IBM had its worst trading day in decades on Tuesday, and within hours the collapse had been absorbed into the market's favorite current anxiety: that AI spending is finally cracking. Shares fell as much as 24% after the company released preliminary second-quarter numbers that missed on both revenue and profit, and the initial read across financial media was reflexive — a bellwether enterprise tech company stumbles, ergo the AI capex cycle is turning. That read is wrong, and the reason it's wrong is more interesting than the panic it's replacing.
IBM's own numbers, filed with the SEC and detailed in a letter from CEO Arvind Krishna, tell a narrower and stranger story. Total revenue actually rose 1% year over year to $17.2 billion, and software revenue climbed 5%. The damage was concentrated almost entirely in Infrastructure, which fell 7%, driven by weakness in the Z mainframe line and its associated Transaction Processing software. Operating earnings per share came in at $2.93, short of the roughly $3.01 analysts had penciled in on revenue expectations closer to $17.9 billion. None of that reads like a company whose customers have stopped believing in AI. It reads like a company whose customers spent their money on something else at the worst possible moment for IBM's mainframe cycle.
Krishna was specific about what that something else was. In the final weeks of June, he wrote, enterprise clients shifted their quarterly capital spending toward servers, storage, and memory purchases in order to lock in supply-constrained infrastructure ahead of expected price increases. IBM had planned for a modest, low-single-digit decline in Infrastructure revenue this year as it lapped an unusually strong launch of its new z17 mainframe. Instead, that decline came in much sharper, and Krishna's own accounting of why was blunt: the company had built some supply-chain disruption into its guidance, but, as he put it, it "did not anticipate the magnitude of the capex reprioritization." That's a company admitting its customers made an emergency purchasing decision that had nothing to do with IBM's product roadmap and everything to do with a scramble happening one level down in the supply chain.
That scramble is the memory chip shortage, and it is the actual story of IBM's quarter. Global demand for high-bandwidth memory, the specialized DRAM stacks that feed AI accelerators like Nvidia's GPUs, has grown so explosively that it is cannibalizing the production capacity that used to go toward ordinary server and storage memory. Samsung, SK Hynix, and Micron control more than 95% of global DRAM output, and all three have been redirecting wafer capacity toward HBM, where hyperscalers like Microsoft, Google, Meta, and Amazon are signing multiyear supply contracts at premium prices to guarantee access. Industry estimates now put data centers' share of high-end memory consumption above 70% for 2026, and that share is expected to keep climbing. IDC projects DRAM supply growth of only about 16% this year, and NAND growth of roughly 17%, both well below the 20-to-30% range that used to define a normal memory cycle. This isn't a shortage that resolves itself once a new factory comes online — new fab capacity from Micron and SK Hynix isn't expected to reach volume production until 2027 at the earliest, and manufacturers have made a deliberate strategic choice to prioritize the higher-margin AI product mix even once that capacity arrives.
The price effect has been severe enough to change corporate purchasing behavior in real time. Conventional DRAM contract prices have been climbing by double-digit percentages quarter over quarter through the first half of the year, with some individual product categories reporting far larger jumps. Enterprise buyers who would normally spread their infrastructure purchases evenly across a fiscal year are instead front-loading them, buying servers, storage, and memory now rather than waiting, because waiting means paying more or facing allocation limits later. That is precisely the dynamic IBM described: clients didn't reduce their overall technology spending, they reordered it, pulling dollars away from mainframe and transaction-processing software purchases and toward securing physical infrastructure before the next round of price increases hit. Thomas Coughlin, a longtime memory industry analyst, has pointed to a prior period of capital-expenditure restraint that left manufacturers poorly positioned for the AI-driven demand surge, while remaining skeptical the current investment boom holds indefinitely — cautioning that the scale of AI infrastructure spending "can't last" and that a future boom-bust reversal is possible once demand eventually cools. Either way, that's a bet on the medium term; the near-term scarcity driving IBM's client behavior isn't in dispute.
The market has already started pricing this distinction correctly in one corner of the trade, even if the commentary hasn't caught up. SK Hynix's American depositary receipts have traded at a premium of more than 50% over the company's Korean-listed shares, an unusual and telling gap that reflects investors positioning around physical memory scarcity as a durable, multiyear constraint rather than a passing shortage. That is the opposite of what an AI-spending-slowdown thesis would predict. If enterprises were pulling back on AI infrastructure investment broadly, memory makers would be among the first to show it, since HBM demand is the most AI-exposed line item in the entire semiconductor supply chain. Instead, memory pricing and allocation tightness are the whole reason IBM's client base behaved the way it did last month.
This distinction matters because it reframes what kind of company gets hurt by the AI memory bottleneck, and it isn't the companies people assume. Nvidia and the hyperscalers building AI data centers are, if anything, protected — they're the ones locking in multiyear HBM supply agreements at the front of the line. The exposed companies are the ones one step removed: enterprise hardware and software vendors like IBM, whose customers have discretionary budgets that can be temporarily diverted toward securing conventional memory and storage before a price hike, at the direct expense of whatever else was on the purchasing calendar that quarter. IBM happened to be the company whose flagship product cycle collided with that diversion in the same three-week window in late June. The same dynamic is quietly reordering budgets across enterprise IT more broadly — anywhere a legacy hardware or software refresh competes for the same capital-expenditure dollars as a server or storage purchase, in a market that increasingly rewards buying now over buying on schedule.
None of this means IBM's Infrastructure business is healthy, or that Krishna's letter fully explains away the miss — the company also acknowledged that several large enterprise deals failed to close within the quarter, a separate execution problem layered on top of the memory-driven timing shift. But the framing that's already taking hold, that IBM's earnings are a crack in the AI capex thesis, gets the causality backwards. IBM isn't evidence that enterprises are losing faith in AI spending. It's evidence of how much AI spending, concentrated in a physical, capacity-constrained input that has nothing to do with software or model quality, is now powerful enough to distort the purchasing calendar of one of the largest legacy technology vendors on earth. The company that should worry investors isn't the one reporting a surprise miss on mainframe software. It's whichever company reports the next one, for exactly the same reason, while the market keeps mistaking a supply chain story for a demand story.
The mismatch is likely to recur. TrendForce and other supply-chain researchers expect memory prices to keep climbing through at least the third quarter, with relief unlikely before late 2027 at the earliest. Any enterprise vendor whose quarter depends on customers making a choice between a scheduled software or mainframe refresh and an unscheduled scramble to lock in memory and storage before the next price increase is exposed to the same distortion IBM just reported. The AI capex story that actually deserves scrutiny right now isn't about whether hyperscalers keep spending on GPUs. It's about how a physical bottleneck in memory manufacturing, three steps removed from any AI model or application, is now large enough to move the reported earnings of a 100-plus-year-old enterprise technology company in a single quarter.
