Can an AI agent design a whole chip? A new benchmark suggests it is still early days. On October 14, 2026, ChipAgents announced Build-A-Chip-Bench, a test built with NVIDIA to see whether AI agents can carry an end-to-end digital chip design from specification through physical implementation. For anyone following AI chip design, the first results show a wide gap between designs that work and designs that are physically viable.
What Build-A-Chip-Bench tests in AI chip design
According to ChipAgents, many existing hardware-AI benchmarks focus on RTL generation or functional verification and largely stop before physical implementation. But a design that simulates correctly can still miss timing, exceed its area or power budget, fail to route or break physical-design rules.
Build-A-Chip-Bench is meant to cover the full flow: architecture, microarchitecture, RTL, functional verification, logic synthesis, floor planning, placement, clock-tree synthesis, routing, timing and power signoff, and physical verification, ending in a verified Graphic Design System (GDS) file. The company says it tests whether an agent can meet functional, timing, power, area and physical-verification requirements together.
The first results: about 40% versus about 10%
ChipAgents reports that, in initial evaluations, tested agents passed functional verification on about 40% of tasks on average. Only approximately 10% met the required physical performance, power and area (PPA) targets. The company says this shows that generating functionally correct RTL is only one part of the engineering challenge, and that agents must also reason across abstraction levels, keep validated behavior through revisions, use specialized tools and plan around expensive physical feedback.
William Wang, founder and CEO of ChipAgents, said the real challenge is carrying engineering intent through a long chain of connected decisions, and that the benchmark gives the industry a way to measure the gap. Tim Costa of NVIDIA said chip design is “a defining test” of whether AI decisions hold up under the demands of the physical world. The benchmark was developed by Surya Gunukula, Jerry Yicheng Wang and Kevin Zhang, with technical input from NVIDIA.
Why an AI chip design benchmark matters
Benchmarks shape what researchers and companies try to improve. A test that goes beyond code generation could point effort at the harder physical steps, where mistakes are costly. It also gives buyers and engineers a shared yardstick, which is useful at a time when many tools claim to speed up chip design. This sits alongside broader work on AI-ready infrastructure, such as AMD’s Helion open rack blueprint for agentic AI.
The other side: reading the numbers carefully
The 40% and 10% figures come from ChipAgents’ own initial evaluations. The announcement does not say which agents were tested, how many tasks were included or how tasks were chosen, and it points readers to the company’s research page for methodology. ChipAgents also sells an agentic AI platform for semiconductor engineering, so it has a commercial interest in how the field is measured. “Approximately” is used for both numbers, so they should be read as rough indicators, not precise scores.
There are security questions too. Chip designs are valuable intellectual property, so teams will want to know where their data goes when agents run design tools. And an AI-generated design that passes tests may still hide subtle flaws, which makes independent review important before tapeout. The release does not cover either concern, so these are our own points to raise. The same accountability debate is playing out more widely, as in why IT leaders expect to be blamed when AI agents go wrong.
My view on AI chip design
As a writer, I feel the most useful part of this story is the honesty of the gap. It would be easy to show only the 40% number, but the roughly 10% physical result is the one that tells us how far autonomous design still has to go. I like that a benchmark is trying to measure it. I would still want to see the full methodology, results from other teams and replication outside the company that built the test before drawing big conclusions. Readers can judge for themselves how close fully autonomous chip design really is.
FAQ: Build-A-Chip-Bench and AI chip design
What is Build-A-Chip-Bench?
It is a benchmark from ChipAgents, developed with NVIDIA, that tests AI agents on an end-to-end digital chip design task from specification to verified GDS.
What did the first results show?
According to ChipAgents, agents passed functional verification on about 40% of tasks on average, while about 10% met the required physical performance, power and area targets.
Who built the benchmark?
ChipAgents says it was developed by Surya Gunukula, Jerry Yicheng Wang and Kevin Zhang, with technical expertise from NVIDIA on the suite and scoring rubric.