Firmus raises another half billion for its Australian AI data centre

The Australian firm's proprietary cooling technology could be the key.

Firmus raises another half billion for its Australian AI data centre
Photo Credit: Unsplash/Joey Csunyo

Firmus has raised another half billion dollars for its Southgate AI data centre in Australia, bringing its valuation to US$5.5 billion. The details hint at something worth paying attention to.

Project Southgate

According to a report on Bloomberg, Firmus has raised US$1.35 billion in the last six months, including this latest transaction. As I wrote previously, Project Southgate is a plan to develop an AI data centre campus in Australia that runs on renewable energy.

What caught my eye in the same report was a reference to the Vera Rubin DSX. This isn't a GPU product, but a reference design for an AI factory, essentially a blueprint for how an entire AI data centre should be architected around Nvidia's next-generation hardware.

From 8kW to 600KW

Vera Rubin is Nvidia's next-generation GPU platform. To be rolled out over the next few years, the roadmap released so far consists of the Vera Rubin NVL72 at 120-130kW, the NVL144 at 190kW, the NVL144 CPX at 370kW, and the Rubin Ultra at 600kW.

To put that in perspective, average data centre rack density was just 8kW a few short years ago. The Rubin Ultra at 600kW per rack represents a 75-fold increase in power consumption. These are numbers that would have sounded absurd not long ago.

Beyond the headline-grabbing power figures, though, Vera Rubin is actually designed for dramatic efficiency gains. The AI token generation capacity of the Rubin Ultra will be even more remarkable, meaning more useful work per watt consumed.

Firmus in Singapore

This is where it gets interesting. Firmus, through a joint partnership with STT GDC called Sustainable Metal Cloud, launched data centres in Singapore to offer GPUs in the cloud. What makes it unique is its proprietary immersion cooling technology. I visited the deployment at STT GDC 6 at Loyang over two years ago.

The design is elegantly simple. Multiple immersion tanks are connected in rows, driven by just one large pump for the entire train. Individual tanks rely on convection circulation, and the system uses condenser water for cooling, which allows for a higher inlet temperature. The result is very high efficiency.

The deployment I saw ran H100 GPUs. I'm assuming the Southgate AI data centre will similarly use immersion cooling, which would mean Firmus has successfully scaled the design well beyond 100kW per rack.

The question I'm most curious about is whether Firmus has cracked immersion cooling at Vera Rubin's power demands, especially at the Rubin Ultra's extraordinary 600kW levels. If it has, that would be a significant milestone for the industry.