Nvidia-Backed Firmus Pulls $5 Billion IPO as AI Infrastructure Doubts Grow

Firmus, an Australian AI cloud and data center operator backed by Nvidia and Blackstone, called off a planned initial public offering of about $5 billion on Friday, October 9, according to TechStartups, which cited reporting by CNBC and Reuters. The listing would have valued the company at roughly $30.6 billion. For enterprise leaders who rent GPU capacity from specialist providers, the episode is a reminder that the companies building AI compute are still financing much of it on promises of future capacity.

What happened to the Firmus IPO?

Firmus had planned to list on the Australian Securities Exchange at a reported A$11 per share. The offering would have ranked as the second-largest new share sale in Australian history, TechStartups reported. The target valuation was close to three times what investors had put on the company only two months earlier.

The board said the terms on offer did not reflect the strength of the business, and that going ahead was "not in the best interests of the company and its shareholders." It pointed to market volatility and unfavorable terms. Firmus said it would now look to private markets and consider other public and private options.

Reuters, as summarized by TechStartups, said investors raised concerns about the valuation, the company's financing needs, its short operating history, construction risk at its data centers, and terms that affected existing shareholders. The Financial Times reported that investors focused on the gap between the capacity Firmus plans to build and the capacity already earning revenue.

How much AI capacity does Firmus actually run?

In a September 8 announcement, Firmus said it had more than 900 MW of contracted capacity across its customers and seven AI factories planned in four countries. Two of them, in Australia and Singapore, were operating. The other five were still under development, with ready-for-service targeted within 24 months.

The customer list is strong. In the same announcement, Firmus said OpenAI had signed a multi-year deal for dedicated compute at two planned sites in Malaysia, making it an anchor customer. Meta announced a separate agreement in September for GPU capacity at Firmus's Southeast Asian facilities, according to TechStartups. In August, Firmus raised $2 billion at a valuation above $10.5 billion from investors including Nvidia, Coatue, Blackstone and Jane Street.

Contracts with frontier AI labs make neoclouds like Firmus attractive, but they also concentrate risk. A contracted data center still has to be financed, powered and built.

Why are investors questioning AI infrastructure now?

Firmus pulled its listing at the end of a week in which the economics of the AI buildout drew close attention. On October 8, TechCrunch reported, citing the Financial Times, that OpenAI had told investors its annualized revenue was approaching $50 billion. A figure near $70 billion had circulated a week earlier. The FT attributed the difference to how the two leading labs count revenue: Anthropic includes sales made through its cloud partners, while OpenAI does not.

The revenue question matters to infrastructure providers because AI labs are among their largest customers. When investors become less certain about how fast lab revenue is growing, they look harder at companies whose growth depends on those labs' compute contracts.

None of this means demand for AI compute is falling. But public market investors signaled that they want to see operating capacity and revenue, not only signed agreements and planned megawatts.

What should enterprise AI buyers take from this?

Few large enterprises buy from Firmus directly, but many rent GPU capacity from specialist providers, or through hyperscalers and model vendors that depend on them. A provider's ability to raise money affects whether promised capacity arrives on schedule.

Data and AI leaders at AI Connect events have warned about this kind of exposure for some time. In a AI Connect webinar on future-proofing AI infrastructure, Stefan Leichenauer, VP of Engineering at SandboxAQ, cautioned that a chosen vendor "may not even exist two years from now." He advised using open frameworks for APIs and testing several vendors in proofs of concept.

Practical steps follow from that:

• Ask what is live. When a provider quotes capacity, ask how much is energized and serving customers today, and what delivery dates depend on construction still underway.
• Keep workloads portable. Containerized inference and open interfaces make it easier to move if a provider slips. Panelists at a AI Connect session on AI infrastructure at scale made the same case for open, swappable layers.
• Write delivery terms into contracts. Service credits, step-in rights and exit clauses matter more when a supplier's next funding round is uncertain.
• Watch total cost, not just unit price. As Varun Tottempudi of Lockton warned at COLLIDE, without monitoring "in six months you might see a bill" as large as the past five years of spending. More on that in our look at AI ROI and GPU costs.

Next to watch: where Firmus finds its capital, and whether other AI infrastructure companies planning listings adjust their terms. The answers will shape how much new GPU capacity enterprises can rent in 2027.

Frequently asked questions

Why did Firmus cancel its IPO?

Its board said the terms on offer did not reflect the business's strength, citing market volatility. Investors reportedly questioned the valuation, financing needs and construction risk.

Who are Firmus's customers?

Firmus has said OpenAI is an anchor customer for two planned Malaysian sites, and Meta announced a GPU capacity agreement for its Southeast Asian facilities in September.

What does this mean for enterprises renting GPU capacity?

A provider's access to capital affects whether promised capacity arrives on time. Enterprises should verify live capacity, keep workloads portable and negotiate delivery and exit terms.

Filed under: Data & AI Infrastructure, AI Strategy & ROI