Oliver Curtis might just be Australia’s luckiest white-collar criminal.
Nearly a decade after he was convicted of insider trading and whisked off to Cooma Correctional Centre to serve a one-year sentence, the former investment banker, Riverview old boy and husband of Sydney socialite Roxy Jacenko is, if you believe the hype, on the cusp of becoming a billionaire after carving out a little piece of the AI gold rush.
Firmus, the Singapore-based AI infrastructure provider co-founded by Curtis and his cousin Tim Rosenfield, is eyeing off a blockbuster listing on the Australian Securities Exchange this year.
The company’s $6 billion promise is to create cooling technology that allows data centres that support AI infrastructure to operate more efficiently, with a landmark project in Tasmania now under construction. That promise has won it the financial backing of billionaires, superannuation funds, institutional investors, and Nvidia, the Silicon Valley-based AI chipmaker that is the world’s most valuable company.
If Firmus’ bet pays off, Curtis will become fabulously wealthy – and proof that, in Australian lives, there are second acts, at least for some. Curtis’ former best friend and co-accused John Hartman, who he’d known from childhood at St Ignatius College, Riverview, the elite North Shore private school that previously educated Tony Abbott, Barnaby Joyce and the like, has also done well.
Hartman, who testified against his former friend in the 2016 criminal trial, was snapped up by billionaire mining magnate Andrew “Twiggy” Forrest within months of his release from prison. He’s now chief executive of the Forrests’ private investment firm Tattarang.
Curtis, too, was back in business months after leaving jail via a waiting Range Rover that transferred him to a private jet. In 2017, he co-founded e-Nome, a cryptocurrency medical start-up with his father, Nick Curtis, a former mining executive.
“Oli is back in the saddle in a big way,” Curtis senior told The Daily Telegraph at the time. “We are really proud to have him leading the team.”
Firmus was approached by this masthead about an interview with Curtis, but the company instead wanted questions in writing. It replied with technical points, detailed in the story below.
E-nome, which purportedly allowed users to access and share their medical data from their smartphones via the blockchain, went nowhere. But two years later, Firmus was born, founded by Curtis, Rosenfield and Jonathan Levee, with Nick Curtis serving as inaugural chair. Back then, AI wasn’t yet a buzzword. Instead, cryptocurrency was the flavour of the month. Firmus had developed a liquid cooling technology that it hoped would allow bitcoin mining operations to become more efficient.
“The original vision was to redesign the way that compute power could be efficiently cooled at scale,” Ted Pretty, a former Telstra executive who took over from Nick Curtis as chair, said in a 2024 company update.
“This vision came from the team in the early days of cryptocurrency mining, where the immense energy requirements highlighted the need for a more efficient cooling solution.”
After bitcoin prices began to fall, Pretty, who stepped down from the board last year, helped pivot the technology towards deployment in AI data centres.
Firmus builds, owns and operates what the industry calls AI factories – purpose-built data centres designed to house thousands of high-powered computer chips that train and run artificial intelligence systems. These facilities are essentially vast, climate-controlled warehouses of computing power, and the hotter and more densely packed the chips become, the more sophisticated the cooling technology required to stop them melting.
Firmus’ pitch is that its proprietary liquid cooling systems – which submerge or pipe fluid directly onto the chips – can do this more cheaply and efficiently than anyone else, giving it a cost advantage it likens to a low-cost mine sitting at the bottom of the commodity curve.
None of the three co-founders have a technical or engineering background. Rosenfield was previously chief executive of lingerie company Simone Perele, while Levee, now Firmus’ director of research and development, had some middle management experience in the resources sector.
But big backers were quickly showing faith.
Key to Firmus’ early success was a 2023 $US100 million ($140 million) investment from Singapore data centre provider ST Telemedia Global Data Centres, to use the company’s cooling technologies in its projects.
Pretty soon, the company was attracting a conga-line of high-profile local investors, Curtis’ criminal past be damned. Early backers include billionaire investor Alex Waislitz, through a company then co-owned with his ex-wife and Visy heiress Heloise Pratt.
Private credit lending group Archibald Capital, run by Ben Madsen, one of Curtis’ friends who attended his 40th birthday bash in Greece last year, is also an investor. Boutique Northern Rivers-based investment manager Tectonic has a 7 per cent stake.
At a recent funding round, last September, Firmus raised $330 million, including crucial backing from Nvidia, the US-based juggernaut, an investment that will afford the company huge legitimacy in the tech world.
Nvidia’s backing is perhaps the single most powerful piece of validation Firmus possesses. When the world’s most valuable chipmaker puts its name and its capital behind a company, it signals to the rest of the market that the technology works and the opportunity is real. But Nvidia’s investment strategy is broad by design. The chipmaker has taken stakes in dozens of AI infrastructure companies worldwide – from US neocloud giant CoreWeave to Middle Eastern ventures – seeding an ecosystem of customers for its own hardware rather than picking a single winner.
Other backers in that round included Ellerston Capital, industry super fund UniSuper and Wilson Asset Management. And last month – a $US10 billion debt financing package from American private equity giant Blackstone.
The company’s Project Southgate initiative – a nationwide rollout of AI factories in partnership with Nvidia and CDC Data Centres – aims to deliver 1.6 gigawatts of infrastructure by 2028, with its flagship project outside Launceston set to open this year.
Firmus’ blue-chip backers and skyrocketing valuation (it tripled to $6 billion between September and November last year), which have paved the way for an IPO this year, reflect the warp-speed of the AI boom, and investors’ faith in the company’s promise of allowing data centres to operate cheaper and greener.
But not everyone is convinced.
In investor presentations last year, Curtis claimed the company’s cooling systems deliver operating costs roughly half those of competitors, and build costs about 50 per cent cheaper.
“We operate these factories at a cost profile that is about half of our competitors’, and we build these things at about half the cost,” Curtis told the Morgans Investment Breakfast last August.
These are extraordinary claims. Globally, hyperscalers such as Microsoft, Google and Meta are collectively spending hundreds of billions of dollars building AI data centres, with billions more flowing annually into operating them. Nvidia, AMD and Intel are pouring further billions into R&D to wring every last watt of performance from their chip designs. The suggestion that a privately held Australian company founded by three people with no prior data centre experience has cracked the code on halving both construction and operating costs – in an industry where the world’s richest and most technically sophisticated companies are fighting for single-digit percentage improvements – is either the investment opportunity of the decade or a claim that deserves far more scrutiny than it has received.
Some industry insiders describe Firmus’ claims around efficiency and power usage as extreme. The most specific claim Firmus now makes is a power usage effectiveness – or PUE – of 1.1 or below at its new greenfield sites.
PUE, or power usage effectiveness, is the data centre industry’s standard measure of energy efficiency – and for AI factories burning through enormous amounts of electricity, it’s one of the most scrutinised numbers in the business. The metric works as a simple ratio: divide the total energy a facility consumes by the energy that actually reaches the computing equipment doing useful work. A perfect score of 1.0 would mean every watt of electricity goes directly to powering chips – nothing wasted on cooling, lighting or keeping the building running.
A PUE of 1.1 would be impressive but within the realm of the achievable. A figure of 1.03, which Firmus has referenced repeatedly on its website, for its Tasmanian operations, is at the extreme end of what is now possible anywhere in the world, says Dr Amr Omar, a senior research associate at the University of NSW who specialises in data centre energy systems.
For context, the global average PUE sits at about 1.5. Australian data centres typically operate between 1.4 and 1.7. Even the hyperscalers – Google, Meta, Microsoft – with their vast engineering resources and optimised facilities, generally achieve about 1.2.
Omar says there’s a limit to what even the best efficiency gains can do. “Lower PUE improves efficiency, but it does not eliminate peak demand,” he tells this masthead. “A 100-megawatt IT load is still a 100-megawatt IT load. Efficiency reduces overhead … It does not change the fundamental scale.”
This masthead questioned Firmus about its cooling technology, and its power usage claims, to which a company spokesperson responded with a series of dot points.
A spokesperson for Firmus said the 1.03 statistic was a “point measurement” taken from a 2021 “immersion deployment” in Tasmania. Firmus’ figures could not be independently verified by a third party, the spokesperson said, because no recognised certification body for PUE exists.
Sam Maher, chief executive of technology and policy advisory GovTech Australia, said she had concerns about potential Firmus investors being swayed by the narrative around the company, as opposed to what it would actually deliver.
“Something I’ve found interesting, this company has attracted a lot of attention through effective profile-building, and from what I can tell, some investors are being swayed by the narrative rather than the actual infrastructure case. That concerns me.
“The Google emissions findings should have sharpened this conversation,” she said, referring to a report that Google’s carbon emissions went up by 65 per cent between 2019 and 2024, not 51 per cent as the tech giant had claimed. “Instead, they’ve been drowned out by investment hype. Australia is not an exception to that pattern; we might actually be behind it.”
But Greg Boorer, chief executive of CDC Data Centres, Firmus’ infrastructure partner on a new multibillion-dollar Melbourne campus in Brooklyn, said that the company’s efficiency claims are grounded in reality.
“A lot of the energy and the water consumption claims are actually accurate, and a lot of that is built on the foundation which we provide,” he said. In practice, that means CDC builds the data centre shell – the building, the power systems, the base-level infrastructure – and Firmus installs its proprietary cooling technology inside as what Boorer described as “the last mile”. When Firmus quotes a low PUE figure or minimal water usage, part of that performance is attributable to the facility CDC has built around it.
He also pointed to the quality of Firmus’ backers as a signal. “Globally recognised organisations in terms of Nvidia, Blackstone and people like that don’t bet on something without a bit of substance behind it.”
‘Globally recognised organisations in terms of Nvidia, Blackstone and people like that don’t bet on something without a bit of substance behind it.’
Greg Boorer, chief executive of CDC Data Centres
Firmus’ local rivals offer another insight. Sharon AI, led by co-founder James Manning, listed on the Nasdaq in February at $US731 million. Days later it launched what it calls Australia’s first Cisco Secure AI Factory – 1024 Blackwell Ultra GPUs deployed inside NextDC’s facilities, with Cisco providing the security fabric and Nvidia’s top installation partner WWT doing the build-out.
Speaking to this masthead, Manning wouldn’t be drawn on Firmus’ claim that it builds at half the cost of competitors.
“I can’t talk to Firmus and their story. I’m not under the hood,” he said. “Every business has their own path to market. We’re not thinking about what other people in the industry are doing.”
What Sharon AI already has that Firmus doesn’t is enterprise and AI-native customers on its clusters and multi-year contracts underpinning its infrastructure spend. Manning described the broader dynamic as “a customer-led infrastructure boom”.
Then there’s Southern Cross AI, which uses specialist chips that don’t need liquid cooling at all. Founder David Keane claims his systems produce 10 times more AI output per unit of energy than traditional GPU setups, a figure he says is drawn from public benchmarks and open to independent verification.
It points to a deeper problem running through the entire sector: there is no single, agreed-upon way to measure and compare data centre efficiency. PUE measures facility overhead but says nothing about what the chips actually produce. Tokens per second per kilowatt – Keane’s preferred metric – captures output efficiency but only for specific workloads. The result is a marketplace where bold claims abound and direct comparisons are almost impossible.
“We’ll leave others to defend their own numbers,” Keane says of Firmus. “But the risk isn’t that customers never arrive. It’s that operators build the wrong infrastructure for the wrong workload and find themselves holding stranded assets.”
Firmus responded to Keane’s comment by saying its infrastructure was designed to evolve with “silicon roadmaps”.
“Liquid cooling is an infrastructure choice for density and efficiency, not a commitment to one chip type,” Firmus’ spokesperson said.
So far, that scepticism about Firmus’ technology hasn’t dimmed the company’s IPO prospects.
But Curtis’ criminal past might. Under ASX guidelines, a proposed chief executive or director must be of “good fame and character”, which includes consideration of criminal history. Failure to satisfy that test allows the ASX to refuse a company to list.
The ASX declined to comment on Curtis’ specific case, but coincidentally released a more detailed guidance note on the character test just last week. The exchange considers the past conduct at the date of a company’s prospective listing. The new guidance note also points to the time since the offending conduct as being relevant, including whether about 10 years have elapsed. In June, it’ll be a decade since Curtis was jailed.
Swinburne University corporate governance specialist Helen Bird said that while investors might be able to overlook Curtis’ past crime amid the AI-driven pre-IPO rush, they may feel less forgiving if Firmus’ fortunes sour in future.
“I was surprised that he was taking such a front-of-house role given his background and criminal conviction,” she said.
Firmus declined to comment on this masthead’s questions about Curtis’ 2016 insider trading conviction, or whether he satisfied the “good fame and character” test.
But Curtis, Firmus’ co-CEO, remains front and centre of the operation, regularly addressing investor conferences, speaking about the company with missionary zeal.
At the Tech Capital APAC Finance Forum last November, Curtis was asked what he hoped the headline would be in 10 years’ time. He wanted Firmus to be known as “the tortoise in the tortoise and the hare”, the company that took the long game and built the foundation layer correctly from the start.
Right now, with its soaring valuation, and investor confidence buoyed by the AI frenzy, Firmus looks every bit the hare.
The Business Briefing newsletter delivers major stories, exclusive coverage and expert opinion. Sign up to get it every weekday morning.