Code Metal raises $125m to rewrite US defense software with AI

Defense tech has discovered a new way to sell the Pentagon what it already owns: its own legacy code, re-packaged as “modernization” and translated by a model.

Code Metal, a startup pitching AI-assisted software “rewrites” for the defense industry, has raised $125 million in a Series B round, according to Wired. The company’s core promise is that large, brittle codebases—often written decades ago in languages like Ada or C/C++ and surrounded by tribal knowledge—can be converted into cleaner, more maintainable systems using AI tooling. In a sector where paperwork is a deliverable and “requirements” are a genre of fiction, the pitch is irresistible: faster refactors, fewer engineers, and a procurement story that sounds like innovation.

The hard part is not generating code. It’s proving that the generated code does exactly what the old code did, across edge cases nobody remembers and test suites that never existed. Defense software is full of systems where correctness is not a nice-to-have: avionics, fire-control, logistics, comms, cyber tooling, and the glue code that keeps classified networks from collapsing under their own bureaucracy.

Wired notes that Code Metal emphasizes verification and assurance. That’s the only plausible angle; the Department of Defense doesn’t merely buy software, it buys liability narratives. Yet AI translation shifts the liability question from “who wrote the bug” to “who validated the behavior.” If a model-derived rewrite introduces a subtle timing issue, numeric instability, or concurrency bug that only appears under operational load, the post-mortem will not be a Git blame—it will be a contract dispute.

There’s also a quieter transformation: vendor lock-in evolves from proprietary source code to proprietary process. If the rewrite depends on a particular model, prompt strategy, fine-tuning data, and internal tooling, then “owning the code” may not mean owning the ability to maintain it. The Pentagon can demand escrow for source; it cannot easily escrow a model’s evolving behavior, the training pipeline, or the tacit promptcraft that made the translation work.

In theory, this is where open standards and reproducible builds should shine. In practice, defense procurement rewards incumbency, compliance staff, and the ability to survive multi-year award cycles. AI doesn’t abolish the military–industrial complex; it gives it a new dialect.

The irony is that the strongest “individual” in this story is not a fearless engineer but the procurement machine itself: a system that can turn modernization into an annuity. If Code Metal truly can deliver verified rewrites at scale, it might reduce costs. If it can’t, it still may succeed—because in government software, the most reliable output is another contract.