AI Datacentres – Greed and Fear

Firmus management shoots itself in the foot.

• Australian data centre operator, Firmus, has cancelled its IPO after it was unable to win enough investor interest at a valuation of $30bn, which I think was caused by existing shareholders getting too greedy rather than a business that has no prospects.
• Firmus is an Australian data centre operator that currently has 42MW of Nvidia Blackwell capacity with plans to increase that to 1GW by the end of 2028 with new installations in Tasmania, Indonesia and Malaysia.
• 1GW of capacity costs around $50bn – $60bn to build, meaning that Firmus needs to raise far more than the $2bn that it raised in August 2026, which is what the IPO was all about.
• Looking at the news flow, it is clear that the management and shareholders of Firmus have made serious strategic errors in the handling of this transaction, although its short- to medium-term business prospects look very healthy.
  • First, Greed: where shareholders and management thought they could capitalise on the insatiable demand for AI and grossly overpriced the IPO.
  • Stakeholders are asking new investors to pay a valuation of around $30bn, which is three times the valuation at which it raised money just 2 months ago.
  • There is no doubt in my mind that the intrinsic value of the company has not increased that much during that time, and so this has been seen by investors as an unreasonable ask.
  • The company currently has 42MW of capacity, which is mostly leased to Meta and was announced in March 2026.
  • This leads me to conclude that the price was probably somewhere around $25bn per GW of capacity, leading to around $1bn of contracted revenues on an annual basis.
  • This should be extremely profitable revenue, as RFM has calculated that at $25bn / GW, 5-year returns on Blackwell are 39% per year.
  • However, everyone else should be producing similar economics, but for some reason shareholders and management decided to price the IPO at 30x current revenues compared to more established and less risky propositions such as Nebius (17.6x 2026 revenues) and CoreWeave (3.5x 2026 revenues).
  • Second, hyperbolic & nonsensical commentary: where the company has stated that it expects to be able to earn 90% EBITDA margins, which in my opinion is deeply misleading.
  • This is because 81% of the cost of running a Blackwell data centre is depreciation, which EBITDA specifically excludes.
  • This is like a software company stating its profitability based on the electricity used to run its computers and ignoring the cost of the developers who write the software.
  • The company has really shot itself in the foot here because its real profitability will look excellent at $25bn per GW (65% operating margin), and now it has blown its credibility.
  • This reflects very poorly on the current management of Firmus and will raise doubts about its capabilities in investors’ minds.
• The company has passed 900MW of contracted capacity, meaning that revenues of around $36bn per year, at excellent margins, but first it has to build the infrastructure, and for this it will need to continue to raise money aggressively.
• Firmus will now look to raise the money privately, which I am pretty sure that it will be able to do given how much demand currently exists.
• The risk here is that at some point, pricing for data centre compute is going to normalise, as the returns currently available are so attractive that everyone who has capacity is making it available for sale while everyone else builds as fast as they can.
• When that happens, there will be a substantial impact on companies like Firmus, as their long-term forecasts will almost certainly be way too high.
• Hence, I think Firmus remains an attractive proposition, but given its early stage, the risk attached to it and how management has damaged its own reputation, $30bn was way too much to ask.
• The market got this one right.