We (briefly) interrupt the relentless flow of AI industry announcements and consequent market volatility for a moment of retrenchment.

Yesterday, Anthropic announced COBOL support for Claude. For those unfamiliar, COBOL is a programming language as legacy as it gets: decades old, but still in active use because it is one of the primary languages running inside mainframes.

Mainframes are legacy hardware also around for decades and still integral to enterprise function in many industries. They have long governed financial services, retail, government backend systems, and other critical infrastructure. Because they are so old, they predate the internet. They were built to be self-contained. They minimize exposure. They are engineered for 99.999% uptime and achieve it reliably. Even during the mass migration to cloud during the 2010s, core mainframe transaction workloads were largely exempted, because cloud vendors could not demonstrate the same level of reliability. IBM was the original manufacturer, still dominates, in fact is effectively the only major manufacturer of enterprise mainframe hardware in the traditional sense, and still makes a significant percentage of its revenue from IBM Z (formerly System/360, 370, 390 and zSeries) although the amounts are no longer broken out, from mainframes.

IBM fell sharply in the wake of the announcement; 14% for the day. Cognizant, with roughly $19–20 billion in annual revenue and significant exposure to legacy modernization work, has also traded under pressure in recent sessions, and Kyndryl, the $15 billion infrastructure services firm spun out of IBM, similarly weakened. That reaction is irrational. The global mainframe modernization services market was roughly $7.9 billion in 2024 and is projected to grow to as much as $18 billion+ by 2033. 

The common thread is clear: investors are repricing companies associated with mainframe and legacy services on the assumption that AI-assisted code comprehension erodes their economic moat. My view is that this is a misread. Lowering the cost of understanding COBOL reduces consulting margins at the interpretive layer; it does not eliminate the infrastructure those systems run on, nor does it meaningfully reduce demand for in-place modernization.

The market reaction to the broader orchestration story (which I have documented in recent coverage) is reasonable. AI models demonstrating credible control across productivity suites, CRM systems, and enterprise tooling represent a genuine platform shift. But there is a moat around mainframes that is almost impregnable, and this particular announcement may actually strengthen IBM’s position rather than weaken it. To understand why, you need to understand how the mainframe actually works.

The Stack That Markets Are Ignoring

Mainframes are layered systems. At the base sits specialized hardware optimized for deterministic throughput and fault tolerance. Above that are hypervisors and logical partitions. Then the operating system layer (most commonly z/OS in large enterprises). Then the transaction engines and middleware — CICS, IMS, DB2 —that guarantee ACID compliance and concurrency control at massive scale.
Only at the very top, at the application layer, do we encounter COBOL business logic. COBOL encodes rules. The mainframe enforces integrity.

When a model reads COBOL, it touches the application layer. It does not displace the transaction layer. It does not replace the operating system. It does not eliminate the hardware architecture. It does not unwind regulatory entanglement or liability structures.

And critically, mainframes are rarely internet-facing. They sit deep inside segmented enterprise networks, insulated behind middleware, API gateways, and controlled identity systems. Most modern breaches occur at the edge (cloud connectors, web layers, SaaS misconfigurations) not inside core transaction engines.

This insulation is not accidental. It is architectural.
Mainframes exist to process payroll, clear settlements, execute government disbursements, and maintain financial ledgers with near-zero tolerance for error. Their value proposition is reliability under constraint, not developer ergonomics. Breakdowns when they occur are often a failure of middleware configuration or communication between layers.

The Orchestration Story Is Real — But Mainframes Are the Exception

Over the past several weeks, the most consequential AI demonstrations have not been about code generation. They have been about control. Models showing credible orchestration across Windows, macOS, productivity suites, CRM systems, legal platforms, and security tooling signal a shift in where value accumulates in the enterprise stack.
If an AI layer can coordinate workflows across tools, execute multi-step processes, and operate applications at scale, then economic gravity begins to move upward. Historically, when a control layer emerges (operating systems, browsers, cloud platforms) value accrues there. Individual tools become composable components rather than primary interfaces.

That is a platform shift. And it makes valuation compression for SaaS companies competing on interaction-layer dominance economically rational.
Mainframes, however, do not compete on interaction-layer dominance. They compete on transaction integrity and regulatory trust. That makes them structurally different from the productivity platforms vulnerable to orchestration-layer abstraction.

Why This Likely Strengthens IBM

The contrarian read that markets are missing: AI-assisted COBOL comprehension does not threaten mainframe incumbents. It reduces their operational costs.

Today, enterprises pay enormous consulting fees, primarily to firms like Accenture, Cognizant, Deloitte, Kyndryl, and specialized legacy shops, for what amounts to code archaeology and updating. Understanding what decades-old COBOL programs actually do, documenting business logic buried in millions of lines of procedural code, and planning modernization pathways around that logic. These engagements are expensive and slow because interpretive scarcity drives the economics.

Claude reading COBOL compresses those consulting margins. It accelerates analysis, documentation, and modernization planning. But the critical question is: modernization to where?

The answer, in most cases, is modernization in place. Enterprises running settlement engines, payroll systems, and regulatory reporting on z/OS are far more likely to evolve within the mainframe ecosystem than to rip out core transaction infrastructure and migrate to platforms that cannot yet match mainframe reliability guarantees. Faster code comprehension reduces friction for in-place evolution (refactoring, optimizing, and extending existing systems) which keeps workloads on IBM hardware.
The consulting intermediaries lose margin. IBM’s platform becomes easier and cheaper to maintain. The moat does not shrink. The cost of defending it does.

The Distinction That Matters

Markets sometimes conflate “AI can read legacy code” with “legacy infrastructure is dead.” That is an analytical shortcut. Mainframes are not fragile relics propped up by obscure syntax. They are deeply embedded transaction platforms insulated by architecture, governance, and switching costs measured in decades.

Claude reading COBOL changes the economics of interpretation. It does not change the physics of enterprise core infrastructure. And in a stack increasingly defined by orchestration, that distinction matters.

Consulting rates tied to digging through old code will probably come down. The sky-high premiums for COBOL programmers, driven for years by a shrinking talent pool and institutional memory locked inside aging systems, are also likely to ease. That’s a natural result of tools that make understanding legacy systems faster and easier. But cheaper interpretation does not mean the infrastructure itself is suddenly vulnerable. IBM’s strength is not that COBOL is hard to read. It’s that the systems running on its mainframes clear payroll, settle trades, and manage government ledgers with extraordinary reliability. Could that position change someday? Of course. Every technology eventually faces a successor. But nothing about this announcement suggests that shift is imminent. Claude reading COBOL changes how quickly humans can understand old code. It does not undo the foundations those systems run on.