In 1991, InfoWorld editor Stewart Alsop predicted the last mainframe would be unplugged by 1996. The client-server revolution was underway. PCs were eating the world. By 2002, Alsop admitted he'd been wrong.

Around that same time, the Industrial & Commercial Bank of China was installing its first mainframes. Today ICBC runs 1.5 billion transactions daily on that infrastructure. Someone was building while the industry was writing eulogies.

Ninety percent of credit card transactions still clear through mainframes. Somewhere between 220 billion and 800 billion lines of COBOL remain in active production, depending on whose estimate you trust. Both numbers are enormous, and the gap between them says something about how poorly understood this layer is, even by the people who depend on it.

Replacing it has turned out to be staggeringly hard. An analysis of 29 mainframe-to-cloud migrations between 2020 and 2025 found 66% failed to meet objectives. Nearly a third were abandoned after discovering insurmountable data conversion problems. TSB Bank's 2018 migration locked customers out of their accounts for weeks, cost the CEO his job, and drew roughly £67 million in combined regulatory fines. The most expensive documented failure: $41 million over 52 months on a COBOL-to-Java banking conversion, abandoned at 40% completion.

The failures have a specific shape. Mainframe applications are designed for sub-millisecond I/O. Move those tightly coupled workloads to distributed cloud environments and network latency becomes a performance cliff. Batch jobs that took an hour stretch to eight. Some large banks run 300 million lines of COBOL as their core banking solution, architected over decades to exploit the mainframe's specific strengths. That conformity is the lock.

So the mainframe sank beneath the conversation. The PC era, the client-server era, the cloud era all arrived on schedule. Each was real. Each deposited its own layer on top without removing what was underneath. The mainframe became invisible precisely because everything above it worked. When Deloitte surveyed mainframe decision-makers and found 91% planned to expand their footprint, the number reflected a rational calculus: infrastructure where replacement had become riskier than continuation, quarter after quarter, until the question stopped being asked.

By most estimates, hyperscalers are spending over $600 billion on AI infrastructure in 2026. That spending creates applications, integration patterns, institutional knowledge that compound year over year, the same way mainframe ecosystems compounded through the 1980s and 1990s. The 2026 Arcati survey found 78% of enterprises still report revenue or transactions totally dependent on mainframe systems. Nobody decided mainframes should persist. The cost of continuation just kept looking more reasonable than the cost of change, one quarter at a time, until decades had passed and the layer was permanent. Cloud AI infrastructure is building that same kind of gravity right now, in real time, at a scale the mainframe era never approached.

The mainframe became the floor everyone else built on. And nobody remembers choosing to keep it.