Compute is the most valuable resource in the AI world we live in today. Nvidia. Google TPUs. Amazon Trainium. OpenAI and Broadcom’s partnership. Elon’s recent post about Tesla’s AI chips.
Designing the most performant chips for AI workloads sits at the heart of accelerating technological progress.
But major hurdles exist.
First, chip design is slow. It takes 12-24 months at mature nodes and 18-36 months at the leading edge for 5nm or 3nm.
Second, chip design is prohibitively expensive. It costs on average $200-250 million for 7nm, $450-500 million for 5nm, and $600-650 million for 3nm. Roughly 50-70% of that is human labor. Another 5-15% is Electronic Design Automation tooling spend in a market long dominated by Cadence and Synopsys, where each generates $5-6 billion in annual revenue and are worth approximately $90-100 billion in market cap.
AlphaChip caught my eye for these exact reasons. It gave us a peek at AI’s potential to transform the entire chip design process, showing we can cut the floorplanning step in physical design from months to hours.
What if we could extrapolate this and build AI to automate the entire flow, from architecture design to RTL to verification, all the way through physical design?
What if chip design took days, not two to three years? Every day is massively costly; some reports from August 2024 indicated that a multi-month Blackwell delay could result in more than $10 billion in lost revenue for 2025 alone. More importantly, imagine the revenue potential unlocked when new generations of chips are designed faster and shipped earlier.
What if each design didn’t cost hundreds of millions of dollars? What if chip companies didn’t need to operate large human teams on top of clunky EDA tooling?
And most exciting: what if we unlocked novel chip designs we might never have explored?
AlphaChip revealed an important human bias: in chip design, we tend to think in Manhattan grid-like structures. AlphaChip’s designs were different, more organic in shape, more like forms inspired by nature. So different, in fact, that humans wanted to reject them at first … Yet AlphaChip went on to shape four generations of the TPU.
We at Sequoia are so excited to partner with co-founders Anna Goldie and Azalia Mirhoseini, leading their very first round from the formation of Ricursive Intelligence. They pioneered AI for chip design by creating and leading the AlphaChip effort and are at the epicenter of this emerging AI for chip design ecosystem. They are visionaries with incredible clarity of thought, intensely ambitious, humble yet exceptionally accomplished, and real talent magnets who move, and inspire others to move, with urgency and velocity.
Anna and Azalia founded Ricursive Intelligence to build the frontier AI lab defining this category. In just the first weeks since company formation, they have assembled a team with the highest talent density you can imagine in the field.
Their core belief: chip design is the compute bottleneck, and progress in AI, hardware and infrastructure is capped by the speed and efficiency of silicon creation.
In their words: “If we get this right, it’s not just faster chip design cycles; it’s a fundamental expansion of what’s possible in hardware. Once chip design becomes fast and accessible, everyone will be able to customize. The automation here will unlock a flood of new hardware innovation.”
Anna and Azalia’s vision for Ricursive is to define a new movement, from “fabless” to “designless.” Fabless, meaning a company designing chips without owning expensive fabs, outsourcing production to foundries. Designless, meaning outsourcing not only manufacturing but the entire chip design process, taking an idea and converting it into a manufacturable design.
We envision a world where Ricursive helps any company design chips for its own workloads faster, more efficiently and more creatively than is possible today. In doing so, Ricursive can help revolutionize the most valuable resource in our era: compute. We could not be more excited to help build a true generational company in the making.