• Firmus, an AI infrastructure company backed by Nvidia (NVDA), is poised to delay its Australian IPO after investor demand weakened, forcing bankers to consider a lower price and reduced size.
  • The offering, initially marketed at A$11 per share for a valuation of about US$30.4 billion, was being discussed around A$8–A$8.25 as order books closed Thursday in Australia without a final price or structure, according to people familiar with the matter.
  • The setback highlights growing investor scrutiny of AI infrastructure valuations and leverage, even with prominent strategic backers like Nvidia.

Firmus, the Nvidia-backed AI data centre operator, is expected to postpone its Australian initial public offering after a sharp deterioration in investor demand, people familiar with the matter said, in a blow to what would have been the country’s second-largest listing.

Bankers are working to reprice the deal and potentially reduce its size after order books closed Thursday morning in Australia without a clear final price or deal structure. The shares were originally marketed at A$11 each, implying an equity valuation of approximately A$43.7 billion (US$30.4 billion). But the offer was subsequently being discussed around A$8–A$8.25, according to Australian reporting, a substantial reset that underscores the gap between private-market enthusiasm and public-investor caution.

The company had sought to raise up to US$5.5 billion, including an over-allotment option, but bankers were also discussing lowering the fundraising target. The Australian Financial Review reported that its prospectus and investor presentation were removed from the investor data room on October 8. The latest pre-crisis timetable called for a prospectus on October 12 and trading on October 23, though those dates are now uncertain.

A spokesperson for Firmus declined to comment. The company did not respond to a request for comment on the status of the offering.

A Valuation Test for AI Infrastructure

The central issue is the gap between a rapidly expanding private-market valuation and public investors’ willingness to finance a largely unbuilt, debt-intensive AI infrastructure portfolio. Firmus, founded in 2019 by Oliver Curtis, Tim Rosenfield and Jonathan Levee, builds and operates specialised data centres—often called “AI factories”—that provide the computing infrastructure needed to train and run AI models. It belongs to the “neocloud” category: specialised providers of GPU-based computing rather than broad, general-purpose cloud services. Reuters identifies Nvidia, Meta (META) and OpenAI among its customers.

Its operating footprint is much smaller than its proposed expansion. Reuters reported in September that it had two operational data centres in Australia and Singapore, with five more under development across Asia-Pacific, mostly at early stages. The company is historically loss-making, with a draft prospectus attributing losses to development and expansion costs. Reuters reported a projected US$77 million pro forma loss after tax for the first half of the financial year ending June 30, 2027—a forecast, not a reported result.

In August 2026, Firmus raised US$2 billion in equity from investors including Nvidia, Coatue, Blackstone (BX) funds and Jane Street, taking its post-money valuation above US$10.5 billion. The company has said it aims to produce combined annual earnings of US$5 billion within five years from its developed data centres, a projection contingent on successful construction.

Financing exposure is significant. The Australian Financial Review reported A$4.7 billion of equity raised and combined debt facilities of A$10 billion. Separately, ABC reported potential debt of approximately US$30 billion once planned facilities are operational. Facilities, future borrowing and debt currently drawn are different measures and should not be conflated.

Partnership Breakup Adds to Execution Risk

The IPO difficulties come after Firmus and CDC Data Centres ended their proposed Project Southgate development partnership. Announced in October 2025, it envisaged up to 1.6 gigawatts of Australian AI infrastructure by 2028, but approximately 42 megawatts were deployed under the partnership. Curtis said the companies mutually agreed earlier this year not to proceed and that this did not affect Firmus’s current development plans or contracted customer capacity.

That breakup coincides with a shift toward Firmus’s own developments and a larger Asian portfolio, including projects in Indonesia and Malaysia. It is important background to the IPO’s execution-risk debate, although sources do not establish that it alone caused the financing difficulties.

The planned transaction would have been Australia’s second-largest IPO, behind Telstra (TLS.AX)’s 1997 listing. Its difficulties consequently represent a meaningful test of whether Australia’s public market will support very large AI infrastructure valuations—not simply whether investors believe AI demand will grow.

Three economic questions dominate. First, valuation: the marketed US$30.4 billion equity valuation was far above the US$10.5 billion-plus valuation achieved in August. Investors must judge whether new contracts and development prospects justify that increase. Second, financing and delivery: these projects require large upfront purchases of GPUs, buildings and power infrastructure before anticipated cash flows materialise. Reuters says IPO proceeds were intended to support further capital expenditure; Bloomberg reporting identifies GPU purchases for a Batam, Indonesia, development with DayOne. Third, electricity supply: Accela Research’s Naomi Wagura told ABC that Firmus’s three planned Tasmanian facilities could require approximately 444 megawatts and make it the state’s largest electricity consumer, requiring additional generation. This is conditional on the projects being realised.

Investor Caution and Political Scrutiny

Bloomberg interprets the sudden loss of demand as a warning about the AI funding boom. The narrower, better-supported conclusion is that public investors are becoming more selective about price, leverage and construction risk, even when a company has prominent strategic backers. One troubled IPO does not establish that demand for AI computing has collapsed.

Nvidia’s overlapping positions as investor, customer and equipment supplier also make transaction economics important: IPO proceeds used to buy Nvidia GPUs would support the supplier as well as Firmus. Those relationships warrant careful disclosure and scrutiny, but are not themselves evidence of wrongdoing.

Public reaction has been sharply divided. Ten Cap co-founder Jun Bei Liu told ABC she was avoiding the float because of inadequate detail and structural risks, and said 97% of the promised capacity had yet to be built. Fund manager Roger Montgomery questioned the apparent inconsistency between earlier claims of excess indicative demand and the subsequent need to reduce the price. These are attributed investor assessments, not independently audited conclusions.

ABC additionally reported that Firmus representatives withdrew from a scheduled parliamentary AI inquiry appearance amid the float’s complications. That is evidence of heightened political scrutiny, not a regulatory finding against the company.

Australia is moving toward stronger oversight of large data centres’ electricity, water and land-use impacts. The federal government issued national expectations in March 2026; National Cabinet subsequently endorsed developing mandatory standards in August. These are a developing regulatory framework, not rules that should automatically be assumed already enacted in full. For Firmus and its peers, the significance is that access to land and power may increasingly require demonstrating that a project adds adequate energy supply, manages water use and avoids imposing excessive costs on communities. The standards could affect project costs and timing, independently of the IPO outcome.

A contemporaneous parallel is emerging in Hong Kong: reporting on October 8 described cornerstone investors becoming more cautious and some issuers postponing offerings after weak post-listing performance. That suggests changing IPO risk appetite is not confined to Australian AI infrastructure, although the markets and issuers differ.

The immediate possibilities are a smaller, cheaper IPO; a postponement while Firmus seeks firmer investor commitments; or withdrawal of the current offering. The verified reporting does not yet establish which outcome has been agreed. The most useful next disclosures would be the final fundraising amount, valuation, listing timetable, available funding if the IPO is delayed, and a clear reconciliation of operating capacity, contracted future capacity and total development ambitions.

Morningstar (MORN) strategist Lochlan Halloway’s analysis focuses on valuation and leverage: borrowing against customer contracts to acquire chips can work if facilities and revenues arrive as expected, but debt can amplify downside when assumptions fail. Wagura’s complementary warning concerns the physical economy—power infrastructure and responsibility for costs if demand is delayed. Together, their analyses explain why the key question is not merely whether AI grows, but whether Firmus can deliver profitable capacity at an acceptable financing and social cost.

Update: This article was updated on October 8 to clarify that while a postponement is imminent, the offering had not been formally postponed as of 1:20 p.m. EDT.