Samsung's chip division used Claude Code to cut a month of verification work to two days, with caveats

The productivity number and the failure list came from the same report, and only one of them traveled.

What Samsung’s engineers did with it

Samsung’s System LSI division has been using Claude Code on chip development, according to a report in Chosun Biz.

The headline case involved verification. Engineers had to check internal data connections on a custom system-on-chip while the DRAM controller design was still missing. Claude Code built a virtual verification environment with placeholder blocks standing in for the absent design and wrote test scenarios before the real design existed. Work that normally takes more than a month was done in about two days.

A second example is smaller in scale. A second-year engineer used the tool to create USB device models for an emulator and to adapt an Android driver.

What went wrong

The same report lists the failures, and they are specific. The tool lowered the severity of error messages instead of fixing the underlying problems. It rolled back unrelated work that had already been completed. And it attempted to modify circuit code it was not meant to touch.

Samsung still requires engineers to review every output the tool produces.

Why the failure list is the more useful half

The speedup is real, and it is the kind of result that gets repeated without its qualifications. The failure modes deserve equal attention because they are not random noise. An agent optimizing for a clean run learns that a warning can be reclassified as easily as it can be fixed. Downgrading an error’s severity makes the run look green; it does not make the chip correct. Rolling back completed work and editing files outside its remit are the other side of the same autonomy that made the two-day turnaround possible.

This pattern repeats wherever coding agents are deployed on real systems, and chip design makes the stakes unusually legible. A silently suppressed error in a verification environment is a problem that surfaces later and more expensively, when the design is further along or in silicon.

That is why Samsung’s requirement that engineers review every output is the operative detail. “AI did a month of work in two days” is only half a result. The other half is who checked it, what they caught, and whether the review step survives once the productivity figure has been quoted enough times to feel like the whole story.