As investors increasingly question whether the billions of dollars poured into data centers can translate into sustainable profits, the artificial intelligence (AI) boom may be entering a more demanding phase.
The brutal share price performance of South Korean AI infrastructure supplier Elice on its trading debut has become the latest warning signal. The Seoul-based neocloud company saw its shares plunge 32% on its first day of trading on Tuesday, despite pricing its initial public offering (IPO) at the top of its offering range and raising 201 billion won (US$151 million). However, the stock rebounded about 30% on Wednesday.
With borrowing costs rising and uncertainty intensifying over the long-term returns of massive AI investments, investor sentiment toward AI infrastructure companies has shifted. Beyond Elice, several AI infrastructure companies have recently suffered setbacks.
The most notable is CoreWeave (NASDAQ: CRWV). The company's shares tumbled 25% last week, marking their worst performance since November 2025. Meanwhile, yields on its junk-rated bonds surged to 15%, typically seen as a critical signal of debt distress, with credit default swap pricing indicating a more than 60% probability of default within five years.
Nscale is also facing investor opposition to its IPO valuation target. People familiar with the matter said the data center developer backed by Nvidia (NASDAQ: NVDA) has been seeking an IPO valuation of about US$30 billion. But some of those people said a group of investors recently urged the company to consider a valuation range of US$16 billion to US$19 billion, believing that range would be more likely to win over buyers.
Additionally, Blackstone's data center operator AirTrunk has decided to postpone its Singapore real estate investment trust (REIT) listing plan to the first or second quarter of 2027. AirTrunk is a hyperscale operating platform spanning Australia, Japan, Singapore, Malaysia, Hong Kong and India. It was acquired by Blackstone and CPP Investments in 2024 for A$24 billion, the largest data center transaction on record globally at the time. Its REIT was positioned as a "steady path" to investing in AI infrastructure — with assets already in stable operation, tenants being large hyperscale cloud computing companies, and generous dividend returns. According to reports, before the offering, AirTrunk had considered raising its yield from an originally planned 5.5% to about 6% to attract investors, but still failed to succeed. The AirTrunk REIT offered an excess return of less than 100 basis points, which investors deemed far from sufficient compensation for a single-industry illiquid instrument carrying currency risk, tenant concentration risk and construction risk.
Earlier still, Nvidia-backed Australian AI data center operator Firmus had originally planned to list at a valuation of about US$30 billion, which would have been one of Australia's largest IPOs. However, due to excessive valuation, insufficient investor subscription demand and market concerns over the massive financing needs of AI infrastructure, the closely watched listing plan was ultimately cancelled on October 9.
These signs highlight a broader shift in overall market sentiment toward AI investment deals, suggesting investors may be scrutinizing ambitious expansion plans and their profit paths more rigorously. Dan Ives, a technology analyst at Yorkville Ives, said: "This is a reality check moment in the tech stock trade. Investors are trying to more clearly distinguish between winners and losers, and to determine whether there are any weak links in the tech AI supply chain." He added: "Investor patience is gradually running out, and scrutiny will intensify further, especially for companies carrying heavy debt loads, such as some neocloud providers."
With borrowing costs constantly rising and growing market concerns about debt-driven expansion, investors may have to distinguish between two types of companies: those that can generate cash flow from the AI boom, and those that spend heavily for anticipated future demand. Billy Leung of Global X ETFs said: "The AI investment trade has moved from a phase where the scale of capital expenditure announcements alone would be viewed positively by the market, to a phase where investors are starting to ask which companies can truly generate sufficient returns from these capital investments." He further noted: "The market is distinguishing between two types of companies — those that need capital to sustain infrastructure buildout, and those that can generate revenue as infrastructure buildout continues." He said the most vulnerable are neocloud providers whose expansion outpaces their own cash generation capacity, especially those heavily reliant on external financing.
John Blank, chief equity strategist at Zacks Investment Research, said CoreWeave is one of the largest neocloud providers and a "poster child for the whole story." He noted: "This is pressure facing the entire market, not just speculative behavior." Companies that borrow heavily to build computing power could face squeezed profit margins if financing costs rise faster than revenue growth. This pressure will be even more pronounced especially when companies have already borne increasing construction costs and depreciation expenses before their facilities reach full utilization.
This shift is particularly important for Asian markets, as Asian tech markets have large exposure to the semiconductor and memory chip industries, which are the core forces driving the global AI investment frenzy. John Blank said: "US semiconductor and AI-related companies have already shown overbuilding. In South Korea, the degree of overbuilding is even higher." South Korea's SK Hynix and Samsung Electronics have been among the biggest beneficiaries of surging demand for advanced memory chips, but as investors reassess the sustainability of AI spending, both companies' share prices have come under pressure. Billy Leung said: "Asian markets are more directly linked to the hardware and memory chip segments of the AI investment trade, so when investors start questioning AI capital expenditure and its financing, South Korea and other North Asian tech markets will be hit disproportionately harder." He added that earnings expectations in South Korea and China's Taiwan region are still rising, suggesting the recent selloff reflects deteriorating sentiment and valuation compression rather than a sudden collapse in demand.
HSBC CEO Georges Elhedery warned that investors and lenders must be more prudent in judging which data center projects deserve financing. He said: "You have to look at the quality of the lease agreements. You have to look at the pricing of these contracts. Not all contracts are of sufficiently high quality to give you full confidence to support them. Some of these projects may face greater challenges."
Comments