Reinventing chip design for the AI era: Our investment in Phinity Labs

Nearly everyone I speak with, from the founder of a 2-person startup to an engineering leader at a frontier lab, keeps running into the same wall when they try to move faster. It’s a physical wall, and it’s made of silicon. We’re in a global compute shortage, and we can’t build chips anywhere close to fast enough to keep up with demand. 

Phinity Labs, an applied research lab for autonomous chip design and discovery, is building a way through that wall. We’re proud to have led their $5.2 seed round, alongside our friends at Pear and Moxxie, and industry leaders like Jeff Dean. 

Chip design today is incredibly capital and time intensive. Designing a chip involves a long sequence of highly specialized steps, from deciding on its architecture to writing the code that defines how it works, verifying that design, and figuring out how to physically lay out billions of transistors before sending it off to manufacturing. Creating a new chip typically means committing to spending hundreds of millions of dollars and the better part of two years on the endeavor. It requires locking in architectural decisions long before anyone even knows what workloads the chip will be expected to support. Changing those decisions midway through can mean redoing months of work across multiple teams. Given this, the typical approach of most in the chip industry is one of extreme caution.

Phinity aims to upend this entire process. They are building the infrastructure that teaches AI models how to design, verify, and tape out chips. Eventually, their goal is to let anyone go from “prompt to silicon” through a closed loop system where agents propose architectures, run them through the flow, learn from what fails, and converge on manufacturable designs without requiring human bottlenecks at each step. 

I first met Phinity founders Sonya Jin and Aadi Nashikkar at the PearX Demo Day late last year and was immediately struck by the scale of their ambition and the clarity of their vision. They reminded me of the Collison brothers from my time at Stripe. Taking on a multi-trillion dollar industry with decades-old entrenched incumbents requires a special combination of fearlessness and youthful naivete, and they had both.

At the time, the company was merely a few months old, with both founders having recently left Stanford, where they met while pursuing Computer Science degrees with a concentration in AI. Sonya had already done AI research and post-training work at NVIDIA and AWS AI Lab, while Aadi had been the co-founder and CTO of a YC startup and later built a synthetic data and assets platform that was acquired. 

What was unique was that they brought deep AI expertise to an industry historically dominated by hardware engineers. Rather than simply applying coding agents to existing workflows, their vision was to train AI models to design chips themselves. 

When we met, they didn’t yet have a signed contract or a dollar in revenue, so the bet was largely on them as founders and their approach to one of the most technically complex problems in existence. Since then, Sonya and Aadi have assembled a team with deep experience across AI research, silicon architecture, design, and verification, including the former director at Intel Emerging Technologies, NVIDIA architects who have led multiple tapeouts, AI researchers who have trained LLMs for chip design, and more. 

Phinity’s agents have outperformed experienced hardware engineers on power, performance, and area, even when those engineers were using frontier coding agents. In one AI inference chip task, a Phinity agent delivered roughly twice the area improvement of the best human result. 

This performance has led to a surge in demand, which the team is moving quickly to onboard. 

Since they released the first version of their platform just a few months ago, they have partnered with the world’s leading frontier labs, going from $0 to eight figures in annualized run rate (the fastest revenue ramp of any investment I’ve made in seven years in venture). 

Part of what sets Phinity apart is that they’ve already built a profitable business working with these labs, while simultaneously developing the infrastructure to train their own increasingly capable chip design agents. These partnerships give them access to unreleased models, while their own environments, graders, and performance data become more valuable over time as they tackle more complex chip design problems. They’ve also begun building custom EDA tools designed specifically for AI agents.

While Phinity has already achieved extraordinary early traction, they’ve barely scratched the surface of what’s possible as their agents take on increasingly complex chip designs and, eventually, the entire process from architecture to tapeout.

If you’re a frontier lab or semiconductor company interested in dramatically shortening the chip design process, the Phinity team would love to chat. And, if you’re passionate about the future of AI and interested in joining one of the fastest-moving teams we’ve had the privilege to partner with, they’re hiring. 

Welcome, Phinity!