Frontier models reach COBOL, and the mainframe modernisation business gets repriced

This space is crowded, but the gap is real and narrow. AI translation is becoming a commodity: Claude Code, AWS Transform and IBM watsonx all do it. Proving equivalence is the bottleneck everyone names, and well-funded startups already claim it: Mechanical Orchard, CloudFrame Continuum (which comes with a guarantee) and Hypercubic. On the testing side, AWS Application Testing and Broadcom Test4z already record and replay mainframe runs. The equivalence-proof migration firm would therefore compete head-on with players that have more capital and more references. A newcomer's best chance is the part none of these vendors appear to have nailed: evidence packages built specifically for regulators and auditors. That means documentation that maps to OCC, FFIEC, NAIC or state audit expectations, independent attestation, and sign-off workflows. It could be sold as an auditor-neutral service to regional banks, credit unions, insurers and state agencies, which are too small for the big SIs and wary of a single cloud vendor. For the test-harness generator, the most defensible angle is to be vendor-neutral and to generate tests from logs and job history, such as SMF records, JCL run histories and dataset snapshots. It would sit before any migration decision and sell to SIs that don't want to be locked into AWS or Broadcom. I found no source describing a product that builds regression suites purely from production batch logs, but cloud vendors could bundle this cheaply, so it may work better as a feature or an acquisition target than as a standalone company. Outcome pricing on its own is not a differentiator, because CloudFrame and some SIs already offer guarantees or gain-share deals.

• Mechanical Orchard (Imogen)

  VC-backed startup from 2022 that is reported to have raised a $24M Series A and a $50M Series B (I did not confirm the funding). Imogen watches the running mainframe, records real data flows, and uses that behaviour as the spec for generating code and continuously proving old and new systems match. This is very close to idea 1.

• CloudFrame Continuum

  Launched July 2026. It records runtime behaviour before converting COBOL and backs the result with a commercial guarantee of functional equivalence. This is effectively the outcome-backed equivalence pitch in idea 1.

• Hypercubic HyperLoop

  Startup with a verification-first approach. It turns COBOL into Dafny and then Go, compares the new app against a simulation twin and against the real mainframe, and records any accepted differences as documented findings. This overlaps with both ideas.

• AWS Transform for mainframe

  Announced December 2025. It generates test plans, test data collection scripts and test automation scripts, and it has AI agents for functional equivalence testing. It also absorbed Blu Age's COBOL-to-Java refactoring. Hyperscaler competition for both ideas.

• AWS Mainframe Modernization Application Testing

  Records test runs on the source mainframe, replays them against the modernised app and compares the outputs. Saved scenarios can be reused for regression runs. This is the record-and-replay core of idea 2.

• Broadcom Test4z

  Generally available since late 2023. It records COBOL batch program runs and replays them for regression testing, with data sets, Db2 and program calls virtualised. It is the incumbent mainframe testing vendor for idea 2.

• IBM watsonx Code Assistant for Z

  IBM's own AI tool for converting COBOL to Java, sold alongside its mainframe estate. It is an incumbent with direct access to the customers.

• Anthropic Claude Code (COBOL modernisation)

  Anthropic published a February 2026 post on using Claude Code to modernise COBOL, and IBM's stock had its worst day in about 25 years. This shows frontier labs are targeting the analysis and translation layer directly.

• Accenture, Cognizant and other large SIs

  Large integrators with existing bank and government modernisation contracts. They are now adding AI tooling, and some offer outcome-based or gain-share pricing.

Checked 2026-10-20