Firmus, an Australian builder of AI data centers backed by Nvidia, on Oct. 9 withdrew a stock market listing that had valued it at about U.S.$30.6 billion, nearly three times the value private investors put on it at the start of August. Its board decided that going ahead with an IPO was no longer in shareholders’ best interests, and the company said it would raise money privately instead. One fund manager who studied the offer and passed told Reuters: “I think the market wasn’t comfortable taking that leap of faith quite at this stage yet.”
Chief Investment Officer Thierry HasseWe believe the AI build-out is a genuine industrial revolution. Revolutions are rarely smooth, though, and last week showed where this one is most likely to hit air pockets, the sudden drops that catch investors off guard. Last week’s trouble was not about whether the world needs more computing power. It was about how the data centers that supply that power are being paid for. Companies that fund their building out of what they already earn can ride out a bad week. Companies that depend on lenders and new investors, or whose finances nobody outside can see, are the ones most exposed.
A Leap of Faith
Look at what that price was buying. Firmus has two data centers running and plans for five more, and Reuters reported that investors were wary of how much it had already borrowed to get this far. Critics also pointed to its losses. Investors were being asked to pay for the finished factory while the first floor was still going up.
This company’s cash flows, for now, are mostly a forecast. A forecast is worth only what someone will pay for it today, and last week fewer buyers were willing.
Who Pays for the Build-Out
Not every company building AI is paying for it the same way, and the difference matters more now than it did when money was cheap. The 10-year Treasury yield is roughly what investors demand to lend to the U.S. government for a decade, and it sets the starting point for what companies pay to borrow. According to Bloomberg, it recently reached its highest level since 2002. Every dollar borrowed to build a data center now costs more than it did a few years ago.
At one end of the spectrum are the largest cloud and internet platforms. As we noted in our quarterly update, they have been borrowing more to build. But by our count from their own filings, the four biggest together spent about 77 cents on buildings and equipment over the past year for every dollar their operations brought in, so most of the build is still paid for from earnings. In the middle are large companies spending well past what their businesses bring in. Oracle’s filings show that over the past four quarters it spent about $1.60 for every dollar its operations brought in, and covered the gap with borrowing and new shares.
At the far end are private companies, from data center developers to the AI labs themselves, whose finances outside investors cannot see. A company that pays for its growth out of what it already earns, even if it borrows alongside, does not need a receptive IPO window, a patient lender, or a rising share price to keep building. Firmus needed all three.
The Numbers Nobody Can Check
Private companies also write their own headlines, and last week one of them was rewritten. On Oct. 8, the Financial Times reported that OpenAI had told investors its annualized revenue for September was close to $50 billion, about $20 billion below a figure the company had signaled earlier. Nothing about its sales had changed. Anthropic counts sales made through its cloud partners as its own revenue; OpenAI does not. According to Reuters, the higher figure came mainly from an attempt to count OpenAI’s sales the way Anthropic counts its own.
Annualized revenue is itself a projection, often one month’s sales multiplied by 12. When a choice of definition can move a figure by $20 billion, that figure deserves less weight than the market gave it. A public company reports audited results every quarter under rules anyone can read. A private company shares what it chooses, defined the way it chooses.
What We Own Instead
This is why we approach AI through businesses whose books we can read: the chipmakers, the power producers and grid suppliers the data centers depend on, the large cloud platforms, and, potentially, smaller companies that use AI to widen their operating margins, though it is early to know which. In our assessment, owning the build-out through companies that pay for it from earnings lets a portfolio take part in the revolution without betting on which private lab or debt-funded developer wins the arithmetic.
When the largest spenders next report their results, the figure we will be reading is how much of that spending each business pays for itself. It is reasonable to find a week unsettling when a $30 billion listing disappears and a $20 billion number turns out to be a definition. The demand behind this revolution did not change last week. What changed was the market’s patience for paying in advance, and owning businesses that already pay their own way is how we have chosen to take part.
Sources
Reuters, “Nvidia-backed Firmus scraps $5 billion Australian IPO on poor demand,” October 9, 2026.
Blockspace, “Firmus withdraws A$7.1 billion ASX IPO, turns to private capital,” October 2026.
The Nightly, “AI hopeful Firmus struggles to sell investors on $44 billion valuation,” October 2026.
Bloomberg, “US Benchmark 10-Year Yield Briefly Hits Highest Since 2002,” October 1, 2026.
Reuters, “OpenAI’s annualized revenue $20 billion less than previously signaled, FT reports,” October 8, 2026.
FactSet Insight, July 23, 2026.
Elevage Partners, “Quarterly market update: The bill is due,” October 7, 2026.
Company quarterly and annual filings; spending ratios calculated by Elevage Partners Research.
Important Disclosure(s)
Holdings Disclosure: Elevage Partners and/or its clients may hold positions in securities referenced in this commentary. The information contained herein represents the views of Elevage Partners at a specific point in time and is based on information believed to be reliable. No representation or warranty is made concerning the accuracy of any data compiled herein In addition, there can be no guarantee that any projection, forecast, or opinion in these materials will be realized. Any statement non-factual in nature constitutes only current opinion which is subject to change. These materials are provided for informational purposes only and do not constitute investment advice. Any reference to a security listed herein does not constitute a recommendation to buy, sell, or hold such security. Past performance is no guarantee of future results. The historical returns of any securities and/or sectors mentioned in this commentary are not necessarily indicative of their future performance.
Air pockets in the revolution
Home / Air pockets in the revolution
Thierry Hasse, Chief Investment Officer
Elevage Partners | October 12, 2026
Firmus, an Australian builder of AI data centers backed by Nvidia, on Oct. 9 withdrew a stock market listing that had valued it at about U.S.$30.6 billion, nearly three times the value private investors put on it at the start of August. Its board decided that going ahead with an IPO was no longer in shareholders’ best interests, and the company said it would raise money privately instead. One fund manager who studied the offer and passed told Reuters: “I think the market wasn’t comfortable taking that leap of faith quite at this stage yet.”
A Leap of Faith
Look at what that price was buying. Firmus has two data centers running and plans for five more, and Reuters reported that investors were wary of how much it had already borrowed to get this far. Critics also pointed to its losses. Investors were being asked to pay for the finished factory while the first floor was still going up.
This company’s cash flows, for now, are mostly a forecast. A forecast is worth only what someone will pay for it today, and last week fewer buyers were willing.
Who Pays for the Build-Out
Not every company building AI is paying for it the same way, and the difference matters more now than it did when money was cheap. The 10-year Treasury yield is roughly what investors demand to lend to the U.S. government for a decade, and it sets the starting point for what companies pay to borrow. According to Bloomberg, it recently reached its highest level since 2002. Every dollar borrowed to build a data center now costs more than it did a few years ago.
At one end of the spectrum are the largest cloud and internet platforms. As we noted in our quarterly update, they have been borrowing more to build. But by our count from their own filings, the four biggest together spent about 77 cents on buildings and equipment over the past year for every dollar their operations brought in, so most of the build is still paid for from earnings. In the middle are large companies spending well past what their businesses bring in. Oracle’s filings show that over the past four quarters it spent about $1.60 for every dollar its operations brought in, and covered the gap with borrowing and new shares.
At the far end are private companies, from data center developers to the AI labs themselves, whose finances outside investors cannot see. A company that pays for its growth out of what it already earns, even if it borrows alongside, does not need a receptive IPO window, a patient lender, or a rising share price to keep building. Firmus needed all three.
The Numbers Nobody Can Check
Private companies also write their own headlines, and last week one of them was rewritten. On Oct. 8, the Financial Times reported that OpenAI had told investors its annualized revenue for September was close to $50 billion, about $20 billion below a figure the company had signaled earlier. Nothing about its sales had changed. Anthropic counts sales made through its cloud partners as its own revenue; OpenAI does not. According to Reuters, the higher figure came mainly from an attempt to count OpenAI’s sales the way Anthropic counts its own.
Annualized revenue is itself a projection, often one month’s sales multiplied by 12. When a choice of definition can move a figure by $20 billion, that figure deserves less weight than the market gave it. A public company reports audited results every quarter under rules anyone can read. A private company shares what it chooses, defined the way it chooses.
What We Own Instead
This is why we approach AI through businesses whose books we can read: the chipmakers, the power producers and grid suppliers the data centers depend on, the large cloud platforms, and, potentially, smaller companies that use AI to widen their operating margins, though it is early to know which. In our assessment, owning the build-out through companies that pay for it from earnings lets a portfolio take part in the revolution without betting on which private lab or debt-funded developer wins the arithmetic.
When the largest spenders next report their results, the figure we will be reading is how much of that spending each business pays for itself. It is reasonable to find a week unsettling when a $30 billion listing disappears and a $20 billion number turns out to be a definition. The demand behind this revolution did not change last week. What changed was the market’s patience for paying in advance, and owning businesses that already pay their own way is how we have chosen to take part.
Sources
Reuters, “Nvidia-backed Firmus scraps $5 billion Australian IPO on poor demand,” October 9, 2026.
Blockspace, “Firmus withdraws A$7.1 billion ASX IPO, turns to private capital,” October 2026.
The Nightly, “AI hopeful Firmus struggles to sell investors on $44 billion valuation,” October 2026.
Bloomberg, “US Benchmark 10-Year Yield Briefly Hits Highest Since 2002,” October 1, 2026.
Reuters, “OpenAI’s annualized revenue $20 billion less than previously signaled, FT reports,” October 8, 2026.
FactSet Insight, July 23, 2026.
Elevage Partners, “Quarterly market update: The bill is due,” October 7, 2026.
Company quarterly and annual filings; spending ratios calculated by Elevage Partners Research.