How Anthropic's Claude Crashed IBM Stock 13% — The COBOL AI Migration Story

The Day AI Crashed a $180 Billion Company

February 23, 2026 will be remembered as one of the most dramatic days in tech stock history. Anthropic published a technical demonstration showing that Claude Code — its AI-powered coding agent — could automatically analyse, understand, and translate legacy COBOL codebases into modern languages like Java, Python, and C#.

The market reaction was swift and brutal.

IBM shares plunged 13% in a single trading session, wiping approximately $31 billion in market value. It was IBM's worst single-day decline since the dot-com crash in 2000. By the end of February, IBM had fallen a total of 27% from its pre-demonstration price.

But IBM was not alone. The shockwave rippled across the entire IT services sector:

• Accenture dropped 9% over the following week.
• Cognizant fell 11%, with analysts highlighting its heavy exposure to legacy modernisation contracts.
• Infosys and Wipro each declined 5-7% on the BSE as Indian markets opened the next day.
• TCS was the most resilient among Indian IT majors, dropping only 3%, partly due to its diversified revenue streams.

Why COBOL Matters So Much

To understand why a coding demonstration crashed stocks, you need to understand the economics of COBOL.

COBOL (Common Business-Oriented Language) was created in 1959. Despite being over 65 years old, it still runs some of the most critical systems on the planet:

• 95% of ATM transactions worldwide still touch COBOL code.
• 80% of in-person transactions at banks run on COBOL-based mainframes.
• The US Social Security Administration, IRS, and most major insurance companies rely on COBOL backends.
• In India, several public sector banks and LIC still run core systems on COBOL mainframes.

The problem? The average COBOL developer is over 60 years old. There are an estimated 800 billion lines of COBOL code in production globally, and the pool of people who can maintain it is shrinking every year.

This created a multi-billion dollar consulting industry focused on COBOL modernisation. IBM's consulting division alone generates an estimated $4-6 billion annually from legacy code migration projects. These are typically multi-year engagements with Fortune 500 clients, involving hundreds of consultants painstakingly rewriting code line by line.

What Anthropic Actually Demonstrated

Anthropic's demonstration was methodical and convincing. Using Claude Code with its agentic capabilities, they showed:

1. Automated Code Analysis: Claude Code ingested a 500,000-line COBOL banking application and produced a comprehensive architecture map, identifying modules, dependencies, data flows, and business logic — a task that typically takes a team of consultants 3-6 months.

2. Intelligent Translation: The AI translated COBOL modules into modern Java with Spring Boot, preserving business logic, edge cases, and even the original commenting style. It handled COBOL-specific constructs like PERFORM VARYING, REDEFINES, and COPY books that have no direct equivalent in modern languages.

3. Test Generation: For each translated module, Claude Code generated comprehensive unit tests and integration tests, along with a validation framework to compare outputs between the original COBOL and the new Java code.

4. Documentation Production: The tool produced detailed migration documentation, including risk assessments, rollback plans, and training guides for operations teams.

Anthropic claimed the entire process — which would typically cost $50-100 million and take 2-4 years with a traditional consulting approach — could be completed in weeks at a fraction of the cost.

The VentureBeat Counterpoint: "Translating COBOL Isn't Modernising"

Not everyone was convinced. VentureBeat published a widely-shared analysis on February 26 arguing that Anthropic's demonstration, while technically impressive, missed the point of legacy modernisation.

The key arguments:

• Translation is not transformation. Converting COBOL to Java line-by-line produces "COBOL written in Java" — you get the same monolithic architecture, the same tightly coupled modules, just in a different syntax. True modernisation involves rearchitecting for microservices, cloud-native deployment, and modern DevOps practices.

• Business logic is the hard part. The real challenge in COBOL migration is not the code translation — it is understanding decades of accumulated business rules, many of which are undocumented and exist only in the code itself. While AI can translate syntax, understanding whether a particular calculation reflects a 1987 regulatory requirement that is still active requires human domain expertise.

• Testing is where projects actually fail. Most COBOL migration projects fail not in the translation phase but in the testing and validation phase. Financial systems require absolute precision — a rounding error in a banking application can cascade into millions of dollars of discrepancies. Automated test generation is helpful but not sufficient for this level of validation.

• Organisational change management — training thousands of mainframe operators to work with modern systems — accounts for 30-40% of migration project costs and cannot be automated by AI.

The debate raged on tech Twitter (now X) for weeks, with reasonable arguments on both sides.

What This Means for Indian IT Companies

The Indian IT services industry has a particularly large stake in this debate. Legacy modernisation is one of the highest-margin, most stable revenue streams for Indian IT majors:

TCS

TCS's Mainframe Modernisation Practice serves over 200 global clients. Their proprietary TCS MasterCraft tooling was already incorporating AI-assisted migration, but the scale of Anthropic's demonstration raises questions about whether clients will prefer an AI-first approach over traditional consulting engagements.

Infosys

Infosys has invested heavily in its Cobalt cloud platform and Topaz AI platform specifically for modernisation workloads. The company's strategy of combining AI tooling with human consultants may actually position it well — if it can move faster than clients building in-house AI migration capabilities.

Wipro and HCL Tech

Both companies derive significant revenue from legacy application management. Wipro was particularly exposed, as it had recently won several large COBOL modernisation deals that are now being re-evaluated by clients.

The Broader Impact

Indian IT companies employ over 5 million people, with an estimated 200,000-300,000 working directly on legacy modernisation and maintenance projects. The fear is not that these jobs disappear overnight — they will not — but that the pricing power shifts dramatically. If AI can do in weeks what takes years, clients will demand lower fees and shorter timelines.

The optimistic view: Indian IT companies that aggressively adopt AI-assisted modernisation tools can increase throughput while maintaining margins, handling more projects with fewer people per project. The pessimistic view: clients bypass consulting firms entirely and use AI tools in-house.

Can AI Really Replace Legacy Code Migration?

The honest answer, as of March 2026, is not yet — but the gap is closing fast.

Here is a realistic assessment:

| Aspect | AI Capability (March 2026) | Human Consulting | |---|---|---| | Code translation (syntax) | Excellent (90%+ accuracy) | Slower but contextually aware | | Business logic understanding | Good for documented logic, weak for tribal knowledge | Strong (domain experts) | | Architecture redesign | Limited — tends to replicate existing patterns | Strong (solution architects) | | Testing and validation | Good for generating tests, weak for edge-case discovery | Essential for financial systems | | Change management | Cannot do this | Required for every project | | Cost | 10-20x cheaper for translation phase | Premium pricing, but covers full lifecycle | | Timeline | Weeks for translation | Years for full modernisation |

The most likely near-term outcome is a hybrid model: AI handles the heavy lifting of code analysis and translation, while human consultants focus on architecture decisions, business logic validation, testing, and change management. This could reduce project costs by 40-60% while maintaining quality.

Lessons for Developers

If you are a developer watching this unfold, here are the key takeaways:

1. Learn AI-assisted development tools now. Whether it is Claude Code, GitHub Copilot, or Cursor — these tools are no longer optional. Explore our guide on Claude Code custom commands to get started.

2. Understand agentic AI workflows. The COBOL demonstration was not just code completion — it was an autonomous agent executing a multi-step workflow. Learn how agentic AI workflows are changing software development.

3. Domain expertise becomes more valuable, not less. As AI handles more of the mechanical coding work, understanding the business domain — banking regulations, insurance calculations, government compliance — becomes the differentiator.

4. Do not panic, but do adapt. The IT services industry has survived multiple disruptions — from the rise of cloud computing to the shift to agile. Companies and individuals who adapt will thrive.

What Happens Next?

As of March 2026, IBM has partially recovered, trading about 18% below its pre-demonstration peak. The company has announced its own "watsonx Code Moderniser" tool, essentially acknowledging that AI-assisted migration is the future and trying to stay relevant.

Meanwhile, Anthropic has announced partnerships with two major US banks to pilot Claude Code for production COBOL migration — the first real-world test of whether the demonstration translates into practical, reliable results.

The COBOL story is far from over. But one thing is clear: AI has permanently changed the economics of legacy code modernisation, and the ripple effects will be felt across the global IT industry for years to come.

Want to stay ahead of the AI curve? Browse our AI skills marketplace for tools and workflows that keep you competitive, and check out our guide on agentic AI workflows to understand the technology behind these disruptions.