AI is coming for one of the hardest jobs in the technology industry: chip design. The nascent startup Architect Labs, founded last year, is betting on the promise of using AI to accelerate design for both chipmakers and companies seeking custom hardware. It says it designed and verified a chip called Redwood in two weeks, with two human chip architects setting and refining the overall plan while AI handled the detailed design work.
It would dramatically speed up a process that typically takes a year or more, costs hundreds of millions of dollars, and requires large, specialized teams of around 100 people, the company said.
This lengthy timeline is especially problematic in AI, where models evolve faster than the chips they run on, said Architect Labs cofounder and CEO Ebrahim Hussain, who is 20.
Architect emerged from stealth in June with a $24 million seed round led by Kindred Ventures. Google's former chief scientist, Jeff Dean, and other executives from OpenAI and Nvidia also backed the startup.
Hao Zheng, an assistant professor of electrical and computer engineering at the University of Central Florida, called the Redwood timeline "impressive," though he said several startups and more established companies, such as Synopsys, also use AI to accelerate the chip design process.
The Redwood experiment revealed something else, said Hussain. After running an AI model on Redwood, the model found ways to improve the chip itself, which he said suggests a self-reinforcing loop.
"It's still very early, but we're very excited about the possibility of growing to larger models and getting AI to essentially recursively improve itself and build better and better hardware," Hussain said.
Redwood can run AI models like Meta's Llama and Alibaba's Qwen, Architect said, though there's a caveat — it hasn't been manufactured yet. Instead, the design has been tested on reprogrammable hardware known as a field-programmable gate array (FPGA), and Architect said those tests suggest the finished product could be more efficient than Nvidia's Jetson Orin Nano, an AI supercomputer used in robotics.
Hussain said the company plans to send Redwood to the Taiwanese chip manufacturer TSMC to be manufactured and hasn't decided when because its AI is still working to improve the design.
Manufacturing is the real test, said Zheng, and it will ultimately reveal how fast, efficient, and reliable the chip is.
Matthew Guthaus, a computer science and engineering professor at UC Santa Cruz, said that while AI is already good at many parts of chip design, "it can also make silly mistakes and needs close supervision and double-checking."
These errors can cost millions if they aren't caught before manufacturing, Guthaus said. Hussain said that Architect's AI system checks its own work for mistakes rather than relying on humans to do so.
Hussain said that AI could usher in a new era for the chip industry. While TSMC helped introduce a "fabless" era, allowing companies like Nvidia to design chips without owning factories, he says the next phase will be "designless" — putting custom chips within reach without the huge teams traditionally required to build them.