Japan's Mainframe Exodus Hits a Wall, and AI Isn't the Answer

The Retreat Is Real

Fujitsu has already announced it is exiting the mainframe business. Hitachi is now following the same path, confirming it will phase out VOS3 — its Virtual-storage Operating System 3 — with sales ending in November 2027 and full maintenance support winding down by December 2034.

This is not a quiet exit. It is two of Japan’s largest IT giants openly admitting that decades of mainframe investment can no longer be sustained on their terms.

The implications extend far beyond product lifecycle management. Mainframe infrastructure has long served as the backbone of Japanese finance, manufacturing, and public-sector operations. Fujitsu and Hitachi built entire divisions around these systems — engineering teams, consulting practices, training programs, and partner ecosystems. Their retreat signals a structural shift: the vendors who spent fifty years selling mainframe assurance are now selling an exit strategy.

The timing matters. Generative AI has been sold to enterprises as the key that unlocks legacy modernization — automating code translation, reverse-engineering undocumented systems, accelerating migration timelines. The pitch has been compelling: let AI handle the hard parts, and companies can finally shed their mainframe dependencies.

Gartner says that pitch is dangerously premature.

The 70 Percent Failure Rate

In a report published June 18, 2026, Gartner predicted that more than 70 percent of mainframe offloading projects initiated in 2026 will fail to achieve their original objectives. The finding is not about technology failing — it is about expectations misaligned with reality.

The core gap, according to Gartner, lies between the marketing claim that generative AI dramatically reduces modernization effort and the actual capability of current AI systems to handle the work. Vendors, under investor pressure, are inserting AI features into their products whether or not those features deliver meaningful results. The result is a wave of migration projects launched with insufficient planning and inflated confidence in what AI can do.

Mainframe applications are not simply old code waiting to be rewritten. They are mission-critical systems shaped by years of incremental modification. The logic buried inside them is often incomprehensible to anyone who did not build it. Change one function and you may not know which downstream process will break. Design documents rarely match the actual system. The engineers who understand these platforms are retiring in large numbers.

AI can translate syntax. It cannot reconstruct institutional memory.

What makes the failure rate so striking is that many of these projects were not abandoned mid-execution. They produced working systems. But those systems did not deliver the promised outcomes — throughput improvements, cost reductions, or operational flexibility. The code ran, but the business case dissolved.

What the Data Shows

ROUTE06, which provides AI-driven development platforms, conducted a survey in April 2026 of 328 managers and engineers involved in core system planning and modernization. The results confirm that the obstacle is not a simple shortage of COBOL programmers.

Among manufacturing companies, 46.7 percent said their modernization efforts remain stuck at the conception stage. Among enterprises served by SIers and IT vendors, that figure rises to 54.7 percent. More than half of Japan’s large corporations have not moved past the planning phase.

Of the companies that did attempt modernization, 70.2 percent reported overall success. But 31.4 percent said only partial success was achieved, with significant issues remaining. The most common failures were schedule overruns caused by system complexity, project stagnation due to institutional knowledge concentrated in a few people, and cost overruns driven by staffing shortages.

Some projects degraded to nothing more than language translation. Several organizations found themselves forced to maintain legacy and modern systems side by side — defeating the purpose of migration entirely.

The reasons for hesitation are revealing. Manufacturing companies cite unclear return on investment (51.9 percent) and inability to estimate migration costs or timelines (50.0 percent). SIers and IT vendors worry about business interruption risk (63.9 percent) and uncertainty around the accuracy and quality of AI-assisted code conversion (52.5 percent).

These numbers point to a deeper problem: modernization is being treated as an engineering task when it is actually a knowledge management problem. The code is the easy part. Understanding what the code does, why it does it that way, and what happens if you change it — that requires institutional knowledge that no translation tool can capture.

Second-Order Effects Already Emerge

The vendor retreat is producing consequences that extend well beyond the companies directly affected. As Fujitsu and Hitachi pull back from mainframe services, they are taking with them specialized training programs, certification pipelines, and internal expertise libraries that smaller organizations depend on.

Consulting firms that built practices around mainframe migration now face a shrinking addressable market at one end and rising demand at the other — clients who need help precisely because early migration attempts faltered. This dynamic is already visible in Tokyo’s enterprise consulting sector, where firms report a growing number of engagements focused on rescue and remediation rather than greenfield modernization.

The retirement of mainframe-era engineers is accelerating the knowledge gap. Many of these individuals held positions that were never formally documented. Their understanding of system interdependencies, business rules embedded in logic, and workarounds developed over decades exists only in their heads. As they leave, they take irretrievable context with them. AI tools can parse code, but they cannot interview a retired systems architect who spent thirty years at a bank.

There is also a procurement effect. Organizations that committed to mainframe contracts through 2030 or beyond now face uncertainty about continued vendor support, pricing, and service levels. Finance teams are adjusting三年-year budget models. Risk officers are reassessing contingency plans. The withdrawal creates ripple effects through supply chains that depend on mainframe-stable data feeds.

The Non-Obvious Takeaway

The prevailing narrative in enterprise tech has been that AI solves the legacy problem. The data suggests the opposite risk: AI enthusiasm is accelerating poorly planned migrations that were already likely to fail.

Gartner’s recommendation cuts against the marketing grain. The firm advises organizations to evaluate each workload individually and assign it to the most appropriate execution platform — not to assume every mainframe application must move. For mid-sized mainframe environments, Gartner recommends modernizing within the existing platform wherever possible and reserving full platform transitions for cases where they are genuinely necessary.

Generative AI, in this view, is most effective as a tool for maintaining and incrementally modernizing mainframe systems — not as a sledgehammer for emergency migration.

The companies most at risk are those that treat AI code conversion as a substitute for architectural understanding. They are the ones launching projects without adequate planning, hoping the tool will compensate for gaps in knowledge that no tool can fill.

Success stories, when they exist, share a common pattern: they invest heavily in discovery before writing a single line of new code. They map data flows, document business rules, identify dependencies, and build internal capability before attempting large-scale migration. AI assists the process, but it does not replace the disciplined work of understanding what is being moved and why.

What Happens Next

Hitachi’s VOS3 phase-out creates a concrete deadline for enterprises still running that operating system. Sales end in November 2027. Maintenance support continues until December 2034 — but support without updates is not the same as viability. Organizations on VOS3 need migration strategies now, not after the announcement.

Fujitsu and Hitachi’s retreat from mainframe business will compress the ecosystem of vendors, consultants, and specialized engineers who supported these systems. That scarcity will drive up the cost of remaining expertise and slow migration projects further.

The 70 percent failure prediction is a forecast, not a law. But it is grounded in observable patterns: vendors overselling AI capability, enterprises underinvesting in discovery and planning, and a retiring workforce taking irreplaceable institutional knowledge with them.

The lesson for Japanese corporations — and for any enterprise grappling with legacy modernization — is straightforward. AI is a tool, not a strategy. Migration projects that succeed are the ones that treat modernization as an architectural problem first and a technology problem second. The companies that skip that step are the ones Gartner says will join the 70 percent. The vendors walking away are not leaving because the work is impossible. They are leaving because the economics of sustaining mainframe ecosystems no longer hold — and the AI-driven migration boom has not delivered the results its promoters promised.