Summary
IBM has not sold you anything in 21 years. No laptop, no phone, no icon on your home screen. The last product an ordinary person could walk into a store and buy left the company in 2005, and nearly everything consumer-facing since has been sold off or wound down. Plenty of people have never knowingly touched an IBM product. They touch one constantly. The machine that approves a card at the gas pump, checks a balance at an ATM, or clears a payroll run at 3 a.m. is very likely IBM hardware sitting in a building nobody photographs. Tech history is crowded with companies that lost the consumer market and vanished, often after fighting dirty on the way down, the way Microsoft did to its DOS rivals. IBM lost the consumer market on purpose and kept the profitable half. IBM walked away from consumers on purpose The ThinkPad sale was the clean break IBM built the machine that created the category. The 5150, released in 1981, used off-the-shelf parts and an open architecture. That decision handed the world a standard and handed every competitor a blueprint at the same time. Clone makers did the rest. IBM’s share of the PC market fell from roughly 75% in the mid-1980s to about 12% a decade later, and the margins went with it. IBM sold its PC division to Lenovo for 2.1 billion in 2014. Both contracts carved out the same exception: Lenovo got the commodity hardware, and IBM kept the mainframe. Sam Palmisano ran the numbers and reached an unsentimental conclusion. Services and consulting returned gross margins above 25%. PCs returned almost nothing, no matter how many the company shipped. In December 2004, IBM announced it was selling its Personal Computing Division, ThinkPad and all, to Lenovo for 2.1 billion. Both contracts carved out the same exception. Lenovo got the commodity hardware. IBM kept the mainframe. The machine it kept is the one that mattered What a z17 does that a rack of servers can’t Mainframes are not fast in the way people usually mean. A decent gaming PC will beat one at nearly anything you would do at a desk. What a mainframe does is move staggering volumes of tiny transactions without losing any of them, and keep doing it while technicians swap out parts of the running system. The current generation is the z17, announced in April 2025 and generally available that June. It runs on IBM’s Telum II processor with AI inference built directly onto the chip, so a fraud model can score a payment in the same place the payment is happening rather than shipping sensitive data elsewhere and waiting for a verdict. IBM rates the system at up to 450 billion inference operations per day at roughly 1ms response time, half again more than the z16 managed. An optional Spyre accelerator card adds 32 cores and 128GB of memory for larger models. None of that is a speed pitch. It’s a promise that the answer arrives before the customer notices a question was ever asked. Your card swipe almost certainly touches one Where the 90% figure comes from IBM says mainframes handle 90% of the world’s credit card transactions. The number gets repeated constantly and deserves a footnote. It comes from a survey of more than 2,500 technology executives that Oxford Economics conducted at IBM’s request. Independent coverage more often cites 87%, which also traces back to IBM. The sourcing is worth knowing, though almost nobody in payments argues with the shape of the claim. IBM has put the volume at close to 8 trillion in payments a year, alongside roughly 29 billion ATM transactions. Independent coverage more often cites 87%, which also traces back to IBM. Infrastructure this important is only visible when it breaks, which is the same reason the plumbing of the internet itself goes unnoticed until something stops resolving. Airline bookings started on the same iron TPF is still there, but the map is changing American Airlines and IBM built Sabre in the early 1960s to replace a reservation process run on paper cards. It worked, and IBM promptly generalized the technology into PARS and sold it to other carriers. The control program underneath was split out in 1969, became ACP, and by December 1979 had been renamed TPF, for Transaction Processing Facility. The modern descendant, z/TPF, is still in production. The picture is genuinely shifting, though. Amadeus decommissioned its last mainframe in 2023. Sabre has spent more than a decade relocating shopping, inventory, pricing, and customer profiles onto open systems. What tends to stay on TPF is the highest-volume core, where a botched migration does not produce a bug report. It cancels flights. Banks haven’t ripped them out because the rewrite is the risk COBOL, and the cost of being wrong once A large bank’s core ledger is frequently decades of COBOL, a language older than the moon landing. Very few of the people who wrote it are still employed there. The business rules encoded in it were never fully documented anywhere else, which means the code is not just the implementation, it is the specification. Old languages outlive their obituaries with some regularity, and C is the obvious parallel. Rewriting all of that is possible. Doing it without a single wrong balance is the part nobody has solved cheaply. On IBM’s July 2026 earnings call, CEO Arvind Krishna told analysts the company sees no evidence of clients leaving the platform, and that 85% of installed processing capacity is holding steady or growing. A 42% drop that isn’t what it looks like Refresh cycles make the numbers lie IBM reported second-quarter results on July 22, 2026, and IBM Z revenue was down 42% year over year. The stock had already fallen 25% after preliminary figures the week before, its sharpest single-day decline on record. Full-year guidance came down to 4–5% growth. Mainframes sell in waves, and that number is a wave receding. Measured against the previous cycle at the same point, z17 is running at nearly 130% of z16, which was itself the strongest program IBM had ever recorded. That works out to about 3 billion more in attached software. Compare that to Adobe, which Wall Street currently prices like a dying business while its revenue keeps climbing. Invisible by design IBM stopped competing for attention two decades ago and started competing for the jobs nobody wants to think about. That trade looks strange until a Friday afternoon at a busy grocery store, when 40 card readers in one building all say approved within a second of each other and not one person in the checkout line wonders why. The company that made that boring is doing exactly what it decided to do in 2004. Anyone curious about how much of daily life runs on hardware they will never see has a good place to start looking.