48 hours on, the $30.6 billion IPO has been called off.

The wave of IPO withdrawals is spilling over into the AI sector.

According to foreign media reports recently, star AI data center firm Firmus has officially announced the withdrawal of its IPO filing. Calculated based on the 2026 global IPO metrics, this IPO, which was set to raise up to 5.5 billion USD, was originally expected to become the fourth-largest IPO in the world this year, second only to SpaceX, CXMT Corp and Cerebras Systems.

However, following Firmus's failed IPO push, its planned IPO valuation of 30.6 billion USD (equivalent to about 204.6 billion RMB) has also collapsed. It is worth noting that Firmus spent less than 48 hours from officially launching investor subscription inquiries to withdrawing its IPO filing. The rapid "subscription abandonment" by investors reversed the once-optimistic expectations almost instantly, sounding a wake-up call for the current frenzy in the AI track.

Firmus Has Withdrawn Its IPO Filing

The story of Firmus's IPO withdrawal starts with the rapid "inflation" of its valuation.

According to disclosures, Firmus originally planned to promote its offering at a price of A$11 per share, corresponding to an IPO valuation of 30.6 billion USD. However, two days before the close of subscription inquiries, investors began to withdraw their orders on a large scale, as Firmus's valuation climbed too fast in the short term, raising widespread concerns among investors.

Retrospective checks show that in 2026, Firmus's valuation started at over 4 billion USD in February, and reached about 5.5 billion USD after receiving investment from Coatue Management and NVIDIA in April this year; later in August, it completed a new round of 2 billion USD financing, pushing its valuation above 10.5 billion USD.

Its planned IPO at a 30.6 billion USD valuation means that on the basis of its previous valuation that had already tripled rapidly, the valuation nearly tripled again within just three months.

In the end, this 30.6 billion USD valuation failed to convince investors. According to John Pearce, Chief Investment Officer of large Australian pension fund UniSuper, Firmus's valuation is "too speculative", and despite the company's attractive business story, it still lacks sufficient persuasiveness.

John Pearce believes that the company requires a large number of conditions to be met simultaneously to support this valuation, and it may continue to rely on debt and equity financing in the future, so the fund chose not to participate in this IPO.

Following the refusal of this large pension fund to subscribe, other investors also began to accelerate their order withdrawals. During this period, the underwriting team considered lowering the offering price to A$8.25, corresponding to an adjusted IPO valuation range of 20 billion to 25 billion USD, but still failed to attract enough investors to subscribe. In addition, Firmus's board of directors itself also determined that continuing to push forward with the offering would not be in the best interests of the company and its shareholders.

As a result, the subscription inquiry process was paused early, and the market left almost no buffer time for Firmus. In just 48 hours, this company, once regarded as a "rising star in the AI data center boom", completely became one of the most tragic IPO failure cases in recent years.

Why Firmus?

Firmus's IPO withdrawal has undoubtedly poured a basin of cold water on the currently red-hot AI financing market. But why did this happen to Firmus?

The answer clearly lies in Firmus itself. The prospectus shows that in fiscal year 2026, Firmus only posted revenue of 51 million USD, with a net loss of 68.2 million USD. Although the company plans to achieve future revenue growth through large-scale data center construction, the scale of its currently operational data centers is limited, with only 46 megawatts completed, which is far from its planned 1.6 gigawatts target.

More importantly, Firmus relies entirely on the forecast of explosive revenue growth in the next three years to support its valuation, rather than real performance data. What is more exaggerated is that some investment institutions have calculated that based on its 51 million USD revenue and 30.6 billion USD IPO valuation, Firmus's price-to-sales ratio (PS) exceeds 600 times.

Even in the current frenzy of AI infrastructure, this price-to-sales ratio is enough to make institutional investors stop and calculate carefully.

The termination of the "Project Southgate" that Firmus cooperated on with CDC Data Centres is also one of the triggers for this IPO withdrawal. Disclosures show that Project Southgate was originally planned to jointly develop AI data centers with a total capacity of 1.6 gigawatts, involving an investment scale of about A$73 billion, but the two sides reached a consensus to stop the project earlier in 2026 due to differences in business direction during the cooperation process.

In addition, Firmus's customers are reportedly highly concentrated in a small number of tech giants including Meta and OpenAI, which also increases the uncertainty risk for Firmus amid the sudden cooling of market sentiment.

Joseph Koh, Investment Manager at Blackwattle Investment Partners, said, "Firmus is charging an extremely high price for a future that can only be achieved with near-perfect execution."

How Long Will the AI Infrastructure Frenzy Last?

In fact, Firmus is not the first company to suffer a failed IPO push.

In the AI infrastructure sector, the IPO of another data center giant also encountered a premature end. This is the second incident in a short period where AI infrastructure assets were blocked by pricing pressure in the public market. Before that, SB Energy under SoftBank suspended its IPO roadshow in late September this year, due to additional inquiries from the U.S. Securities and Exchange Commission, as well as investor doubts about its roughly 500 billion USD valuation and its high reliance on the single customer OpenAI.

Earlier, the post-IPO market performance of AI infrastructure companies was also not optimistic. Data shows that in May this year, BXDC, the AI data center REIT under Blackstone, completed its IPO and raised 1.75 billion USD, but this AI data center stock broke its offering price immediately after listing, closing at 19.81 USD on the day, below its 20 USD issue price.

As the world's largest alternative asset management firm, Blackstone's BXDC IPO was originally regarded by the market as an important test for the assetization and pricing of AI data center assets. However, its post-listing performance means that even the most mature players in the market cannot find enough underwriters for this seemingly "stable" asset in the public market to arouse market interest.

Furthermore, Csquare, a Texas-based data center operator that landed on the New York Stock Exchange in July this year, set its IPO price at 21 USD per share, which was not only lower than the 23 to 27 USD range it previously promoted to investors, corresponding to a market capitalization of 3.2 billion USD. On its first trading day, Csquare also failed to hold its offering price, and finally closed at 20.75 USD.

It is worth noting that Csquare raised more than 1 billion USD in its IPO, the vast majority of which was used to repay debt — this in itself shows that this type of AI infrastructure company has accumulated a large amount of debt and considerable leverage before listing, and the primary task of the IPO is not to raise funds for growth, but to reduce liabilities.

From post-IPO price breaks, to listing with reduced fundraising scales, to direct IPO withdrawals... this series of market coldness reflects the loosening of the capital market's belief in the narrative of heavy computing power assets, as well as growing caution on valuations.

Earlier, the Holt Unit under UBS released a report in August 2026 showing that the volatility of the global technology industry has climbed to the highest level since the burst of the dot-com bubble. "Investors are reassessing the sustainability of cash flow returns from hyperscale data centers and semiconductor companies," the report noted.

The UBS team estimates that the top five hyperscale data center operators — Microsoft, Meta, Alphabet, Amazon and Oracle — will face a total capital demand of about 227 billion USD in 2027. If they cannot meet both operational and financing commitments at the same time, these five giants will probably not be able to bear the huge capital expenditure and the tension with free cash flow, let alone those new "players" that survive relying on the giants and cannot generate stable revenue.

This article is from the WeChat official account "ChinaVenture", written by Chen Mei, and published with authorization from 36Kr.