Borrowed Demand — Anything for AI, Part II?

What does a mainframe company’s crash have to do with the AI boom?

On July 14th, IBM — a company older than the Federal Reserve — had the worst day in its trading history. The stock fell 25%, erasing roughly $67 billion of market value in a single session and taking out a record that had stood since Black Monday, October 19, 1987. [1,2,3] The S&P 500 barely noticed. Apparently there is nothing to see here — just the largest one-day repricing of a blue chip in nearly four decades.

So did AI finally kill the mainframe? No. Something stranger happened: the AI boom got its bill paid — by everyone else. IBM’s CEO told investors that clients abruptly shifted spending late in the quarter toward servers, storage, and memory ahead of announced price increases, and that “numerous large deals failed to close on the timelines we expected.”[4] The software and mainframe budgets those deals were coming from did not shrink because customers stopped needing software and mainframes. They shrank because the money was needed, right now, to panic-buy AI hardware.

Get our insights direct to your inbox: SUBSCRIBE

We think July 14th gave investors their first clean look at something we are calling borrowed demand: the AI hardware boom is now being funded, in part, out of the rest of the technology sector’s own revenue line — and part of what shows up as “demand” is hoarding.

We have written about this before: in April we asked whether there was any margin of safety at 12 times sales for the AI beneficiaries, and noted that 16 of the 20 best performers in the S&P 500 were tied to the AI infrastructure boom — the “picks and shovels” names. We titled that paper Anything for AI. It turns out we underestimated how literally the market would take the title. Companies are not just paying anything for AI — they are taking the money from everything else.

The chart below shows where 2026’s technology budgets are going. Gartner expects overall IT spending to grow 10.8% this year, but data-center systems to grow 31.7% and servers 36.9% — while software growth gets revised down to 14.7% and devices limp along at 6.1%.[5] When the overall pie grows 11% and the AI slice grows 32–37%, someone else’s slice is smaller than planned. On July 14th, we found out whose.

The receipt: what did the worst day in IBM’s history actually say?

Start with the numbers that caused the damage. IBM pre-announced preliminary second-quarter revenue of

$17.2 billion, up just 1%, against a consensus near $17.9 billion; operating EPS of $2.93 against $3.01 expected. [6,7] A roughly 4% revenue miss does not normally cost a company a quarter of its market value in a day.

What spooked the market was the reversal. In the first quarter, IBM’s Infrastructure segment grew 15%, with mainframe revenue up 51% on the z17 cycle, and Software grew 11%.[8]One quarter later: Infrastructure down 7%, Software decelerating to 5%, Consulting flat.[9] Gross margin slipped 100 basis points, and year-to-date free cash flow stands at $4.8 billion — against the $12.9 billion the business generated over the trailing twelve months. [10,11] A mainframe cycle that was shipping 51% growth does not die of natural causes in ninety days. The chart below shows the cliff.

IBM did not lose its customers’ budgets to a competitor. It lost them to the AI buildout’s shopping list. The company’s own explanation is the thesis of this paper in one sentence: clients raided the budgets that were earmarked for software renewals and mainframe deals to buy servers, storage, and memory ahead of price increases — spending pulled forward not by demand for computing, but by fear of what computing components will cost next quarter.[12,13] The market understood immediately that this is not an IBM-specific problem: ServiceNow, Accenture, and Microsoft all sold off in sympathy the same day.[14]

Note what the company did not say. It did not say demand disappeared. It said deals “failed to close on the timelines we expected” — the demand was deferred, not destroyed, while the customers’ cash went to the hardware queue. That distinction matters enormously for how you read the crash: a company whose customers still need its product but temporarily cannot pay for it is a different animal from a company being disrupted out of existence. It is also, we would note, exactly what “borrowed” means. IBM’s missing quarter did not vanish — it is sitting in server racks and memory inventories on its customers’ loading docks.

Is IBM now the bargain of the decade, then? Here is where we disappoint the bottom-fishers. The worst day in IBM’s history took the stock all the way down to… average. After the crash, IBM trades at 19.4x trailing earnings (17.9x forward), 3.0x sales, a 6.3% free-cash-flow yield, and a 3.2% dividend — carrying roughly $58 billion of net debt.[15] Cheaper than an S&P 500 at 27x earnings? Sure. But “cheaper than one of the most expensive indexes in history” is a low bar, and nearly $60 billion of net debt does not sweeten it. To screen the way the wreckage-tier software names we profiled in April screen, IBM’s free-cash-flow support would need to be half again richer than it is today. It isn’t there. The data says: average, not cheap — wreckage prices without, yet, wreckage-grade value. Whether it becomes the former or stays a value trap is a question the next two quarters of deal-closure data will answer; we would want to see the deferred deals actually close.

Where the money went: the memory tax

Follow the panic-buying to its source and you arrive at the memory market. DRAM contract prices rose roughly 90% in the first quarter of 2026 — a record — then another ~58% in the second quarter, with double-digit increases still forecast for the third. [16,17,18] The chart below shows the trajectory. Supply is not riding to the rescue: industry forecasts put 2026 DRAM bit-supply growth at just 16% (NAND 17%), below prior-cycle norms, with tightness persisting into late 2027,[19] in part because roughly a quarter of DRAM wafer capacity has been diverted to high-bandwidth memory for AI accelerators. [20]

The shortage is now so broad it has reached backwards through time: even DDR2 — a memory standard introduced in 2003 — is seeing contract prices rise as manufacturers abandon legacy lines for AI product.[21] And the deceleration in the chart above is not relief arriving; TrendForce attributes it to long-term agreements capping how fast sellers can reprice, and to consumer buyers simply hitting the limit of what they can absorb.[22] Prices are still rising into 2027 — just against a base that has already doubled.

Step back and consider what this means. For thirty years, technology’s quiet gift to every corporate budget was deflation — the same compute, storage, and memory got cheaper every year. That dividend just inverted. Enterprise IT is now an input-cost cycle: the same server that credentialed a CFO’s budget last year now costs meaningfully more, and everyone knows the price goes up again next quarter. We have written before about technology’s earnings elephant — the sector’s quiet slide from capital-light to capital-intensive. The memory tax is that elephant, presenting its bill.

And software — the segment whose budgets are being raided — is squeezed from both ends. At the same moment its buyers’ wallets are being emptied into the hardware queue, Gartner estimates $234 billion of enterprise application-software spend is at risk of displacement by agentic AI. [24] The SaaSpocalypse we described in April was a valuation event. This is a budget event.

How much of the demand is real? The hoarding evidence

Here is the remarkable part: the people with the best view of the order book — the memory makers themselves

— are behaving as if they do not believe it. Samsung, SK Hynix, and Micron are now demanding that customers disclose end-customers and order details to weed out double-ordering and stockpiling, and are shifting large buyers onto short-term, post-settlement contracts. [25, 26] Think about that. When the seller starts asking who the end customer really is, the seller suspects the order book is padded.

The contract terms tell the same story. The shift to “post-settlement” pricing — where the final price is set after delivery, at whatever the market then bears — is what sellers do when they believe prices are still going up and buyers are ordering ahead of need.[27]No seller offers those terms to a customer whose orders they trust.

Panic-buying memory ahead of announced price increases is the enterprise version of hoarding toilet paper in 2020: individually rational, collectively ruinous — and worthless as a demand signal. The semiconductor industry has run this movie before: the double-ordering wave of 2021 was followed by the inventory collapse of 2022, and procurement desks report that pandemic-era double-orders were still rippling through the channel when this cycle’s panic began stacking new ones on top.[28] What did the industry learn? Nothing.

The hoarding is one layer. Look at how the whole buildout is financed and the pattern extends. We have called the AI buildout a debt-fueled echo of the dot-com era and documented its circular financing loops before; the official sector is catching up to that view. The Bank for International Settlements reports that data-center debt issuance nearly doubled last year to $182 billion — hyperscalers alone issued $121 billion of bonds, four times their five-year average — with private credit projected to add another $800 billion over the next two years.[29] Meanwhile the revenue side of the ledger still has not shown up: Sequoia’s well-known arithmetic puts the gap between AI infrastructure spend and the annual revenue needed to justify it at roughly $600 billion,[30] Allianz Research measures the AI capex-to-revenue divergence at ~46% — wider than the 32% at the peak of the 2001 telecom excess [31] — and roughly 95% of enterprise GenAI pilots still produce no measurable P&L impact. [32]

Add it up and the demand signal that crashed IBM is borrowed three ways: