Another One Bites The Dust: Nvidia-Backed Firmus Slashes IPO Price 25% As AI Data Center Window Slams Shut

Two weeks ago, when SoftBank’s SB Energy pulled the marketing of its giant IPO meant to find its even more giant data center, we wrote that “slowly the data center dream is turning into a nightmare” (“SB Energy Delays IPO Funding World’s Largest Data Center Amid Investor Revolt, Public Outcry”, Sep 22). Two days later we went a little further:

SB Energy is next and then it’s all over

— zerohedge (@zerohedge) September 24, 2026

“All over” may be premature. “Next,” however, has arrived right on schedule.

Overnight, Bloomberg reported that Australian data center operator Firmus closed the books on its struggling IPO “without clear indication of the price or the deal structure,” as investors grew “increasingly concerned that the deal could be pulled.”

The Nvidia-backed company had been trying to raise as much as $5.5 billion at A$11 a share, which valued it at A$43.7 billion ($30.4 billion) and would have made it one of the largest listings in Australian history. By the time the books closed, the A$11 was gone: per the Australian Financial Review via Reuters Breakingviews, Firmus cut the price to as low as A$8, a 25% haircut, and was “scrambling” to hold it at A$8.25.

How’s the SB Energy IPO going?

— zerohedge (@zerohedge) October 8, 2026

Put differently, the second mega data center IPO in a month has just hit the same wall as the first. This time the wall came with a twist, and that twist is why this story belongs in our long-running coverage of the debt holding up the AI supercycle.

Below we explain how Firmus got from a $5.5 billion private valuation to a $30 billion IPO in six months, why the “creative” valuation metric failed to sell it, and why at the reduced price the equity is worth less than the debt.

Books Closed, Price… TBD

The cracks were visible before the books closed. On Wednesday evening Goldman’s Sydney futures desk wrote in its morning comment that “the AFR reports bankers managing the IPO of Firmus are considering a potential cut in the listing price because of weak demand from foreign investors.” Within hours that “consideration” was a 25% cut.

Equity investors with direct exposure did not wait around. Shares of Firmus backer Maas Group fell as much as 30% in Sydney, the most on record, which prompted the company to note “significant market speculation and commentary” about whether the IPO would go ahead. Meanwhile Plato, an Australian fund manager running about $6 billion, called Firmus a “screaming short“, citing 30 “red flags” ranging from valuation to senior management.

Ten Cap’s Jun Bei Liu summed up the bookbuild on Bloomberg TV:

“I’ve never seen an IPO so polarizing. There was a lot of international investor interest, however, when it comes to the crunch, the demand seems like it isn’t there when they were asked to put up the capital that’s required.”

Translation: everyone loves AI infrastructure until they are asked to pay for it.

For the record, the bookbuild was pulled forward from Friday to Thursday just three days ago because of “early indications of demand in excess of the offer size.” That’s the same “oversubscribed” (just as the Paramount bond deal had over $100 billion in “demand” and its bonds are now crashing every day) deal now negotiating with itself over A$0.25 a share.

From $5.5 Billion To $30 Billion In Six Months

The rerating is the whole story. Firmus started out as a Bitcoin mining operation in 2019. In February it landed a $10 billion debt package led by Blackstone and Coatue. In April a Coatue-led round valued it at $5.5 billion. In August a $2 billion round with Nvidia, Coatue, Blackstone and Jane Street nearly doubled that to $10.5 billion. Then, less than two months later, the IPO asked public investors for three times the August price.

Behind that price tag, according to Breakingviews, Firmus has so far built just 42 megawatts of capacity out of a pipeline of about 1 gigawatt, or just over 4%. It recently shifted away from a mostly domestic partnership with CDC Data Centres toward Malaysia and Indonesia, to serve hyperscalers such as OpenAI and Meta. Per Reuters, the draft prospectus forecast a $77 million loss in the first half of FY27 and no forecasts at all beyond June 2027.

Australian fund managers were less than impressed. Morningstar’s Lochlan Halloway saw signs of a “boom phase,” pointing to the “wild increase in valuation in such a short period of time.” Merlon’s Kirit Hara said his process “prevents us from effectively buying into, kind of, the hopes and dreams.” Katana’s Romano Sala Tenna admitted “we are struggling with the fundamental arithmetic,” and Blackwattle’s Joseph Koh, who will not bid, put it more simply: “There are so many unknowns.”

Introducing “EV+1/EBIT+2”

When the arithmetic doesn’t work, you change the arithmetic. As Breakingviews’ Antony Currie explains, the deal leaned on a new valuation tool, “EV+1/EBIT+2.” It takes enterprise value using the net debt the company will have in 12 months and divides it by the EBIT it hopes to earn two years from now. 

The logic is that data center developers fund up to 90% of construction with debt and can start generating revenue within a year, so today’s numbers flatter nobody. The problem, as Currie notes, is that the metric is “easy to manipulate.” Capex assumptions can be inflated to pump up EV, and any delay in permits, equipment, power or water means “EV+1 becomes +2 or +3, with EBIT delayed to +3, +4 or beyond.” 

This is pure batshit insanity, and shows just how big the AI bubble has become for a company to realistically think that such “vibes” garbage can be credibly used an investment highlight for institutional investors. 

Regular readers know how that goes: Project Jupiter declared force majeure and Oracle’s 1.3GW Wisconsin “Lighthouse” campus slipped earlier this week, so “+1” is the optimistic case. Then again, nothing says “mature asset class” like a valuation multiple whose denominator is two years in the future.

Half The Deal To… The Existing Shareholders

A second detail also spooked buyers. Firmus planned to allocate about half of the IPO to “selected existing strategic and financial investors”, a list that includes Nvidia, Coatue, Blackstone and Jane Street. At the same time, pre-IPO holders could reportedly sell up to 20% of their stakes from day one. Bloomberg says this overhang was what turned potential investors cautious, on top of “what some investors see as an aggressive pricing strategy.”

Then there is where the money goes. IPO proceeds would fund purchases of Nvidia GPUs for Firmus’ first project in Batam, Indonesia, built with DayOne as part of an eight-year partnership with Nvidia. In other words, Nvidia invests in Firmus, Firmus raises money partly from its own backers, and the money goes back to Nvidia for chips. We have seen this kind of Nvidia circular financing before, just never quite this small.

One more detail: the four joint lead managers are Bank of America, JPMorgan, Morgan Stanley and Morgans. That would explain the absence of sell-side research on the deal from three of the largest research shops in the world.

When The Equity Is Worth Less Than The Debt

This is where the deal stops being an IPO story and becomes a credit story. As one reader put it in a reply to our tweet, “At that point an AI data centre stops being a growth stock and starts being a credit.”

The napkin math is simple, and alarming. Morningstar estimates Firmus’ debt at about $30 billion, roughly six times its own forecast earnings. At A$11 the equity was worth about $30.4 billion, so the company was valued at about one dollar of equity for every dollar of debt. At A$8.25, and assuming the same share count, it is worth roughly A$33 billion, or about $23 billion, some $7 billion less than the debt it carries. At A$8 it is $22 billion.

In other words, at the price the bankers are now “scrambling” to defend, the shareholders’ stake would be worth less than what Firmus owes its lenders, before the stock trades for a single day. That is a leveraged credit with an equity ticker, and it lands at the worst possible moment for AI credit, which we summarized yesterday afternoon:

Credit remains the big story: here everything is falling apart. 1. Oracle bond yields, CDS record high 2. SpaceX bonds new all time low/YTW high 3. Bond behind Meta’s Hyperion data center (proj Beignet) hit record low 4. Paramount CDS trading at financial crisis levels https://t.co/bxEqe8IQFw

— zerohedge (@zerohedge) October 7, 2026

Throw in 10Y Treasury yields that hit 5.36% on Wednesday, a 24-year high, per the same Goldman Sydney note, and the cost of carrying $30 billion of data center debt only goes one way.

SB Energy, Then Firmus… Then Who?

Firmus is the second data center mega-IPO to stumble in three weeks. SoftBank-backed SB Energy postponed marketing its IPO in late September after more SEC questions and investor pushback on a valuation of around $60 billion, and on its dependence on a single customer, OpenAI. Nvidia had agreed to guarantee up to $105 billion to help OpenAI lease SB Energy’s Ohio data center, while also investing $1.5 billion in SB Energy itself. Back then we asked the obvious question:

Is it a coincidence that the 2 key data centers of OpenAI’s Stargate are suddenly on the ropes? Project Jupiter (4.5GW) loans trading < 90, major delays… SB Energy (8GW) delaying IPO, major delays. Without these two DCs having clear timeline/funding, OpenAI can just cancel IPO

— zerohedge (@zerohedge) September 23, 2026

And the line behind them is long. Singapore’s DayOne, which also happens to be Firmus’ partner in Batam, wants to raise up to $5 billion at a $20 billion valuation in a US listing as soon as November. Switch has filed confidentially, Vantage and CyrusOne are exploring listings, and London’s Nscale just raised $3.36 billion in pre-IPO convertibles. As Shenton Research’s Ke Yan told Reuters, “the dividing line is whether demand is contracted and already energised, or only planned.” Firmus, with 42MW built against a gigawatt pipeline, is firmly on the “planned” side of that line.

Goldman: $1 Trillion Of AI Spending… And Everyone Hiding In The Mag 7

None of this means the AI spending boom is slowing. Goldman’s Global Corporate Access team flagged in Wednesday’s IR Kick-Start (available to pro subs) that GS Research now estimates global AI investment will exceed $1 trillion this year, with the US and Asia accounting for more than 80% of net data center additions (Europe gets about 10%).

The question was never whether the money gets spent. It’s who funds it, and at what price. Goldman’s Asia sales desk answered part of that overnight in “Power Back On” (also available to pro subs): hedge funds have “certainly re-engaged in tech as a result of macro uncertainty,” with net exposure to the Mag 7 as a share of total US exposure on GS’ prime book at ~22%, the highest since the start of 2022.

Put differently, investors still want AI, but they want it through companies with balance sheets that can fund a trillion dollars of capex, not through a 42-megawatt former Bitcoin miner levered six times its forecast earnings. The same GS desk notes that “the positioning within AI and large-cap tech no longer seem to be tailwinds,” which means the crowd is all on one side of the boat, and it’s not the side Firmus is on.

Bottom Line

This week, Mike Novogratz told the Greenwich Economic Forum that AI is the “biggest bubble of our lifetime,” but that it can’t burst yet because it “isn’t spectacular enough” (which we discussed earlier). Ray Dalio disagreed and speaking at a CEO conference in Singapore, called it a “classic bubble” nearing its bursting point because of rising rates and the debt behind AI infrastructure.

We side with Dalio, and the reason is the IPO window. Bubbles don’t end when the bulls run out of narrative; they end when the marginal buyer refuses to take paper off the hands of the early investors at the asking price. That has now happened twice in three weeks, first with SB Energy and now with Firmus. Each time the issue was the same: too much debt, too little contracted revenue, and existing backers looking for an exit.

Having warned since October 2025 that AI is a debt bubble too, something that virtually everyone – even the WSJ – now accepts, we’ll allow ourselves another prediction: when the equity cushion under a data center is smaller than the debt, the equity investors stop pricing a growth story, and it becomes the bondholders’ problem. Just ask Oracle’s lenders.

Firmus is scheduled to lodge its prospectus on October 12 and start trading on October 23, assuming there is still a deal. We’ll see whether the “+1” in EV+1 refers to years, or to the number of IPOs that follow it into the wall.

Much more in the full Goldman “Power Back On” and “IR Kick-Start” notes, both available to pro subs.