Hypercubic's $5.3M bet: the hard part of COBOL migration is reading the code, not writing it

Hypercubic Inc. has raised $5.3 million in seed funding to point AI agents at COBOL modernisation, in a round led by CIV with participation from Y Combinator, Afore Capital, Pioneer Fund, Multimodal Investors and angels including Opendoor chief executive Kaz Nejatian and Walmart Labs co-founder Venky Harinarayan ^[1][2]. The interesting part is the sequencing rather than the size: the company, started by former Apple engineers Sai Gurrapu and Aayush Narik, treats reading and documenting a mainframe codebase as the main body of work, with the rewrite falling out of it ^[3][11].

The demand side is well rehearsed. COBOL was first written in 1959 for processing business and financial data, and an estimated 200 billion-plus lines still support more than 95% of the world's swipe transactions and ATMs ^[4][5]. One estimate cited by SiliconANGLE holds that 90% of current COBOL engineers will retire within the next five to 10 years ^[6]. The usual answer is a rewrite into something like Java, which younger engineers know ^[17], but the risk profile is why it keeps getting deferred: a single undocumented change in a monolithic COBOL program can take down a bank ledger or a government social security payments system and take days to trace and fix ^[7]. The code has been compared to asbestos, embedded in vertically integrated mainframe stacks such that replacement pulls in data architecture, storage and transaction processing ^[9]. Manual rewrites run for years, and most organisations postpone them on cost and risk grounds ^[8].

Hypercubic's agents run on models trained on large volumes of COBOL and are meant to map an entire codebase first, recovering the business logic buried in it and generating documentation that in most cases was never written ^[10][11]. "The problem is not simply translating COBOL syntax," Gurrapu said. "The real challenge is recovering decades of hidden business logic, understanding how systems behave in production, and proving that a modern replacement preserves what matters" ^[12]. The company says that documentation then makes it easier to generate modern code that replicates the original application, including its historical data ^[13], and that its agents can modernise a legacy app in a couple of months against years for manual transformation ^[14]. That timeline is Gurrapu's own claim, not an audited result.

The gap in the account is the third clause of his own sentence. Recovering logic and generating documentation are artifacts a customer can inspect; proving that the replacement preserves behaviour is what a bank's risk committee will actually gate on, and the source material does not describe how the agents establish that equivalence ^[12]. Two deployments are cited without names: one of the largest banks in Latin America and a leading Caribbean retailer serving more than 50 million customers a year, both on critical applications ^[15]. No line counts, cutover dates or parallel-run results are given. Nor is the scale mismatch small: spread across the estimated 200 billion lines in service, $5.3 million is roughly 2.7 cents per thousand lines ^[18].

Watch whether the documentation layer gets priced and sold on its own, since an accurate map of an undocumented ledger has value to a bank that has no intention of migrating this decade. Watch for a named reference customer that has cut over in production rather than run a pilot. And watch what the new money buys: the company says it goes to agent capabilities and capacity for more enterprise customers, which is where a services-heavy migration business usually starts to show its true margins ^[16].