AI data-center IPOs hit a valuation test

Firmus's withdrawal on October 9 is the clearest example here of an IPO that did not proceed. Data Center Dynamics reported that the listing had been scheduled for October 23. In the company statement carried by that report, the board said proceeding would not serve shareholders' interests and outlined a search for private-market capital and alternative routes.(Data Center Dynamics)

Reuters Breakingviews described a planned A$11 share price and a market value approaching A$44 billion, alongside investor concerns about the growth plan. That valuation was a proposed transaction price, not an amount already invested by the public. Likewise, the intended capital raise and the value of the whole company are separate figures.(Reuters Breakingviews)

My interpretation is that the failed offering exposes a financing dependency. If expansion requires fresh equity, the availability and price of that equity are part of the business case. A company can have attractive customers and still face a gap between the projects it wants to build and the money investors will supply on acceptable terms. Withdrawal does not, by itself, establish insolvency or a permanent end to the business.

On October 13, Reuters reported that Blackstone-backed AirTrunk had postponed a proposed Singapore REIT IPO amid difficult market conditions. The report cited two people with direct knowledge and described a potential US$1.5 billion raise. AirTrunk had not immediately responded to Reuters and Blackstone declined to comment. This is a sourced report of a postponement, rather than a confirmed permanent cancellation.(Reuters via MarketScreener)

The REIT structure matters. A property vehicle is assessed through the cash flows and obligations of the assets placed inside it; it should not automatically be valued like a company renting out GPU computing services. The proposed proceeds also should not be mistaken for a valuation of the entire AirTrunk group.

For a listing built around income-producing infrastructure, financing costs and the yield buyers require can affect the price they will pay. A delay may therefore reflect a disagreement over price, timing or structure. The reported delay alone cannot tell us that tenants have stopped needing data-center capacity.

Elice did reach the public market. Seoul Economic Daily reported a ₩90,500 offer price, at the top of the indicated range, and a retail subscription rate of roughly 408 times. Its October 20 report recorded a ₩63,000 trading price at 9:30 a.m. local time. The same report described the company's expansion from an education platform into GPU cloud services and data centers.(Seoul Economic Daily)

Calling this a cancelled or uncompleted IPO would be wrong. The issuer raised capital; the immediate setback was the price available to shareholders in secondary trading. A large subscription multiple measures bids relative to the shares allocated in that process. It is not a promise that buyers will support the offer price once trading opens.

The distinction is practical. A weak debut can make later fundraising and shareholder exits more difficult, but it does not automatically reverse the original raise. The opening-day decline also cannot, on its own, quantify the company's future utilisation, profitability or financing needs.

These businesses sit near the same AI buildout but do not sell identical products or carry identical risks. Their common challenge is paying for capacity before the eventual cash generation is fully demonstrated. Shareholders must decide both whether the capacity will be useful and whether enough of the resulting value will remain for them after construction costs, operating expenses and financing obligations.

One plausible feedback loop is that a lower equity valuation makes a given capital raise more dilutive. If debt is unavailable or expensive, the company may have to stage construction more slowly or bring in partners. Slower delivery can then change the revenue timetable used to support the original valuation. This is a financing mechanism to examine, not a claim that every step has occurred at all three companies.

That is why an announced customer, planned megawatts and revenue already being collected should be kept separate. A contract can be valuable while remaining conditional on delivery, power availability or other milestones. Similarly, an existing data-center asset and a proposed campus require different assumptions about time, spending and execution.

For the next AI infrastructure prospectus, these are the questions I would use to separate a compelling demand story from a financeable equity proposition.

A completed financing on clearly disclosed terms, evidence that capacity is being delivered, and cash generation catching up with spending would make the funding story easier to assess. Conversely, repeated delays that increase the amount of uncommitted capital needed would deserve closer attention. For Elice, subsequent trading and operating disclosures matter more than treating the first session as a permanent verdict.

My reading of these setbacks is that public markets are testing who earns an adequate return from the AI buildout, at what entry price and with whose capital. Growing demand can support an industry while leaving some offerings unattractive. Keeping the transaction outcome, operating evidence and valuation assumptions separate makes that debate more useful than declaring every difficult debut a failure of AI.