The technology boom version 2.0 has been driven by insatiable spending on data center infrastructure. That spending continues, but many of the companies providing the hardware necessary to run AI chatbots, like ChatGPT and Gemini, are in the midst of a brutal sell-off that’s knocked the technology-heavy Nasdaq 100 down nearly 10% and flirting with correction territory.
The pain has been particularly hard felt among semiconductors and cloud plays, many of which have been taken to the proverbial woodshed, including these five, which tumbled over 50% from summertime highs to recent lows:
- Intel: -58%
- Arm Holdings: -55%
- Oracle: -52%
- Sandisk: -53%
- KLA Corporation: -52%
- Source: Yahoo! Historical Quotes
AI spending boom fears take root
Stocks tend to move based on emotion, and now, into its fourth year, the AI trade has gotten pretty crowded. I’ve seen this movie before during my 30-year career, so it’s not surprising to me to see the drawdowns. The usual progression: A new idea lands with many doubters and few believers. As stocks run, more converts jump on board until most of the money on the sidelines has joined the party, leaving few new buyers to step in to support stocks on pullbacks.
Tech stocks like Intel didn’t drop because their businesses have suddenly imploded. Instead, they’ve dropped because supply from sellers locking in gains has dwarfed demand. Basic economics sounds simplistic, but sometimes the simple answer is the correct one.
If there’s a specific catalyst beyond protecting gains, ratcheting back peak data-center spending is the most compelling. There’s mounting fear that we’re about to hit peak capex for data center spending in 2027, setting the stage for a slowdown.
The stakes are high, given the dollars at stake.
In mid-October, JPMorgan said 2026 AI capex for the five biggest hyperscalers (Amazon, Alphabet, Meta, Oracle, and Microsoft) would be $900 billion, climbing above $1 trillion in 2027.
More Tech:
- Meta and Microsoft just sent employees a memo about Anthropic
- Microsoft cutting off popular software with no extension Oct. 13
- Apple’s biggest AI advantage may face new privacy challenge
Most of that spending to date has been financed by free cash flow (and, arguably, by circular financing deals, thanks to investments from the likes of Nvidia). However, companies are increasingly turning to debt, exposing balance sheets to risk.
For instance, Goldman Sachs pegged the percentage of AI capex financed by debt at 33% in 2026, up from 26% in 2025. In 2027, Goldman Sachs says 35% will be paid for by debt, totaling roughly $400 billion.
If revenue growth underdelivers or lenders get antsy and borrowing gets pricier, we could end up with a major capex reset that unravels the AI trade, or so the bear argument goes.
Bank of America says demand justifies spending
There’s a silver lining to more judicious AI capex, though, suggests Bank of America. Analyst Vivek Arya thinks that if capex budget cuts happen, it would “lower peak profit but preserve compute scarcity, extend the AI cycle and support higher multiples over time.”
Think a shorter peak and a flatter runway for the entire AI cycle, which could mean less boom-and-bust action for underlying stocks.
Arya thinks there’s plenty of incentive to continue investing in AI, given surging backlogs.
“With $2.5tn+ backlog and surging enterprise adoption, demand looks durable: scarce compute, not excess capacity, is the risk,” wrote Arya in a research note shared with me.
Also read: Elon Musk sends incredible news to Micron stock investors
“Slower supply growth against resilient demand could lift utilization, support pricing and reduce overbuild risk. The result could be lower peak earnings, but a longer earnings runway and potentially higher multiples for chipmakers already valued as if the cycle is ending,” added Arya.
Major hyperscaler revenue certainly seems to support the argument for spending.
Amazon’s AWS revenue rose 39%, Microsoft Azure sales grew 49%, and Google AI revenue increased over 100% year over year in the third quarter, according to Arya.
Are beat-down tech stocks bargain bin buys?
The risk of a slowdown or debt reckoning is reflected in the most punished technology stocks. Intel has been late to the party in terms of AI-optimized chips, but optimism over U.S. fab capacity builds, and deeper pockets from U.S. and Nvidia investments caused its shares to soar before the recent drop, despite operational headwinds.
At Oracle, the fear of overspending to become one of the biggest hyperscalers has raised major debt concerns, contributing to its recent drop, despite the company’s argument that its backlog is solid and justifies the push.
ARM Holdings makes its money by selling semiconductor designs, and a slowdown in data center spending would crimp its upside. It has historically traded at a premium valuation because of its ‘asset-light/no-fab’ status, making it more susceptible to drawdowns as investors press sell on so-called expensive players first.
A downthrottling in capex would also be bad news for KLA Corporation, which makes inspection and measurement machines used to create semiconductors. If the data center buildout slows, so would chip demand, and the need for new fab equipment. Like ARM, KLA has long been considered a premium play in the semiconductor equipment space, trading at a higher valuation than peers such as Applied Materials and Lam Research, making it susceptible to sell-offs.
SanDisk sits in a different space, given that it makes memory, but lower spending on high-end server capacity means less memory is needed. Given the 303% surge in NAND flash memory this past year, and the historical cyclical nature of memory market demand, a bit of investors’ caution is warranted.
It’s hard to judge whether any of these individual stock drawdowns have run their course and are ready to reverse, but Arya thinks semiconductor plays aren’t overpriced relative to history.
“Since CY19, the SOX index has traded at 12.5x-30x forward PE multiple, ranging from a 20% discount to a 40% premium versus SPX. Today, 16.4x represents a 10% discount, consistent with AI-cycle troughs in CY23-1H26 and only 10pts above the deepest relative discount since CY19.”
Of course, stocks can always get cheaper, but Arya’s math shows that these stocks aren’t historically priced excessively, and that should give some modicum of comfort to bargain hunters, especially if Arya is right that a capex reset wouldn’t trigger demand destruction nearly to the extent some fear.
“The market therefore appears to be discounting capex risk before the demand data have cracked. If paying AI demand stays resilient, lower capex could reduce overbuild, preserve scarcity, extend the EPS cycle and justify higher multiples, in our opinion,” concluded Arya.
