Three AI-Infrastructure IPOs Collapsed This Month. CoreWeave's Bonds Show Why
Three initial public offerings tied to the physical infrastructure behind the AI boom have collapsed or been shelved indefinitely in October, a cluster of failures that lines up almost exactly with the moment CoreWeave's bonds began trading like distressed debt. Taken together, the IPO retreats and the borrowing costs tell a more skeptical story about AI infrastructure financing than the industry's own talk of a continuing boom would suggest.
Three deals, three weeks, one message
SK hynix's NAND flash unit, Solidigm, ended its review of a US listing that bankers had floated at a valuation of up to $150 billion — a deal that would have ranked among the largest semiconductor IPOs on record. SK hynix cited market conditions and share price trends, though it insisted the decision did not reflect any change in Solidigm's growth prospects. The IPO was meant to help fund a new US fabrication plant; SK hynix's own market value has roughly halved since its June peak.
Nvidia-backed Firmus Grid, an Australian data-center operator, scrapped a roughly $5 billion offering that would have been the country's largest IPO in nearly three decades. The reversal came fast: days before pulling the deal, Firmus had signaled strong investor demand at a $30 billion valuation, up from just $10.5 billion less than three months earlier. Bankers couldn't convert that signaled interest into actual orders once global investors grew wary of the valuation multiples AI infrastructure deals were commanding.
AirTrunk, the Blackstone- and CPPIB-backed operator that runs more than 3 gigawatts of data-center capacity across Australia, India and Asia-Pacific, pushed its planned $1.5 billion Singapore REIT listing into the first half of next year. The company had considered sweetening the offering by raising its yield from 5.5% to 6% to attract buyers, but demand stayed below target even after that concession. Bisnow reported that Philippine data-center REIT Vitro, plus SB Energy, Holtec and Aggreko, have separately slowed or paused their own fundraising plans in the same window.
The bond market is already pricing the risk
The clearest signal sits in CoreWeave's debt. The company's 9.625% notes due 2032 have yielded as high as roughly 15%, and its 9.75% notes due 2031 have traded near 14.9% — levels associated with distressed credits, not a company S&P still rates B+. Debt CoreWeave issued in June at a 9.675% yield was trading near 15% last week. S&P Global Market Intelligence summarized the mismatch bluntly: CoreWeave's capital spending is close to hyperscaler scale without the balance sheet of a hyperscaler.
Wall Street is split on what that means. Goldman Sachs, led by analyst Gabriela Borges, upgraded CoreWeave from Neutral to Buy on October 19, calling it one of the strongest operators in the AI cloud industry and arguing its asset-backed project financing should keep future borrowing costs only modestly above current levels — shares rose about 6% in premarket trading on the call after having fallen roughly 25% the prior week. Not everyone agrees: D.A. Davidson's Gil Luria has warned that shareholders could lose everything if AI spending falters. The disagreement among analysts over the same set of bonds is itself a sign of how unsettled the financing picture has become.
A financing paradox at the top of the market
The stress is showing up in equities too. The Philadelphia Semiconductor Index was on pace for an eight-session losing streak in mid-October that erased roughly $2 trillion in chip-stock market value, matching its longest skid since the 2008 financial crisis. More than half of the index's constituents were sitting at least 10% below their 52-week highs even as the index remained up nearly 60% for the year.
Three withdrawn or delayed IPOs in a single month, a split verdict from Wall Street's own analysts, and a chip-stock index giving back trillions in value all point the same direction: the capital that spent the past two years rushing to fund AI's physical buildout is now pricing that buildout's risk much more like ordinary infrastructure debt than like a sure thing. Whether that caution spreads to the model layer — where valuations for the labs themselves have kept climbing even as the hardware underneath them gets a more skeptical reception — is the open question the rest of this quarter will answer.