Firmus Pitches A$7B ASX IPO With OpenAI as Anchor and Nvidia in the Stack

Firmus Technologies launched investor roadshow meetings in Asia this week for an initial public offering on the Australian Securities Exchange targeting as much as A$7 billion, or roughly US$5 billion, with a listing aimed at late October, Bloomberg reported on Monday. At that size, the deal would be the largest Australian float since Telstra's 1997 privatisation — surpassing Medibank Private's A$5.68 billion offering in 2014 — and the first Asia-Pacific AI factory operator to reach a major exchange with OpenAI as a named anchor tenant. No prospectus has been lodged; size and timing remain subject to change.

The company's contract book is the spine of the roadshow. On September 8, Firmus said it had passed 900 megawatts of contracted capacity across its portfolio and named OpenAI as an anchor customer under a multi-year agreement for dedicated compute at two sites in Malaysia. It did not disclose the megawatts, pricing, or dollar value of the OpenAI deal. "These new data centers in Malaysia will help us serve growing demand for OpenAI's products," Sachin Katti, OpenAI's vice president of compute strategy, said in the announcement. [LINK: Firmus September 8 2026 900 MW contracted OpenAI anchor press release] Across seven planned AI factories in Australia, Singapore, Indonesia, and Malaysia, two are operating and five are targeted for service within 24 months. All five will run on Nvidia's Vera Rubin GPU platform, deploying successive generations of compute inside Firmus's proprietary HyperCube modules.

The week that roadshow meetings began, OpenAI's chief executive said on Saturday that his company would not go public in 2026 and that the AI industry should pace frontier development. [LINK: Reuters Sam Altman OpenAI no IPO 2026 pace frontier] The company his firm is anchoring in Malaysia is planning to list in six weeks.

What Firmus Actually Builds: HyperCube Architecture and the Nvidia DSX Stack

Firmus describes itself as an AI factory developer rather than a data center operator, and the distinction is not marketing semantics. Conventional data centers were designed for general-purpose computing and retrofitted for AI workloads, which means cooling is an afterthought, power distribution is overbuilt for redundancy rather than optimized for compute density, and the relationship between GPU clusters and physical infrastructure is governed by separate management systems. Firmus built its infrastructure from the opposite direction: compute, cooling, power, and grid connection designed together from the start.

The physical unit of that design is the HyperCube, a proprietary module housing 32 NVL racks in a primarily liquid-cooled, high-density configuration. [LINK: Firmus infrastructure HyperCube specifications] In Singapore, the system uses immersion cooling — servers submerged in dielectric fluid that draws heat away far more efficiently than forced air. In Tasmania, for next-generation Nvidia GB300 chips, Firmus uses direct-to-chip cooling, where liquid is delivered directly to the chip surface via cold plates. The company reports a power usage effectiveness (PUE) as low as 1.1 at its existing factories — a figure that compares favorably to the hyperscale industry average of roughly 1.3 to 1.4, though this claim is company-stated and has not been independently audited across all sites. [LINK: Firmus engineering principles PUE efficiency data] Firmus also claims its liquid-cooling approach cuts construction cost roughly in half compared to air-cooled builds, providing an economic argument for why a company that raised its first A$330 million one year ago can underwrite 900 megawatts of contracts.

The HyperCube modules are prefabricated in regional New South Wales through a manufacturing process that Firmus accelerated in August when it acquired Benmax's fabrication, design, and projects businesses. [LINK: Firmus Benmax acquisition August 2026] Software governing the physical plant runs on AI FactoryOS, Firmus's proprietary platform that coordinates cooling, power systems, and GPU telemetry as a unified control layer rather than a collection of bolted-on management tools.

The technology backbone for all new deployments is Nvidia's DSX platform, a reference architecture stack that Nvidia launched formally at GTC 2026 and describes as covering the full AI factory in five layers — energy, chips, infrastructure, models, and applications. [LINK: Nvidia DSX platform announcement] DSX Reference Design provides validated architectures from chip to grid; DSX OS provides open-source software for factory lifecycle management and multi-tenant operations; DSX Sim allows operators to model the factory in simulation before spending on construction; and DSX Flex connects AI factories to grid signals including load shedding, demand response, and renewable energy pricing. Under the Batam agreement signed in June, Nvidia supplies DSX as the technology standard for a 360-megawatt campus designed to house 170,000 accelerators. The partnership runs through 2034 and is structured so that Nvidia earns product revenue plus a share of cloud revenue generated from the facility — a commercial arrangement that goes substantially beyond the chip-supply and equity-investment relationships Nvidia maintains with CoreWeave and Nebius. [LINK: Firmus Nvidia Batam 360 MW 170000 GPU campus announcement]

Read more: Nvidia Vera Rubin: first benchmarks show 30x agentic efficiency gain

Contracted vs. Running: Five Sites in 24 Months Is the Entire Bet

The single number that matters most for the IPO is the one that is hardest to verify: how much of 900 megawatts of contracted capacity translates into operating revenue before the valuation's assumed earnings window closes.

Two factories are operational — the Singapore SMC Cloud, which runs on Nvidia A100 and H100 hardware in immersion-cooled HyperCubes inside STT GDC facilities and previously trained the SEA-LION family of Southeast Asian language models, and an early Tasmania site. The flagship Launceston campus (Project Southgate, 90 MW in its first stage, 36,000 Nvidia GB300 chips) is under construction and targeted for 2026 completion. The Batam campus — 360 megawatts, 170,000 accelerators — is not expected to be operational until the first quarter of 2027. The Malaysian sites anchored by OpenAI have no confirmed completion date in publicly available material. Five sites in 24 months is a delivery schedule that requires Nvidia's Vera Rubin platform to arrive on time, power connections in South Australia to be energized on schedule, and construction timelines in Indonesia and Malaysia to hold.

The capital structure behind that schedule is substantial by any measure. Firmus has raised more than US$3 billion in equity since September 2025 from Coatue Management, Nvidia, Blackstone, Jane Street, and Ellerston Capital, reaching a post-money valuation above US$10.5 billion after an August round. The US$10 billion Blackstone-led debt facility, which closed in February with Bank of America, JPMorgan, Morgan Stanley, and Morgans participating, is structured to draw against executed long-term customer contracts — meaning each disbursement is tied to a specific signed agreement. [LINK: Gilbert Tobin Firmus USD10 billion Blackstone debt facility] That structure is not unusual for infrastructure project finance, but it means the debt cannot simply scale with ambition; it scales with contracts that clear the lender's underwriting criteria.

One earlier milestone quietly illustrates the execution risk. Melbourne was targeted as an early Project Southgate site, with up to 150 megawatts expected by mid-2026. As of spring 2026, independent commentators noted the cluster had not been publicly confirmed live on schedule. The company has not addressed the status publicly.

How Firmus Sits in the AI Infrastructure Market

The neocloud sector — specialized AI infrastructure providers distinct from hyperscaler general-purpose cloud — went public in 2025 and 2026 with CoreWeave and Nebius as the two reference points that will frame every Firmus conversation with institutional investors.

CoreWeave listed on Nasdaq in March 2025 at US$40 per share, raising US$1.5 billion. By mid-2026, it operated roughly 1.5 gigawatts of active data center capacity and held a contracted revenue backlog of US$104 billion. Its second-quarter 2026 net loss was US$626 million on approximately US$33.8 billion in total debt, with full-year 2026 capex guidance of US$35 to US$39 billion. Its stock has climbed roughly 18 percent year-to-date. Nebius, which Nvidia backed with a US$2 billion investment in March, reported second-quarter 2026 revenue of US$582 million — up 454 percent year-over-year — with a 50 percent adjusted EBITDA margin and potential multi-year agreements with Meta (US$27 billion) and Microsoft (US$17.4 billion through 2031). Its stock is up 146 percent in 2026. [LINK: CoreWeave Q2 2026 results CRWV Nasdaq] [LINK: Nebius Q2 2026 revenue 454 percent Nvidia investment]

Firmus's pitch against those comparisons is built on three claims: geography, hardware, and grid.

Geography: CoreWeave and Nebius are predominantly US and European operators. Firmus's entire network is in Asia-Pacific — Australia, Singapore, Indonesia, Malaysia — in jurisdictions where sovereign AI infrastructure policy is a driving factor and where US-centric hyperscalers face structural limits. AI researchers at national institutions in Southeast Asia do not always have access to US-domiciled GPU compute under standard commercial terms.

Hardware: The HyperCube's liquid-cooling design claims roughly half the construction cost and 40 percent better energy efficiency than standard air-cooled builds. If those figures are accurate — and they remain company-stated — they imply Firmus can deliver more compute per dollar of capital than a conventional neocloud at scale. Nebius and CoreWeave use GPU-dense racks inside standard data center buildings with enhanced cooling; Firmus designs the entire building-chip-grid system as one engineered unit.

Grid: The Gunvor Group agreement, signed in June 2026, secures 600 megawatts of firm electricity for South Australian campuses at Tailem Bend and Stirling North, linked to 1.2 gigawatts of new wind and solar generation and 1.5 gigawatt-hours of battery storage by 2032. [LINK: Firmus Gunvor 600 MW power agreement South Australia] The 220-hour annual demand-response commitment and the DSX Flex integration mean Firmus's factories can respond to grid signals — potentially earning frequency control ancillary service (FCAS) revenue from the South Australian grid while throttling compute during peak pricing periods. No comparable published AI factory operator has disclosed this level of grid-service integration.

IREN, the other Nvidia-backed AI infrastructure company listed on the ASX, provides the closest domestic market comparator. Its smaller scale and earlier stage mean Firmus would enter the exchange as a substantially larger entity — but also with a substantially more complex capital structure to explain.

The Conflicts and Constraints a Prospectus Has to Price

Three structural issues in the Firmus cap table have no clean resolution in any public document yet filed, and a prospectus will need to address all of them.

The first is the Nvidia overlap. Nvidia is simultaneously Firmus's largest institutional equity investor, its primary chip supplier, its technology-platform provider through DSX, and, under the Batam agreement, a revenue-sharing partner earning a percentage of cloud compute revenue generated from 170,000 of its own accelerators. Roger Montgomery, an Australian fund manager who examined the structure after the April 2026 raise, described the arrangement in public commentary as one that "frankly does not pass the pub test." [LINK: Roger Montgomery Firmus structural risk analysis] That criticism is not dispositive — investors in CoreWeave accepted a similar Nvidia dependency — but the revenue-sharing structure at Batam, where Nvidia earns from the factory it also supplied and partly financed, is a more direct entanglement than CoreWeave's chip-purchase-and-equity model.

The second is the OpenAI anchor. OpenAI is the most scrutinized private company in the world, has not published audited financial statements, and has told investors it will not go public in 2026. A multi-year compute agreement with OpenAI is the strongest possible demand signal for AI infrastructure; it is also the demand signal most dependent on a single counterparty's continued growth and financial health. A Firmus prospectus would be the first document to disclose what it costs to build AI factory capacity for OpenAI's products at scale in Southeast Asia — the megawatts, the pricing structure, the contract term — which makes it, incidentally, the first public pricing window into AI lab infrastructure economics at that geography.

The third is the governance question. Oliver Curtis, co-CEO and co-founder, was convicted of insider trading in Australia for offenses committed in 2007 and 2008, was sentenced to two years' imprisonment in 2016, and served approximately 12 months. [LINK: Oliver Curtis insider trading conviction Australia 2016] The ASX updated its "good fame and character" listing requirements in February 2026 — a timing that multiple independent analysts linked to the anticipated Firmus application, though the exchange did not confirm a connection. The ASX's process for evaluating officer character is confidential, no publicly documented case of ASX refusing a listing on character grounds has been found, and Curtis's backers include some of Australia's most established institutional investors. Whether the ASX imposes conditions on any Firmus prospectus — including requiring disclosure of Curtis's history and the company's representations about it — remains open.

A fourth question, less about conflict than about capital adequacy, concerns scale. More than US$3 billion in equity, US$10 billion in Blackstone debt, and a proposed A$7 billion IPO add up to well under half of the A$73.3 billion programme Firmus announced for its Australian build-out alone. The rest is assumed to come from the offtake model: customer revenue funds construction of the next tranche of sites, which contracts more customers, which funds the next tranche. This is structurally similar to the logic behind every large AI infrastructure commitment signed in 2026, including Anthropic's US$13.7 billion GPU agreement, CoreWeave's US$22.4 billion OpenAI backlog, and the Stargate consortium's multi-year power commitments. It works as long as AI compute demand continues growing at current rates and the anchor customers continue to require more capacity than they can build themselves.

The Prospectus as Price Discovery

Firmus has grown from a A$1.85 billion valuation in September 2025 to a reported A$30 billion target in current roadshow conversations in 12 months. That trajectory reflects the broader repricing of AI infrastructure, but it also means the IPO will be the first moment when the company's financial mechanics are open to public scrutiny rather than closed institutional negotiation.

The prospectus will disclose, for the first time, what OpenAI has agreed to pay per megawatt-hour or per rack in Malaysia. It will show how the Blackstone debt facility's drawdown schedule maps to specific signed contracts. It will describe what happens to the company's financial projections if Vera Rubin delivery slips a quarter or if a South Australian grid connection takes longer than permitted. It will explain how Nvidia's revenue share at Batam is calculated and whether it creates a structural impediment to Firmus switching chip suppliers at any future point. None of that information exists in any current public filing.

That prospectus is still weeks away. The roadshow running this week through Sydney and Melbourne is the institutional education phase — the moment when fund managers examine the story before the document arrives. Australian AI infrastructure investors who cannot buy OpenAI equity directly and who find CoreWeave's US$35 billion debt load difficult to underwrite are being asked to decide whether Firmus's proprietary cooling technology, grid-integrated power model, and APAC-first geography justify a valuation premium or whether the absence of operating history, the concentrated counterparty exposure, and the unresolved governance question require a discount. The October listing, if it proceeds, will produce the first answer to all of those questions simultaneously.

Tim Rosenfield, Firmus's co-CEO, described the OpenAI agreement as "the moment Asia-Pacific becomes a producer of intelligence, not just a consumer of it." The prospectus will show what that production costs, who bears the risk, and at what price public investors are invited to join.