IBM lost $40 billion in market value in one day. The trigger?
A February 2026 blog post from Anthropic claiming AI can “rapidly refactor” COBOL code, turning what used to take years into a matter of quarters. Shares plunged 13 percent—IBM’s worst single-day crash since October 2000.
Investors saw the headline and panicked: if AI can translate COBOL overnight, IBM’s mainframe cash cow is dead.
The problem? Anthropic’s claim solved a problem that doesn’t exist.
Translation isn’t modernization. And the $40B wipeout exposed how badly the market—and apparently, AI companies—misunderstands what enterprise legacy systems actually require.
The market panicked over a language problem that was never the real issue
Anthropic’s blog post promised teams could modernize COBOL codebases “in quarters instead of years.” Wall Street heard “IBM’s revenue model is obsolete” and hit sell. But IBM SVP Rob Thomas’s response cuts through the hype: “Translation captures almost none of the actual complexity”—the modernization challenge isn’t a COBOL language problem.
He’s right.
AI can translate COBOL syntax to Java or Python in hours. That’s impressive. It’s also roughly 5% of what mainframe modernization requires. The other 95%? Redesigning data architecture that’s been optimized for decades. Preserving transaction integrity across systems processing billions of operations daily. Replicating runtime environments where hardware and software are so tightly coupled that separating them breaks everything.
Investors reacted to the translation part while ignoring the infrastructure part. Anthropic’s aggressive enterprise push into healthcare and now legacy modernization reveals ambitious market positioning—but the technical claims need verification against actual production systems.
COBOL isn’t dying—it’s running $3 trillion in daily transactions while IBM’s mainframe revenue hits 20-year highs
The “COBOL is ancient and dying” narrative collapses when you check the numbers. COBOL is 66 years old, designed in 1959. It also powers 95% of US ATM transactions and roughly $3 trillion in daily commerce. These aren’t legacy systems limping toward retirement—they’re the backbone of global finance.
And they’re growing.
IBM Z mainframe revenue jumped 48% year-over-year in 2026—the highest growth rate in two decades. Evercore ISI analyst Amit Daryanani noted that clients are sticking with mainframes despite having migration options, suggesting AI hype is outrunning reality in enterprise environments. Banks aren’t avoiding modernization because they’re technologically conservative. They’re avoiding it because the systems work, and the alternatives don’t offer equivalent reliability at scale.
The market heard “AI can translate COBOL” and assumed demand for mainframes would collapse. But translation was never the constraint. The constraint is that no general-purpose cloud infrastructure can match what mainframes do.
AI can translate code, but it can’t replicate the hardware—and that’s where the real cost lives
Here’s the honest limitation: mainframes scale to tens of billions of encrypted transactions per day on a single system. General-purpose cloud servers can’t match that transaction density or reliability. Full replatforming to cloud costs $2 million to $15 million+ upfront, with 50-90% annual infrastructure savings post-migration—but only if you can accept the performance trade-offs.
Most enterprises can’t.
Analyst Steve McDowell frames it clearly: “Applications don’t run on mainframes because they’re written in COBOL. They run [there] because mainframes deliver… reliability that general-purpose servers can’t match.” AI-assisted coding works for prototypes and greenfield projects. Production-grade systems engineering—especially at the scale of global financial infrastructure—remains a human domain requiring deep architectural knowledge.
No documented cases exist of AI successfully completing a full mainframe-to-cloud migration in 2024-2026. That research gap isn’t an oversight. It’s the smoking gun. Anthropic made a claim, the market believed it, but there’s zero evidence it works at enterprise scale.
Rob Thomas again: “The gap between translation and modernization is where most enterprises run into trouble.” Anthropic: “AI can rapidly refactor COBOL codebases.”
One of these statements is backed by $3 trillion in daily transactions. The other triggered a $40 billion panic.
