft contemplated a 90-day review window, which Trump scrapped just hours before a planned signing ceremony last month over concerns it would hurt American competitiveness. The timeline was cut to 30 days to win over skeptics within the administration, including David Sacks, the former AI czar who had opposed the order. "Advanced A.I. capabilities make our nation stronger, but also introduce new national security considerations that require coordinated action across executive departments and agencies," the order said.
The Mechanism: Why These Three Stories Belong Together
On the surface, an Australian IPO, a steelworks auction, and a U.S. executive order have little in common. Beneath the surface, they are three expressions of the same structural shift: the AI economy is forcing a re-pricing of physical and institutional infrastructure, and governments are discovering that markets move faster than policy.
The Firmus valuation is the cleanest read. A company founded seven years ago to cool crypto-mining rigs is now valued at more than US$10.5 billion because it owns the scarce inputs the AI boom needs: power connections, land, cooling expertise, and hyperscaler contracts. The 1.6GW Project Southgate and the OpenAI capacity agreement are not abstractions - they are the collateral that makes the equity story financeable. Blackstone's US$10 billion facility signals that lenders are willing to underwrite AI infrastructure on the strength of contracted cash flows rather than speculative demand.
Whyalla is the same logic applied to a declining industry. Steel is not an AI input in the way GPUs are, but it is an input into the defense and infrastructure buildout that the AI era is accelerating - from AUKUS submarine facilities in Adelaide to data-center construction. The A$1.9 billion government package is, in effect, a subsidy for sovereign capacity: taxpayers are absorbing the risk that a private buyer will not, because the alternative - losing domestic steelmaking entirely - is judged politically unacceptable.
And the Trump order is the recognition that the technology itself is outrunning the guardrails. The voluntary 30-day window is a confession that mandatory preclearance would be both legally fragile and competitively suicidal in a race against China. But the classified NSA benchmarking process creates something real: a government view of what the most capable models can do, and a channel through which national-security agencies can see frontier systems before they hit the market.
Cyclical or Structural? The Call
The central question for investors is whether the AI infrastructure boom is cyclical - a capital-expenditure supercycle that will mean-revert - or structural, a regime shift that will not reverse on its own. The evidence points to structural, but with a cyclical overlay that will produce casualties along the way.
The structural case rests on three pillars. First, the demand is contracted, not speculative: Firmus's 900MW-plus of committed capacity is tied to named hyperscalers, and the OpenAI agreement is a multi-year commitment rather than a spot-market bet. Second, the supply constraints are physical and slow to build: data-center power connections take years to permit and construct, and the specialized cooling and modular-construction expertise Firmus has accumulated cannot be replicated quickly. Third, the policy direction is one-way: both Canberra's A$1.9 billion steel package and Washington's AI framework signal that governments will keep subsidizing and regulating this infrastructure rather than stepping back.
The cyclical overlay is real, however. The neocloud IPO window - CoreWeave in March 2025, Nscale and Firmus in 2026 - is a classic late-cycle financing pattern: private valuations have run ahead of public-market appetite, and the IPO is the exit valve. If hyperscaler capex slows, or if the AI revenue that justifies today's spending fails to materialize at the expected pace, the most leveraged players will find their contracted capacity suddenly looks expensive. The mean-reversion risk is not in the demand for AI itself, but in the pricing of the infrastructure that serves it.
The second-order implication is the one the market is not fully pricing. Everyone understands that AI capex benefits data-center builders. The less obvious chain runs through power, steel, and sovereign risk. A 1.6GW data-center buildout needs transformers, switchgear, and grid connections that are themselves in global short supply. Whyalla's magnetite and the A$1.9 billion steel subsidy are the Australian answer to the same bottleneck. And the Trump order's NSA benchmarking creates a new form of regulatory moat: once the classified threshold for "covered frontier models" is set, compliance becomes a fixed cost that favors large, well-capitalized players and raises the barrier for smaller developers. The voluntary framework, in other words, may end up consolidating the industry it claims to leave free.
The Counter-Thesis
The strongest argument against the structural read is that the AI infrastructure boom is a financing bubble dressed up as industrial necessity. The counter-thesis runs like this: hyperscalers are signing long-term capacity agreements today to lock in supply, but those contracts are only as good as the revenue from AI services that underpins them. If AI monetization disappoints - if enterprise adoption is slower than expected, or if the marginal value of additional inference capacity falls as models become more efficient - the contracted capacity becomes stranded, and the US$10.5 billion valuation looks like the top of a cycle rather than the base of a secular trend. CoreWeave's public-market debut was strong, but it came in a window of exceptional risk appetite that may not repeat for Firmus in October.
This argument has force, and it is backed by the historical pattern of infrastructure supercycles: they end not when demand disappears, but when supply arrives faster than demand and pricing power evaporates. The falsifying signal for the structural thesis is specific: if Firmus prices its IPO at or below the August private valuation of US$10.5 billion, or if the stock trades down more than 20 percent from the offer price within three months of listing, the market will have signaled that the AI infrastructure premium is cyclical, not structural. Conversely, an oversubscribed offering that prices above the private round would confirm that public markets are still willing to pay up for contracted AI capacity.
What to Watch
Three signals will separate the structural from the cyclical over the next year. First, the Firmus IPO itself: the final raise size, the pricing relative to the US$10.5 billion private valuation, and the first-quarter trading performance. Second, the Whyalla outcome: whether M Resources or Jindal Steel wins, whether BlueScope exercises its right of last offer, and whether the A$1.9 billion package proves sufficient or requires top-up funding - a sign that the commercial case remains fragile. Third, the implementation of the Trump order: whether the NSA benchmarking framework is delivered on schedule, and whether major model developers actually volunteer for the 30-day pre-release review. High participation would validate the voluntary model; mass opt-out would expose it as theater.
The base case is that Firmus lists successfully at a premium to its private valuation, Whyalla sells to one of the two shortlisted bidders with BlueScope standing aside, and the Trump framework operates as a soft-touch coordination mechanism rather than a hard constraint. The downside case is a delayed or downsized IPO, a Whyalla deal that requires additional taxpayer support, and a regulatory framework that fails to keep pace with model capability - leaving governments reacting to incidents rather than shaping them.
The AI boom is being built in concrete and steel as much as in code, and 2026 is the year the bill comes due. Firmus's IPO is the test of whether public markets believe the infrastructure trade has further to run; Whyalla is the test of whether governments can convert industrial policy into commercially viable assets; and the Trump order is the test of whether a voluntary framework can govern technology that no one fully understands. The common thread is that the physical and institutional foundations of the AI economy are still being poured - and the market is paying for them before the foundations have set.
Explore more exclusive insights at nextfin.ai.
