Scarcely known outside its home market until this week, Elice Inc. on Tuesday found itself at the heart of a roiling debate about AI valuations and whether companies will be able to deliver on them.

The $750 million data center and AI developer became a surprise poster child for the loss of faith in so-called neoclouds when it tanked 32% on its debut in Seoul. It joined a half dozen other companies in the sector that just over the past few weeks have shown investors souring on artificial intelligence. Australia’s Firmus Grid Ltd. and AirTrunk have delayed public offerings, Britain’s Nscale Ltd. and US peer SB Energy Inc. faced investor pushback on their valuations, and CoreWeave Inc.’s bond yields shot up to 15%.

In a sharp reversal from the broad-based pile-in of recent years, global markets are now being much more selective and demanding with AI bets. Neoclouds, the firms that specialize in running dedicated facilities for AI computation delivered over the internet, are an essential part of the supply chain that customers like Anthropic PBC and OpenAI tap in multibillion-dollar deals for capacity. But worries about financing, delays in data center construction and the uncertainty of returns have jolted the entire tech ecosystem, triggering the longest losing streak in semiconductor stocks since the global financial crisis.

“Elice’s disastrous debut is another confirmation that the era of ‘buy everything AI’ has given way to ‘question everything AI,’” said Hebe Chen, senior market analyst at Vantage Global.

Neoclouds are particularly vulnerable. New Jersey-based CoreWeave is the best known among them, pioneering the business model of taking on significant debt to load up on Nvidia Corp. graphics processing units (GPUs) that it later rents out in multiyear contracts. Its 18 months on public markets have shown the wild swings of confidence in this strategy: the company’s market value approached $90 billion last year, with Nvidia a vocal backer and investor, but it’s endured a brutal selloff since then and is now more than 60% off its peak.

Firmus’s initial public offering valuing it at $30 billion collapsed, while Nscale is facing investor skepticism on a listing at a similar valuation. The two have run into the real-world challenges of converting AI computing demand and orders into functional operations, as both have only a small fraction of their planned data center capacity live. As the cost of capital increases, investors are becoming more wary of companies depending heavily on future revenues.

Soaring borrowing costs were the catalyst for CoreWeave’s most recent selloff. Its bonds this month sank into distressed territory — leading Nvidia Chief Executive Officer Jensen Huang to step in to reassure markets — while AI investor SoftBank Group Corp.’s credit risk rose to the highest level in nearly four years. SB Energy is SoftBank’s vehicle for setting up AI infrastructure in the US, but its IPO talks have slowed to a crawl and regulatory approvals have also taken longer than expected.

One of the world’s most enthusiastic investors in AI, SoftBank faces pressure on several fronts. Its largest holding is a significant share in OpenAI, which already pushed its IPO into next year, and the SB Energy debut was designed to help it raise funds to support its construction projects in the US. SoftBank pushed through the largest corporate junk-bond offering on record last month, raising roughly $11 billion, shouldering yields it’s never had to pay before. Its credit-default swap level is now the highest among major Japanese companies.

Australia-based AirTrunk, backed by Blackstone Inc., had been planning to gauge interest in a $1.5 billion real estate investment trust in Singapore offering a roughly 6% yield, but it pushed that back into next year after September’s interest rate rise in the US.

This has all had a knock-on effect, as junk bond buyers are increasingly freezing out tech startups looking for financing, compounding the problem. On Thursday, JPMorgan Chase & Co. and Goldman Sachs Group Inc. hiked the yield on a $5 billion leveraged loan sale for Volta Infrastructure Holdings Ltd. as investors demanded greater compensation to buy risky debt tied to the artificial intelligence boom.

“When the sentiment is super bullish, no one cares, and when it’s super bearish, it’s all doom,” said Leonid Mironov, a portfolio manager at Gavekal Capital Ltd. in Hong Kong. “The answer is cash flow: If there’s enough of it to go around, neoclouds will be fine. Positioning being this bearish makes it difficult to gauge actual fundamental impacts.”Even as enthusiasm has radically cooled off, indications about underlying demand for AI hardware remain strong.

Chip industry linchpins Taiwan Semiconductor Manufacturing Co. and ASML Holding NV both beat expectations with their earnings and near-term outlook and said they have visibility for high demand through the end of this decade. But the AI halo effect — which in April sparked a nearly 600% rally in sneaker maker Allbirds when it pivoted to building AI infrastructure — appears to have conclusively worn off and potentially turned into a form of contagion.

Elice has government contracts, data centers already in place and profitable operations, yet it still got caught up.

“When major neocloud names face turbulent swings in investor sentiment, their smaller competitors in Asia are bound to face the repercussions,” said Homin Lee, a strategist at Lombard Odier. “This is an example of that dynamic. We maintain our focus on large-cap tech names in North America and Asia.”To be sure, the difficulty with new debuts isn’t limited to the AI sector. Smart-ring maker Oura Inc. also shelved its IPO plans in September after interest rates climbed higher and the Iran war dragged on.

But the AI trade has been the engine powering growth across every major market since the frenzy sparked by ChatGPT’s debut. It delivered the biggest IPO ever with SpaceX’s debut earlier this year and potentially a bigger one still with Anthropic as soon as next month. Anthropic, like neoclouds, is essential for sustained demand for AI hardware like GPUs, power supplies, networking and cooling.

The timing and success of Anthropic’s debut on public markets will be key to watch for anyone hoping the AI boom sustains its flagging momentum. The company, which has already committed to hundreds of billions of dollars of AI infrastructure contracts and is looking for a market valuation in the trillions, is in talks with Nscale for $45 billion of additional capacity. Its ability to sustain its high levels of funding and spending will be essential.

“Any negative data point or hint of nervousness causes panic for the tech sector as this multitrillion-dollar AI foundation buildout is tied to a handful of major tech players,” said Dan Ives, senior managing director at Yorkville Ives & Co.