IBM Just Gave the AI Capex Story a Number—Here's How to Test If It's Real

IBM's July 14 warning didn't stay contained to IBM. Within hours, a wide swath of software and IT services names sold off—Accenture, Cognizant, and Infosys among the consulting names; ServiceNow, Workday, Salesforce, SAP, and Adobe among the software names; even Microsoft, HubSpot, and Datadog dipped despite reporting nothing that morning. The moves ranged from roughly 2% to as much as 10% intraday depending on the name and the hour, before several names pared their losses by the close. Vital Knowledge analyst Adam Crisafulli called it a "devastating blow" for the sector in a note that morning, tying the reaction directly to fear that the capex pivot he described would spread industry-wide.

The iShares Expanded Tech-Software Sector ETF (IGV)—a reasonable proxy for the group—fell as much as 4.5% intraday on the news alone. Zoom out further and the divergence gets sharper: the S&P Software & Services Index is down roughly 20% year-to-date in 2026, while the Philadelphia Semiconductor Index has gained more than 70% over the same stretch. That gap predates IBM's warning by months. What Tuesday did was give traders a specific, dated event to point to as evidence for why.

One detail worth sitting with: several of the stocks that fell hardest that day, including ServiceNow, Salesforce, and Workday, don't share IBM's mainframe exposure and have no z17 supply chain problem of their own. What they share with IBM is a customer base that's reprioritizing hardware over software contracts—and investors sold first, on the theory that the same mechanism could show up in their own numbers next.

If software budgets are getting squeezed, the money has to land somewhere—and the memory and semiconductor sector has been the visible destination all year. Micron crossed a $1 trillion market cap on May 26, 2026, one of the fastest runs to that valuation in market history, driven by high-bandwidth memory (HBM) demand tied directly to AI accelerators. According to Counterpoint Research, memory prices across DRAM, NAND, and HBM rose somewhere in the 80–90% range between the fourth quarter of 2025 and the first quarter of 2026. Micron has said its HBM production is sold out through 2026 and into 2027.

The structural piece matters more than the price spike. Manufacturing a unit of HBM consumes roughly three units of conventional DRAM wafer capacity, so every additional unit of AI memory built is conventional memory that doesn't get made. That's not a short cycle that eases once one product launch passes—it's a capacity trade-off baked into how the chips are manufactured, and it's a large part of why data centers are now estimated to consume something like 70% of all memory chips produced worldwide.

Servers themselves tell the same story. Dell booked $24.4 billion in AI-related orders in a single recent quarter and ended that period with a $51.3 billion AI backlog—management has said demand is outrunning supply, with memory again cited as the binding constraint. Hewlett Packard Enterprise has posted similar acceleration in its server business. Krishna's letter named "servers, storage, and memory" as the three things clients were racing to secure. Two of those three categories now have publicly disclosed, multibillion-dollar backlogs at IBM's own competitors in the hardware business.

Zoom out to the hyperscalers, and the scale gets clearer still. Microsoft, Amazon, Alphabet, and Meta are on pace to spend a combined $700–725 billion on capital expenditure in 2026, up roughly 77% from around $410 billion in 2025, with the overwhelming majority of that directed at AI data centers, GPU clusters, and custom silicon. That figure has been public and rising all year. What IBM's warning adds isn't a new number—it's evidence that the same reallocation logic hyperscalers apply to themselves is now reaching ordinary enterprise customers deciding between a software renewal and a server order.

The hyperscaler capex boom has been the dominant market narrative for most of 2026, and this site covered the resulting divergence between AI-infrastructure winners and richly valued software names in our AI-trade revaluation coverage from earlier this month, when semiconductor and AI-cloud names cracked even as the broader index hit new highs.

What's different about the IBM story is who's doing the reallocating. Up to now, the capex conversation was almost entirely about four or five giant companies spending on their own infrastructure. IBM's letter describes its own enterprise customers—banks, insurers, manufacturers, and government agencies, the kind of accounts that make up the bulk of IBM's book—making the same trade-off inside their own budgets. That's a meaningfully larger and more diffuse pool of spending than four hyperscalers deciding how much to spend on their own data centers, and it's the detail that turns a hyperscaler story into a software-sector story.

Whether that reallocation is now genuinely widespread, or whether IBM's mainframe-heavy customer mix made it unusually exposed to this particular dynamic, is exactly what the rest of Q2 earnings season should clarify. Our Q2 2026 earnings calendar already flagged AI capex scrutiny as a theme heading into this reporting cycle—IBM's July 14 warning was the first data point to arrive, not the last.

None of this requires predicting whether the AI trade is a bubble. It requires watching a specific, falsifiable pattern develop or fail to develop over the next several weeks.

Build a two-sided watchlist rather than a one-sided one. Pair the software and IT-services names that sold off in sympathy—Accenture, ServiceNow, Salesforce, Workday, SAP, Adobe—against the memory and semiconductor names benefiting from the same reallocation, like Micron, and track their relative performance day to day rather than in isolation. A scanner that lets you build and monitor that kind of paired watchlist in real time, such as Trade Ideas, makes it easier to catch the divergence early rather than after the headlines.

Treat the July 20–24, 2026 Big Tech earnings cluster as the next real test. Our Big Tech earnings and oil market preview covers the specific names and dates reporting that stretch—listen for whether their own capex guidance keeps climbing, and whether any of their enterprise or cloud customers are cited as delaying purchases for the same reason IBM gave.

Watch IBM's own July 22 call for the follow-through detail that actually matters: whether the deals that slipped in June were delayed or lost outright. A delayed deal that closes in Q3 supports the "timing, not trend" reading. A lost deal that goes to a competitor, or simply doesn't get replaced, supports the "structural crowding out" reading. That distinction, not the headline miss itself, is what should move the stock and its peers on the call.

The single biggest risk in trading this theme is overfitting one company's quarter into a sector-wide law. IBM's Infrastructure segment is unusually concentrated around its Z mainframe cycle, a business line most of its software peers don't share. Krishna also cited contract delays and cybersecurity-related spending pauses as separate factors—reasons that have nothing to do with AI capex and would resolve on their own timeline. A trader who reads every subsequent software miss as proof of the same mechanism, without checking bookings and customer commentary first, is pattern-matching on a sample size of one.

The reverse mistake is just as costly: dismissing the pattern entirely because IBM has idiosyncrasies. The sympathy selloff crossed too many unrelated business models—consulting, SaaS, enterprise software, all with different customers and different sales cycles—to wave away as coincidence. The honest position sits between the two: this is a real, testable hypothesis, not a confirmed trend, and it stays that way until more companies report.