Cognichip founder Faraj Aalaei is building AI that designs chips faster — and could return semiconductors to the startup era

Faraj Aalaei, Cognichip

Faraj Aalaei has spent four decades in semiconductors, including two companies he took public — the first from inception to Nasdaq IPO in exactly three years, which he says still holds the record for the fastest a semiconductor company has gone from opening its doors to going public. The second he sold to Marvell in 2019. Cognichip, founded two and a half years ago, is his attempt to fix what he sees as the structural problem that made both of those companies harder than they needed to be.

The core problem

Chip development costs have compounded relentlessly. Aalaei's first company raised $50M and still had $17M left when it went public. The second required $200M, and he took it public partly to avoid another private round. Today a competitive chip can require several hundred million dollars to develop, takes two to three years to design, and another year to get customers running — meaning a startup that begins development today won't see meaningful revenue for five to six years. Software moves on a fraction of that timeline, and the gap is widening.

A shrinking pipeline of electrical engineering graduates compounds the pressure. Aalaei's assessment is blunt: the industry cannot sustain a model where chips take four years to design, cost hundreds of millions, and carry deep uncertainty about whether the market will still want them on arrival.

“90% of the time in the chip business, our engineers are spending doing things that can be done by these models now. The reason large LLMs are not good in chip design is because chip design data as open source is actually not available — we're the first company that took that on. My vision is to return our industry back to a point where four or five of us can go to Sand Hill Road, raise a reasonable amount of money, and bring a chip to market.”

What Cognichip is building

Cognichip is a frontier model lab focused exclusively on semiconductor design — what Aalaei calls ACI, artificial chip intelligence. The pitch is that 90% of chip engineering time goes to work these models can now do, freeing engineers to focus on architecture and product decisions rather than execution.

The central bet is on proprietary data. General-purpose LLMs perform well on software because decades of open-source code gave them rich training material. No equivalent data set exists for semiconductor design — chip design data is largely closed, and what is publicly available Aalaei describes as not very useful. Cognichip has spent two and a half years building what he claims is now the largest proprietary semiconductor design data set in the industry, pairing mathematicians and physicists with engineers who have collectively completed hundreds of tape-outs.

The argument against generalist models catching up is straightforward: reasoning alone cannot substitute for domain-specific training data. A model that has never seen enough chip design examples cannot reliably produce correct output, and in a domain where hundreds of billions of transistors must be placed without a single error, hallucination is not an acceptable failure mode.

The market case

Beyond cost reduction, compressing design timelines changes the risk profile of chip startups entirely. A five-to-six year development window forces teams to over-engineer chips as a hedge against unknowns, bloating power and cost without guaranteeing market fit. Aalaei's vision is to return semiconductors to something closer to the conditions of his first company — a small team, a manageable raise, and a realistic path to market. Whether Cognichip's tooling can compress timelines enough to make that credible is still unproven, but it is the explicit goal.

Aalaei describes the broader ambition as bringing innovation back to an industry that has effectively priced out startups — particularly as global competition in semiconductors intensifies and the stakes of falling behind rise.