Cognichip wants AI to make AI-powered chips, and just raised $60M to test it

Advanced silicon chips have fueled the growth of artificial intelligence. Now can AI return?

Cognichip is developing a deep learning approach to work with engineers as they develop new computer chips. The problem it’s trying to solve is one that companies have had for years: Chip design is too complicated, too expensive, and too slow. Advanced chips take three to five years to go from conception to mass production; the preparation phase alone can take up to two years before physical appearance begins. Consider that the latest line of Nvidia GPUs, Blackwell, has 104 billion transistors – that’s a lot.

In the time it takes to develop a new chip, Cognichip CEO and founder Faraj Aalaei says the market could change and cause the entire investment to disappear. Aalaei’s goal is to bring AI-based tools that software engineers can use to speed up their work in the semiconductor industry.

“These machines are now smart enough that just by guiding them and telling them what the results are and what you want, they can create beautiful code,” Aalaei told TechCrunch.

It is said that the company’s technology can reduce the cost of chip development by 75% and cut the time in half.

The company quietly exited last year and said Wednesday it raised $60 million in new funding led by Seligman Ventures, with participation from Intel CEO Lip-Bu Tan, who will join the Cognichip team. Umesh Padval, managing partner at Seligman, will also join the team. Cognichip has now raised $93 million in total since its launch in 2024.

However, Cognichip wouldn’t point to a new device developed by its machine and didn’t reveal any of the customers it says it has been working with since September.

The company says the advantage is that it uses its model which is taught on chip graphics, rather than starting with a demanding LLM. This requires gaining access to domain-specific academic knowledge, which is no small feat. Unlike software developers, who share much of their code openly, chip makers closely monitor their IP, creating the kind of open environment that training AI coders often lacks.

Cognichip had to create its own datasets, including production data, and license data from partners. The company has also developed ways to allow chip makers to securely train Cognichip models on their own data without exposing it.

In the absence of proprietary data, Cognichip leaned on other open sources. In one demonstration last year, Cognichip invited electronics engineering students from San Jose State University to test the prototype in a hackathon. The team was able to use the model to build CPUs based on the open source RISC-V architecture – a freely available architecture that anyone can build on.

Cognichip is competing with players like Synopsys and Cadence Design Systems, as well as startups with more money like ChipAgents, which closed a Series A round of $74 million in February, and Recursive, which raised. $300 million Series A round in January.

Padval said the current flood of investment in AI is the biggest he has seen in 40 years of trading.

“If it’s a high-end cycle of semiconductors and hardware, it’s a great move for a company like (Cognichip),” he said.