A picture shows the logo Bitcoin in the first Italian Bitcoin crypto currency shop “Bitcoin Compro Euro” (meaning I Buy Euro), on December 11, 2017 in Rovereto, northern Italy. Bitcoin surged past $18,000 after making its debut on a major global exchange but was trading lower on December 11, 2017, highlighting the volatility of the controversial digital currency that has some investors excited but others nervous. (Photo by PIERRE TEYSSOT / AFP) (Photo by PIERRE TEYSSOT/AFP via Getty Images)
AFP via Getty Images
Jane Street is a trading firm. Blackstone Tactical Opportunities exists to buy complexity other investors avoid. Neither runs a venture portfolio, and both just wrote checks into an Australian company that started out mining Bitcoin. Firmus, which now builds AI data centers, raised $2 billion in equity this week at a valuation above $10.5 billion, roughly double where it stood in April.
The valuation is the headline. The signal is the roster of who is paying and what kind of asset they think they are buying.
The Round And The Roster
The new money came from existing backers Nvidia and Coatue, joined by funds managed by Blackstone and by Jane Street. Counting this round, Firmus has raised more than $3 billion of equity in the past year. The April round valued the company at $5.5 billion, so the price of admission roughly doubled in four months.
A doubling that fast usually reads as froth, and skepticism is the right reflex. What complicates the froth reading is the character of the buyers. Trading firms and private equity credit desks underwrite cash flows, contracts, and collateral rather than stories. Their presence says the AI factory is being evaluated as infrastructure, the way ports, pipelines, and power plants get evaluated, and that evaluation is happening on a continent most American investors never think about.
From Mining To AI Factories
Firmus began as a Bitcoin mining operation and repurposed itself into an AI infrastructure developer built around Nvidia’s DSX reference architecture and its own HyperCube hardware platform. Its flagship program, Project Southgate, started in Tasmania and is expanding into Melbourne, Sydney, Canberra, and Perth, targeting 1.6 gigawatts of AI compute across five sites by 2028 in partnership with CDC Data Centres. The first phase, valued at A$4.5 billion, began with a 150-megawatt Melbourne build carrying 18,500 Nvidia GB300 GPUs. The full program is projected to reach A$73.3 billion of investment through 2028. A company most American investors have never heard of is building compute at a scale that belonged exclusively to the biggest technology companies on earth two years ago.
Co-chief executive Oliver Curtis described the round’s purpose in the announcement:
“This investment allows us to move on multiple fronts at once. We’re scaling across Australia while fast-tracking our capacity to expand into the wider Asia-Pacific region.”
The expansion target after Australia is Indonesia, then the broader Asia-Pacific. The pitch to those governments is straightforward: sovereign AI compute, built regionally, powered largely by renewables, without waiting in line behind American hyperscaler demand.
The Debt Underneath
Equity is the smaller half of the story. In February, Firmus closed a $10 billion debt facility led by Blackstone Tactical Opportunities and Blackstone Credit & Insurance, with Coatue participating. An AI startup borrowing ten billion dollars from an insurance-backed credit platform is the AI factory becoming a financeable asset class, underwritten the way toll roads and transmission lines are underwritten.
That structure is spreading, and it concentrates risk in a specific place. CoreWeave, the largest of the American GPU cloud operators, ended the first quarter with $25.1 billion of debt and another $10.1 billion in lease obligations. The operator layer of the AI buildout runs on leverage almost everywhere you look. Firmus at least pairs its borrowing with long-dated government-adjacent tenancy through CDC, but the basic shape is the same: borrow now, build for years, and count on tenants filling capacity that does not exist yet.
Where Risk And Reward Split
The bear case on AI capex has always rested on concentration, the idea that a handful of American hyperscalers carry the entire buildout and will eventually flinch. Every new buyer class that shows up weakens that premise. Sovereign-adjacent platforms in Australia, private credit desks, insurance capital, and trading firms are now funding compute demand that exists whether or not any single hyperscaler trims its budget. The demand base for AI infrastructure is widening geographically and financially at the same time, and that breadth is what makes the buildout durable.
The reward and the risk do not land in the same place, though. Suppliers get paid in cash when the hardware ships: Nvidia collects on 18,500 GB300s regardless of Southgate’s eventual occupancy, and the same goes for the networking, power, and cooling vendors behind it. The operators carry a decade of utilization risk and refinancing risk on borrowed money. Companies selling into this buildout are collecting certainty, while companies operating it are absorbing leverage. That distinction, more than any valuation print, is the durable way to read every announcement in this category.
Capital is treating AI factories the way it once treated power plants. The equipment makers get paid on delivery either way, and the operators find out later whether the tenants show up.