Faraj Aalaei raised $50 million for his first semiconductor company and still had $17 million in the bank when it listed. He raised $200 million for the second. He says a chip now costs several hundred million dollars to develop.
The industry's answer to that has been to design for a market five or six years out and hedge by adding features nobody has asked for. His answer is to shorten the five years.
"Our industry cannot go on taking four years to design a chip and cost hundreds of millions of dollars and not know whether there's gonna be a good market for it or not."
Aalaei has spent more than forty years in semiconductors, took two companies public, sold the second to Marvell in 2019, and now runs Cognichip, which he describes as a frontier model lab that works only on chips.
The full interview is covered here so you can skip it. 18 minutes of audio, 9 minutes of reading.
Here are the 8 takeaways that matter.
👤 Guest: Faraj Aalaei, founder and CEO of Cognichip, who has taken two semiconductor companies public
🎙️ Hosts: John Coogan and Jordi Hays, who run TBPN's daily live tech show
👥 Also on: Nico Wittenborn of Adjacent, Scott Keogh of Scout Motors, Mitchell Green of Lead Edge Capital, and Ben Gilbert and David Rosenthal of Acquired, in separate segments of the same episode
📰 Published: 14 September 2026 on YouTube (TBPN)
🔴 YouTube | 🟣 Apple Podcasts | 🔗 Episode page | ⏱️ 18 min | ✅ Time saved: 9 min
Key Takeaways
Nine tenths of what a chip engineer does today is work Aalaei says current models can do
The point of automating it is to move those people onto new products and markets
A chip designed now is a guess about a market five or six years away
The industry hedges by packing in capability that may never be used, which raises power draw and cost
The general-purpose models are bad at chip design because the training data does not exist in public
Software models work because software has had thirty years of open source
The cost of one chip has gone from $50 million to several hundred million inside one career
He wants a return to four or five people raising a reasonable amount on Sand Hill Road and shipping a chip
1. Two IPOs And A Marvell Exit
The hosts asked for an introduction, and got a career that has run alongside the cost curve he is now attacking.
He has been in the industry more than forty years, and joked that he started when he was two years old
The first company was founded in the late 1990s and listed on Nasdaq
The second went public in 2017 and was sold to Marvell Semiconductors in 2019
He then spent two or three years investing and helping other founders, which is where he learned enough about AI for what he called the light to go on
The idea came from his own experience. He realized that AI might solve the problems he had lived through while building two chip companies, and that has been the mission for the last two and a half years
2. The Cost Curve Broke
The reason he started the company is a set of numbers from his own funding history.
The first company needed $50 million and did not spend it. "When I started my first semiconductor company, I raised 50,000,000. When I took it public, had 17 left in the bank still."
The second needed four times as much. He raised $200 million, and said he took it public because he did not want to do another private round and would rather raise public money
The current figure is an order of magnitude beyond that. He said it now takes several hundred million dollars to do a chip
Complexity is what is driving it, together with the time involved: two to three years to design a chip, then about another year to get customers running, and then five or six years before the company makes money
3. Six Years Behind Software
The gap he is trying to close is between how fast software moves and how slowly silicon can answer it.
The lag is his central claim. Software is moving so fast that it leaves chips behind by almost six years
The second constraint is people. He said the industry has a diminishing number of electrical engineering students graduating, against rising complexity and cost
Put together, he described an industry out of sync. Fewer engineers, more complex chips, higher costs and longer timelines, while the thing the chips are built for moves every few months
The flywheel he wants runs the other way. Collapse the time and the cost, build chips faster, and the software moves faster in turn
4. Artificial Chip Intelligence
Cognichip is built as a lab for one industry rather than as a general model applied to it.
He describes it as a frontier model lab that works only on semiconductors, in contrast to the large labs working on general intelligence
The name for what it builds is ACI. Artificial chip intelligence, as against artificial general intelligence
The staffing is the strategy. Scientists from mathematics and physics, paired with people who have twenty or thirty years of chip design behind them
He measured that experience in tape-outs — the process of finishing a chip and sending it to a fabrication plant — and said people at the company have done hundreds of them
The third group is the software engineers who wrap that knowledge into an enterprise-class product a chip designer can use
5. 90% Of Engineer Time
The claim that carries the business is about what the existing workforce spends its day doing.
His figure is nine tenths. "What a lot of people don't realize is that 90% of the time in the chip business, our engineers are spending doing things that can be done by these models now."
The purpose is not headcount. He wants those engineers moved onto the creative side — new products, new markets, new capabilities
The tools he described are staff, not software features. He called them digital designers, working alongside the human ones to get a project through
The host's framing was the risk that the architecture dates. New chip startups take technical and execution risk, and then the further risk that the design is no longer relevant by the time it ships — and that the buyers now want proof a startup can reach gigawatt scale
6. The Five-Year Guess
Asked how anyone manages that risk, Aalaei said the industry's method is to guess and then over-build.
He called it one of the hardest jobs in the industry. Starting a project knowing the chip does not go into production for another five years, when nobody has that kind of clarity about what will be needed
The standard defense is to add things. Devices get packed with capability that may or may not be necessary, as an insurance policy against what changes
That insurance has a price. The chips get bloated, power consumption goes up, and cost is unnecessarily high — and even then the bet may not come in
His argument is that speed removes the need to hedge. Collapse the development time and the unknowns go with it, raising the probability of hitting the market when you want, with the right power and performance
The host compared the economics to drug development, a large upfront investment on close to a coin flip about whether a market exists at the end
What he wants back is the old shape of the business. Four or five people going to Sand Hill Road, raising a reasonable amount of money, bringing a chip to market and taking the company public — the way his first company was built
He tied that to competition. The rest of the world has woken up to chips, so there will be far more competition, and getting innovation back requires this kind of change
7. Why Generalist Models Fail
The hosts pointed out that general models have recently solved hard open problems, and asked why chip design would be different.
His answer starts with training data. "So the models are as good as the data that you use to train them." If a model has never seen enough examples, it does not know how to do the work
Reasoning does not fill the gap. "And so no amount of reasoning is going to get you to something you've never seen."
The specific shortage is open-source chip design data. Very little is available, and what is there he said is not very useful
The contrast that proves the point is software. Models are good at code because software has had thirty-odd years of open-source data widely available, and the large labs are all consuming the same material
So the company built the dataset first. Two and a half years of work on what he called a data moat, which he claims is now the largest semiconductor dataset in the industry
Data alone is still not enough. Getting a chip from idea to architecture to physical transistors requires the workflow a mature industry uses to make sure everything works
The tolerance is what makes it unlike other applications. With hundreds of billions of transistors on a piece of silicon, they work because not one of them is misplaced. "You can't, make things up. It's not like poetry."
The conclusion is a judgment about the industry's viability. "Our industry cannot go on taking four years to design a chip and cost hundreds of millions of dollars and not know whether there's gonna be a good market for it or not. That's not sustainable. That's not investable."
8. DSL, Then 10 Gig
The hosts asked why he was able to raise $50 million on Sand Hill Road in the first place, and the answer was a record that still stands.
The first company was built for the dial-up era's successor. The team was inventing a new way of building a chip for DSL, which is how broadband still reaches many homes
The speed is the part he is proudest of. "And we went from opening the door to having a chip in hand that we were selling and generating revenue, and we took the company public almost three years to the day we opened the door."
He said that record has not been beaten for a semiconductor company going from inception to a public listing
His list of what it took has luck in it. A good idea, a great team, good timing and good luck, all of which he said came together on the first company
The second company was a data center bet on a speed transition. Data centers were moving from one gig to 10 gig, and the company built what he called the world's first 10 gig chips
Aalaei's bottom line is that the semiconductor industry's economics have stopped working — too much money, too long a wait, too little visibility — and that the fix is to compress the design cycle with models trained on data that only exists inside the industry.
Bonus Insights
He has never built the thing he is now selling into. As he put it, this time he is not building a chip but a system that helps everybody else build chips faster
The chat asked about the art behind him. It is Tom Brady; Aalaei is a Patriots fan, and a host suggested he try to get Brady into the next funding round
Products, Companies & Tools Mentioned
Cognichip (His company: a frontier model lab for semiconductors, building what he calls artificial chip intelligence)
Marvell (Bought his second company in 2019)
Nasdaq (Where his first semiconductor company listed in the late 1990s)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: