Inside the Mainframe: Where AI Meets COBOL
For decades, the mainframe has been a black box: systems written in COBOL through the 1970s and 1980s still compile and run on modern IBM z/OS hardware with minimal changes, powering trillions in global transactions every day. The people who understood those systems — the ones who knew which JCL parameters to tweak, how to read a return code, or why a particular copybook layout mattered — are retiring faster than anyone can hire replacements. What remains is an enormous, opaque codebase that nobody fully understands anymore, and a growing fear that the systems running global finance, airlines, and government will become unmaintainable.
That fear is now driving a new class of AI-native tooling built specifically for the mainframe environment. Hypercubic, founded in 2025 by ex-Apple engineers and backed by $500K in pre-seed funding from Y Combinator's Fall 2025 batch, positions itself as the first platform designed to operate inside the mainframe rather than alongside it. Unlike general-purpose AI coding assistants that assume access to Git repositories, file systems, and CI pipelines, Hypercubic's Hopper environment connects directly to a real TN3270 terminal and presents mainframe-aware panels for datasets, job output, and CICS regions. Its AI agent can navigate ISPF, inspect VSAM files, write and edit COBOL, generate JCL, submit jobs, parse JES spool output, and query CICS transactions, all within the z/OS environment itself.
The company launched an early version of its tool on Hacker News in November 2025, where it drew attention from the developer community for tackling a problem most AI tooling ignores. Standard AI coding tools treat legacy systems as static repositories, missing the reality that mainframe work lives across terminal screens, datasets, copybooks, and decades of operational knowledge that only experienced operators carry. Hypercubic's approach, as described on its website, is to capture both the codebase and the expertise of retiring engineers, turning undocumented code into a searchable system of record. Its HyperLoop system promises behavioral-equivalence verification, proving that migrated or refactored code behaves exactly like the original, which the company frames as the missing piece between today's AI coding tools and mission-critical infrastructure.
IBM has taken a different path. The company's z17 mainframe, announced in April 2025, is the first generation fully engineered with AI capabilities across hardware, software, and systems operations. Built on the new IBM Telum II processor and IBM Spyre accelerator unveiled in August 2024, the z17 supports AI inferencing directly on the system, enabling real-time transaction scoring without moving data off-platform. IBM's framing is not about replacing the mainframe but integrating it into the enterprise AI stack. As of that announcement, 88% of global IT executives surveyed by IBM said application modernization was crucial to their digital transformation plans, and 78% affirmed that mainframes would remain central to those efforts. The z17 represents a hardware-level bet that the mainframe will coexist with AI rather than be replaced by it.
AWS has staked its claim in the middle ground. AWS Transform for mainframe, described as an agentic AI-powered service, accelerates the migration of monolithic COBOL applications into modern architectures like microservices. The service uses automated code analysis and AI-assisted refactoring to reduce the time, cost, and risk of mainframe modernization, with partners like Royal Cyber enabling enterprises to move COBOL applications to Java. Unlike Hypercubic's inside-the-mainframe approach or IBM's hardware-integrated AI, AWS Transform treats the mainframe as a source to be transformed rather than preserved.
Together, these tools reflect a fundamental shift: AI is no longer a peripheral aid to mainframe modernization but a core component of how enterprises approach their most critical legacy systems. The question is no longer whether these systems can be modernized, but how quickly and at what cost to the workforce that has kept them running for decades.
The Vanishing Experts: A Workforce Crisis Unfolds
The mainframe workforce is aging out faster than anyone can replace it. The average active COBOL developer is 58 years old, according to a 2024 Micro Focus/Rocket Software survey, and 75% of COBOL developers plan to retire by 2035. At the current rate (roughly 10% of the workforce exits every year) the global pool of active COBOL developers has already shrunk from over 300,000 in the early 2000s to approximately 50,000 today. Fewer than 500 new COBOL developers enter the North American workforce annually, creating a gap that grows wider with each passing year.
This exodus carries more than résumés. When a senior COBOL developer retires, three things leave with them: institutional knowledge about how systems actually behave, exception-handling logic that exists only in muscle memory, and the historical context behind decades of design decisions. A 2023 Deloitte survey found that 68% of organizations with mainframe systems report that "significant business logic exists only in the minds of current developers."
The problem compounds because COBOL systems were never built for easy handoffs. Many were written 30 to 40 years ago, when documentation standards were minimal. Systems have been modified thousands of times since, but documentation was rarely updated. Experienced developers often struggle to articulate knowledge they've internalized over decades — they know what to do instinctively, but explaining why and how is a different skill entirely. Many are also focused on finishing their careers, not training replacements.
Universities stopped teaching COBOL in droves starting in the 1990s, when the language was widely considered "dying." Over 85% of U.S. computer science programs dropped it from their curricula. Fewer than 20 universities still offer COBOL courses, down from 200+ in the 1990s. Students followed the market toward web development, mobile apps, and more recently AI engineering — COBOL doesn't appear on any "hot tech" list, and a 25-year-old developer who invests two years learning it is betting on a shrinking market.
But market forces are starting to push back. COBOL developers now earn an average of $125,525 per year in the United States, with senior consultants commanding $200–350 per hour — up from $100–150 five years ago. Some consulting firms have waitlists stretching six months or longer. The Bureau of Labor Statistics projects that software developer roles broadly will grow 15% over the next decade, and within that category, mainframe specialists occupy a niche where demand consistently outstrips supply. As of March 2026, Indeed, Dice, and other job boards show hundreds of open COBOL positions across banking, government, insurance, and healthcare.
Organizations are responding with a mix of retention efforts and new training pipelines. Coursera, IBM, and several community colleges now offer COBOL training programs. Some companies are implementing retention bonuses, part-time consulting arrangements, or phased retirement plans to keep key developers engaged through early modernization stages. But the real shift is toward hybrid roles — professionals who understand COBOL well enough to manage migration projects, train AI tools on legacy codebases, or consult on modernization strategy. These bridge roles often command even higher salaries than pure development positions, because they require both deep technical knowledge and the ability to navigate complex organizational dynamics.
The tension here is clear: AI tools promise to automate many aspects of COBOL modernization, but the research consistently shows that human oversight remains essential. Organizations need developers who understand legacy systems to validate translated code, ensure nothing breaks during migration, and preserve business logic that exists only in retiring experts' minds. AI can extract knowledge from code and accelerate timelines, but it cannot yet replace the judgment of someone who lived through decades of system evolution. The workforce impact isn't just about replacing retiring talent — it's about redefining what that talent looks like in an AI-assisted world.
Banks and Airlines Rewire Talent Pipelines Around AI
JPMorgan Chase's approach to AI integration illustrates how the largest financial institutions are rewiring their talent pipelines around mainframe modernization rather than simply replacing legacy systems. The bank's chief data analytics officer, Derek Waldron, described the transformation as a "fundamental rewiring" for the AI era, with ambitions to provide every employee an AI agent and automate every behind-the-scenes process. That scope covers the bank's 317,000-person workforce — about 250,000 of whom now have access to the firm's internal LLM Suite platform, with roughly half using it daily.
The productivity gains from AI coding assistants have reached 10–20% across the software development lifecycle, according to the bank's 2024 shareholder letter. That improvement matters because JPMorgan employs more than 60,000 technologists globally and operates a technology estate spanning over 6,000 applications and nearly an exabyte of data. The consumer banking division signaled in May 2025 that operations staff would fall by at least 10% over the next five years as AI deployment accelerates — a reduction that translates to thousands of roles. The Commercial & Investment Bank alone runs more than 175 AI use cases in production, including know-your-customer processes that have cut unit costs by nearly 40%.
Doug Petno and Troy Rohrbaugh, named co-presidents in June 2026, now steer the bank's technology strategy. Their mandate includes balancing first-mover advantage against the value gap between AI's potential and enterprises' ability to capture it. As Waldron noted, companies "do work in thousands of different applications, there's a lot of work to connect those applications into an AI ecosystem and make them consumable." That integration work demands engineers who can read COBOL logic and translate it into AI-ready services — a skill set that did not exist in meaningful numbers a decade ago.
American Airlines and Delta Air Lines have followed similar paths, though with less public disclosure. Both carriers outsource portions of their reservation and crew-management systems to IBM Z mainframes while simultaneously layering cloud-native booking engines on top. IBM's 2025 survey found that 78% of global IT executives consider mainframes central to digital transformation, and 88% view application modernization as crucial. The tension here is real: airlines need to retain COBOL expertise for flight-operations systems that regulators will not allow to fail, while simultaneously building teams fluent in AI-assisted refactoring tools.
Insurance presents the starkest hiring shift. Allstate and State Farm have partnered with vendors like Amdocs to automate claims-processing workflows that previously required deep mainframe knowledge. Amdocs' automated mainframe modernization program helped one of the world's largest insurers reduce costs and unlock cloud agility, according to the vendor's case materials. But those savings come with a personnel cost: traditional COBOL programmer roles are declining even as demand grows for engineers who can train AI tools on legacy codebases, manage migration projects, or consult on modernization strategy.
The hiring math is shifting toward hybrid profiles. Enterprises now seek professionals who understand COBOL well enough to guide AI tools through refactoring but can also write cloud-native code that interfaces with modern APIs. That bridge role commands premium compensation — far above what pure COBOL programmers commanded at the peak of their market in the 1990s. The board data from semiconductor and fintech firms confirms this premium:
| Company | Median Salary Band | Number of Salaried Roles |
|---|---|---|
| ASML | $164,000 | 31 |
| Stripe | $235,000 | 19 |
Both companies are aggressively hiring engineers with AI and infrastructure overlap skills.
Regulatory scrutiny is beginning to shape these hiring decisions. Financial services teams must now prove to auditors and risk committees that system behavior, data handling, and controls remain intact after AI-driven changes. That requirement favors candidates who can document migration logic and explain how AI tools interpret legacy business rules — a skill that pure COBOL programmers historically never needed.
Compliance Catches Up With AI-Powered Modernization
On June 27, 2024, the Financial Industry Regulatory Authority (FINRA) issued Regulatory Notice 24-09, formally reminding member firms that existing supervisory obligations apply to their use of generative artificial intelligence and large language models. The notice landed as financial institutions began deploying AI-driven tools to modernize COBOL-based mainframes, a shift that introduces new attack surfaces into systems that process 95% of ATM transactions and run 43% of banking infrastructure. Regulators are not debating the utility of these tools — they are demanding that firms prove they can manage the risks before production deployment.
FINRA's guidance applies whether firms develop AI tools internally or license them from third parties, including platforms like Hypercubic, IBM z17, and AWS Transform that automate COBOL analysis and refactoring. The rules implicate virtually every area of a firm's regulatory obligations, depending on how the technology is deployed. For firms using generative AI as part of supervisory systems (reviewing electronic correspondence, for instance) policies and procedures must address technology governance, model risk management, data privacy and integrity, and the reliability and accuracy of the AI model. Each respective application of generative AI needs to be comprehensively evaluated and tested before production, according to compliance firm 17a-4.
The risks are not theoretical. Generative AI technology has been marked by concerns about accuracy, privacy, bias, intellectual property, and possible exploitation by threat actors. When a chatbot session becomes a recordkeeping object under Rule 4511, bot transcripts must export into the firm's books-and-records system. AI-generated client letters and market commentary must flow through existing Rule 2210 pre-publication review. Firms remain responsible for all communications, including AI-generated ones, meaning hallucinated, misleading, or non-compliant AI output remains the firm's regulatory liability the moment it reaches a client.
Risk teams are responding with a responsible, risk-based approach to experimentation and introduction. The critical steps include adjusting existing risk and control frameworks for potential generative AI threats, ensuring clear governance aligned with FINRA regulatory guidance, and educating firm employees on generative AI fundamentals. As one compliance advisory noted, implementations should be governed by an AI committee consisting of IT, compliance, and outside consultants.
The supervisory burden extends across five key domains. First, firms must maintain a written AI evaluation framework documenting how tools are assessed before deployment. Second, they need an AI inventory tracking every tool, every use case, and every supervisory owner. Third, communication supervision under Rule 2210 must specify where AI-generated content enters the supervisory review queue, who approves it, and what audit trail exists. Fourth, chatbot supervision under Rule 3110 requires retention plans, sampling methodologies, and prohibited-content screens. Fifth, recordkeeping under Rule 4511 mandates that AI prompts and responses are retained per the rule's requirements. Additional controls cover Reg BI suitability and best-interest analysis where AI assists in recommendations, and vendor due diligence aligned with Reg S-P for AI service providers.
Despite the urgency, many firms are falling behind. FINRA's 2026 Annual Regulatory Oversight Report warned that member firms' use of generative AI is outpacing the controls, documentation, and supervisory frameworks needed to manage technology risks. Examiners are now testing for documented supervisory procedures specifically, and the most common use case under scrutiny is summarization and information extraction — exactly the function that AI-driven mainframe modernization tools perform when analyzing millions of lines of undocumented COBOL code.
FINRA will consider providing further guidance or proposing amendments to its rules as appropriate, but firms should continue to carefully monitor the developing regulatory landscape. Other federal and state laws may also apply to AI use, and member firms are encouraged to engage their Risk Monitoring Analysts as AI-related business changes arise. The message is clear: modernization accelerates, but compliance does not bend.
Market Momentum: Growth Projections Through 2029
The mainframe modernization market is entering its fastest growth phase yet, with analyst data pointing to a convergence of AI-driven tooling, hybrid cloud adoption, and workforce transformation that will reshape enterprise infrastructure decisions through 2029.
Kyndryl's 2024 survey of 500 senior business and IT leaders found that 86% are either deploying or planning to deploy generative AI within their mainframe environments, with a third already treating the platform as a foundation for AI-enabled workloads. That rapid uptake has translated into measurable returns: respondents reported one-year ROIs ranging from 114% to 225% on modernization initiatives, alongside collective annual savings of $11.9 billion. The same survey showed that 96% of organizations are migrating workloads off mainframes, averaging 36% of their total workload portfolio to cloud platforms.
Security concerns now top the list of drivers for modernization investments. Two-thirds of survey respondents identified security as the most crucial mainframe feature, while nearly all flagged it as the key factor influencing modernization decisions. Yet skills shortages persist: more than a quarter of organizations report the talent gap as a barrier, particularly in AI and cybersecurity specialties.
External providers are absorbing much of the implementation burden. Kyndryl's data shows 77% of organizations rely on third-party vendors to execute mainframe modernization projects, and 92% consider enterprise-wide observability essential — even as 85% struggle to achieve it across hybrid environments.
Gartner's AI service operations forecast for 2026–2029 projects up to a 50% cost reduction by 2029 as enterprise AI tools accelerate legacy modernization. HyperFRAME Research's first-half 2026 Lens data tempers that optimism slightly: only 14% of organizations classify their core data architecture as fully modernized for AI workloads, revealing a gap between ambition and infrastructure readiness. Their broader 2024 survey of over 1,000 global mainframe professionals confirms a generational workforce shift and widespread adoption of generative AI and hybrid cloud technologies.
The cost-benefit case is strengthening. As Gartner notes, the real breakthrough in 2026-era modernization tools is not code generation itself but proving that refactored systems behave exactly like their originals — a capability that reduces risk and justifies larger-scale migrations.
Still, the talent pipeline remains the bottleneck. Zero G Talent's live board data shows demand for AI-literate systems engineers and mainframe architects accelerating faster than traditional COBOL roles. ASML posted 48 new roles in seven days, with salary bands spanning $237,000 to $355,500 for senior engineering positions. Stripe added 57 roles in the same window, with engineering managers and data scientists commanding median compensation above $235,000.
The analyst consensus points to a bifurcated market: enterprises that can staff AI-mainframe hybrid teams will capture most of the projected cost savings, while those dependent on aging COBOL specialists without upskilling programs risk falling behind. The next three years will determine whether that workforce transition happens fast enough to meet infrastructure demand.
Working in frontier tech? Zero G Talent tracks the openings: see every open ASML role, browse frontier tech jobs, openings at Stripe, and the people building the field.