IBM’s Invisible Empire: How Mainframes Still Process Trillions While the Company Vanished From View

IBM has not sold a consumer product in more than two decades. Its last PC left the lineup in 2005. The x86 server business followed in 2014. Yet every time a credit card clears at checkout or an airline reservation system books a seat, the odds point to hardware few people ever see.

That hardware sits in climate-controlled rooms. It runs code written before many current engineers were born. And it moves money and data at scales that make cloud outages look like minor inconveniences. The tech giant nobody sees.

Jonathon Jachura laid out the case in a recent piece for MakeUseOf. IBM walked away from consumer markets on purpose. Sam Palmisano, then chief executive, ran the numbers. Services and software delivered gross margins above 25 percent. Personal computers did not. The 2005 sale of the PC division to Lenovo for $1.75 billion marked a clean break. Nine years later the x86 server unit went the same way for about $2.1 billion. Mainframes stayed behind.

Those systems never left the critical path. IBM claims they handle 90 percent of global credit card transactions. The figure comes from an Oxford Economics survey the company commissioned, though independent reports often cite 87 percent. Either way the volumes are enormous. Close to $8 trillion in card payments a year. Roughly 29 billion ATM transactions. More than 70 percent of worldwide transaction value by IBM’s count.

But a single quarter can shake confidence. On July 22, 2026, IBM reported second-quarter revenue of $17.16 billion. Infrastructure fell 7 percent. IBM Z revenue dropped 42 percent year over year. The stock tumbled 25 percent in its sharpest single-day decline on record. Full-year guidance slipped to 4 to 5 percent growth. Analysts cut estimates fast.

Arvind Krishna, IBM’s chairman and chief executive, explained the drop in a letter to investors released days earlier. Clients had rushed purchases of servers, storage and memory to beat expected price hikes. That pulled spending away from mainframe refreshes. Transaction processing software took a hit too. “What played out was worse than our expectations,” he wrote. Yet he pointed to the bigger picture. The z17 launch ranks as the strongest start to any mainframe program in company history. Measured against the prior cycle, it runs at nearly 130 percent. Eighty-five percent of installed processing capacity is holding steady or growing.

Krishna repeated the message on the earnings call. “While clients continually evaluate workload placement, we see no evidence of clients moving off the mainframe.” The z17, he insisted, remains on track. And the long-term incumbency looks solid.

Announced in April 2025 and available from June, the z17 packs the Telum II processor. On-chip AI acceleration delivers more than 450 billion inference operations a day at about one-millisecond response. A 40 percent increase in cache helps. An optional Spyre accelerator card, expected later in 2025, adds 32 AI cores and 128 gigabytes of memory. It targets larger models and generative workloads that once required separate systems.

IBM detailed the chips at Hot Chips in 2024. The company positioned them as proof that AI belongs in the transaction layer, not as an afterthought. Fraud detection can now score a payment in the same cycle that processes it. No sensitive data leaves the secure enclave. Response times stay low enough that customers never notice the question was asked. That matters when every second costs money.

Airlines have trusted this iron for generations. American Airlines and IBM built Sabre in the early 1960s to replace paper-based reservations. The underlying Transaction Processing Facility, or TPF, evolved into z/TPF. It still runs core high-volume workloads at several carriers. Amadeus retired its last mainframe in 2023. Sabre has shifted shopping, inventory and profiles to open systems over more than a decade. What remains on the old platform tends to be the part where a migration failure cancels flights.

Banks face even thornier math. Core ledgers often consist of COBOL written decades ago. The original programmers have retired. Business rules live inside the code rather than separate documentation. Rewriting everything without a single balance error has proven expensive and risky. Few institutions attempt it at scale. Krishna told analysts he sees no signs of departure. The platform’s resilience, he said, continues to drive decisions.

Recent coverage reinforces the tension. CNBC reported the lowered forecast and the mainframe slide. It noted how organizations redirected budgets ahead of price increases. Channel Dive quoted Krishna directly on the lack of evidence for workload migration. Both pieces highlight that the 42 percent drop reflects timing more than structural decline.

IBM’s own newsroom posted the preliminary letter on July 14 and full results eight days later. The documents show software revenue rose 5 percent to $7.76 billion. Consulting held flat at $5.33 billion. Free cash flow jumped 70 percent in some accounts, helped by AI-driven productivity. The mainframe business, despite the headline number, sits well ahead of historical refresh cycles.

Critics still ask whether the model can last. Cloud providers push distributed architectures. Open-source alternatives nibble at edges. Yet replacing decades of audited, battle-tested code carries its own cost. One wrong decimal in a payroll run or settlement system can trigger regulatory nightmares and customer fury. Mainframes deliver five-nines reliability with hot-swappable components. Technicians replace parts while the system keeps running.

That invisibility forms the point. Consumers never see the machine approving their gas purchase. They don’t thank the COBOL routine that clears their direct deposit. They notice only when it fails. IBM stopped chasing attention in 2004. It bet on the jobs nobody wants to think about. So far the wager has paid in margins and market position that consumer brands envy.

And the AI additions suggest the company intends to extend that run. By baking inference directly into the processor, IBM aims to keep sensitive workloads inside the trusted perimeter. Banks can run fraud models without shipping card data across networks. Airlines could personalize offers in real time without extra latency. The Spyre card expands capacity for larger language models, including IBM’s own Granite family.

Recent X discussions touched on data centers and AI infrastructure but rarely mentioned IBM Z specifically. One thread noted Amazon’s big bet on OpenAI and the resulting demand for cloud and chips. Another highlighted the sheer number of U.S. data centers versus other countries. The conversation stays focused on visible hyperscalers. The quiet machines that settle those cloud bills stay out of frame.

IBM’s strategy looks stranger the longer one considers it. Most tech giants fight for consumer mindshare. They build phones, laptops, streaming services and social networks. IBM exited that arena deliberately. It kept the back office. The result is a business that processes trillions, powers legacy systems still too risky to replace, and now layers modern AI onto iron older than many startups. The quarterly numbers will fluctuate with refresh cycles. The underlying demand for reliable transaction processing has not gone away.

Next time the card reader beeps approval in under a second, remember the system that made the decision. It probably runs on hardware from a company whose logo vanished from store shelves years ago. That company planned it that way. And the world keeps turning on machines few will ever see.