Firmus Raises $2B at $10.5B for AI Factories

The money chasing AI is increasingly flowing into buildings, power, and cooling rather than models. On Thursday, August 6, 2026, Firmus, an Nvidia-backed developer of energy-efficient AI data centers, said it had raised a fully subscribed $2 billion equity round that lifts its post-money valuation above $10.5 billion — nearly double the level it commanded just four months earlier.

The round drew a roster of investors more often associated with mega-deals than with a company that started life as a bitcoin miner. It includes follow-on participation from Coatue and Nvidia, alongside new capital from funds managed by Blackstone Tactical Opportunities and other Blackstone vehicles, with additional participation from the trading firm Jane Street. Firmus said the proceeds will accelerate the rollout of its flagship Project Southgate AI-factory network across Australia and fund early expansion into other Asia-Pacific markets, including Indonesia.

From bitcoin mining to AI factories

Firmus is not a household name, but its trajectory captures how quickly the infrastructure layer of the AI boom has re-priced. The Singapore-based company was founded in 2019 and cut its teeth on cryptocurrency mining before pivoting to AI compute — a path that gave it early expertise in the two things AI data centers need most: dense power and aggressive cooling.

That pivot has been rewarded on a compressed timeline. In April 2026, Firmus raised roughly $505 million in a round led by Coatue at a post-money valuation of about $5.5 billion. Four months later, the new $2 billion raise values the company north of $10.5 billion. Doubling a valuation in a single quarter is the kind of move that, outside AI infrastructure, would draw skepticism; inside it, it has become almost routine as capital races to fund the physical buildout behind the AI capex boom.

Equity is only part of the capital stack. Firmus has separately secured a $10 billion debt facility led by Blackstone and Coatue — described as one of the largest private debt financings in Australian history. The combination of fresh equity and a large debt line is telling: data centers are capital-intensive, long-lived assets, and financing them increasingly resembles project finance for power plants more than venture funding for software.

What Project Southgate is

Firmus’s flagship asset is Project Southgate, a network of what the company calls “AI factories” — purpose-built campuses designed to run dense clusters of Nvidia GPUs at high utilization. The initial construction program, centered on a campus in Launceston, in northern Tasmania, has been described as a roughly $4.5 billion build, structured as modular, fully liquid-cooled facilities. Firmus develops Southgate in partnership with CDC Data Centres and Nvidia, and builds on Nvidia’s DSX reference architecture for AI factories.

The company’s pitch rests on efficiency. Firmus uses liquid cooling — variously described as immersion and direct-to-chip — in which coolant carries heat away from the silicon far more effectively than moving air. Firmus claims its combined software and cooling approach uses 33% less energy and 99% less water than conventional data centers. Those numbers are the company’s own, but the underlying problem they target is real: as our look at AI data center economics laid out, power and cooling have become the binding constraints on how much compute a given site can actually deliver.

Efficiency claims aside, the strategic logic is that a site drawing less power per unit of compute can pack more accelerators into the same grid connection — a decisive advantage in markets where new power is scarce.

Why the smart money is buying infrastructure

The investor list is the story as much as the number. Blackstone, Coatue, and Jane Street are not thesis-stage venture funds; they are large pools of capital looking for durable, cash-generating assets. Their presence signals that AI data centers are increasingly being underwritten like infrastructure — with debt, long contracts, and physical collateral — rather than bet on like startups.

Nvidia’s participation is its own signal. The chipmaker has spent the past year threading its capital and reference designs through the companies that deploy its hardware, ensuring demand for its accelerators is matched by the power, cooling, and buildings needed to run them. Backing an efficient AI-factory developer in a power-constrained region extends that strategy. It is the same instinct visible across the industry, from hyperscaler joint ventures like the Meta–BlackRock data center venture to purpose-built “neoclouds” such as SoftBank’s SB Neo.

The timing tracks the demand curve. Each new generation of Nvidia silicon, including the Rubin platform, raises the power and cooling requirements per rack, which in turn raises the value of operators who can actually deliver dense, efficient capacity. Firmus is selling exactly that scarce capability, in a part of the world where AI infrastructure is still relatively thin.

The Australian and Asia-Pacific angle

Firmus’s geographic focus is deliberate. Australia offers access to renewable power, cool climates in places like Tasmania that reduce cooling loads, and a stable regulatory environment — attributes that matter more as the strain AI puts on electricity grids becomes a first-order constraint on where compute can be built at all. The company has signaled ambitions beyond Australia, with expansion into markets including Indonesia as part of a broader Asia-Pacific push.

A public listing has hovered over the company for months. Firmus has discussed an eventual float — most often mentioned in the context of the Australian Securities Exchange — though no IPO has been confirmed. The latest raise, at more than double the spring valuation, both reduces the urgency of a near-term listing and raises the stakes of any eventual one.

What it means

Firmus’s raise is a clean illustration of where AI capital is going: not into another model, but into the power-dense, water-sipping real estate that models run on. When Blackstone, Coatue, Nvidia, and Jane Street collectively write a $2 billion equity check and back it with a $10 billion debt line, they are treating AI compute capacity as an infrastructure asset class — one they expect to throw off contracted cash flows for years.

The winners are operators who can solve the physical bottleneck. Firmus’s bet is that efficiency — less power and water per unit of compute — is a durable edge in markets where grid connections, not chips, are the scarce resource. If its efficiency claims hold up in production, it can build where less efficient rivals cannot, and Nvidia gains another well-capitalized channel for its hardware in an underserved region.

The risks are the same ones stalking the entire buildout. Valuations in AI infrastructure are moving faster than revenue, and a company doubling its worth in a quarter is exposed if demand growth or financing conditions soften. Heavy reliance on debt makes the model sensitive to interest rates and to the pace at which capacity gets contracted and filled. And efficiency figures of “99% less water” are self-reported — the kind of claim that invites scrutiny as these facilities scale from pilot campuses to grid-straining clusters.

What to watch next. Track how quickly Project Southgate’s Tasmanian and mainland Australian capacity comes online and gets contracted, since utilization is what turns an expensive building into a cash-generating asset. Watch whether the Indonesia and wider Asia-Pacific expansion materializes on schedule. And watch the IPO question: a listing would force Firmus to disclose the economics behind its efficiency pitch, and reveal whether a $10.5 billion valuation is underwritten by contracts or by the momentum of the moment.