How Recursive Intelligence raised $335M at a $4B valuation in four months

The co-founders of Recursive Intelligence seem destined to be co-founders.

Anna Goldie, CEO, and Azalia Mirhoseini, CTO, are so well-known in the AI ​​community that they were among the AI ​​engineers who “got these weird emails from Zuckerberg driving us crazy,” Goldie told TechCrunch, laughing. (They did not receive donations.) The two worked on Google Brain together and were early employees at Anthropic.

They have been credited by Google for creating the Alpha Chip – an AI tool that can create solid chip configurations in hours – a process that would normally take human designers a year or more. This tool helped create three generations of Google Tensor Processing Units.

Native explains why, just four months after launching Recursive, last month he announced $300 million Series A round for $4 billion led by Lightspeed, just months after raising a $35 million seed round led by Sequoia.

Recursive is making AI tools that make chips, not chips themselves. This makes them very different from almost all AI chip startups: they are not Nvidia wannabe competitors. Instead, Nvidia is an investor. The GPU giant, along with AMD, Intel, and the rest of the chip makers, are the target customers.

“We want to give any chip, whether it’s a chip chip or a traditional chip, any kind of chip, to be built in a way that’s sustainable and very fast. We’re using AI to do that,” Mirhoseini told TechCrunch.

Their paths crossed at Stanford, where Goldie earned his PhD while Mirhoseini taught computer science classes. Since then, their work has been closed. “We started at Google Brain the same day. We left Google Brain the same day. We joined Anthropic the same day. We left Anthropic the same day. We rejoined Google the same day, then we left Google the same day. Then we started the company again the same day,” Goldie said.

During their time at Google, the co-workers were so close that they worked together, all enjoying the learning of the community. The pun was not lost on Jeff Dean, the famous Google engineer who was a co-founder. They called their Alpha Chip project a “chip circuit tutorial” – a role-playing game for sharing. Internally, the duo was renamed: A&A.

The Alpha Chip introduced them to the industry, but it also attracted controversy. In 2022, one of his colleagues at Google was fired, Wired saidafter years of trying to discredit A&A with their chip project, even though the project was used to help create some of Google’s most important products, bet-the-business AI chips.

Their Alpha Chip project at Google Brain proved a potential Recursive concept – using AI to speed up chip design.

Making chips is difficult

The problem is, computer chips have millions to billions of components embedded on their silicon wafers. Developers spend a year or more putting the components on a chip to ensure functionality, power efficiency and any other requirements. Determining the systematic placement of such small units vertically is difficult, as you might expect.

The Alpha Chip “can do very high-quality designs in, like, 6 hours. And the great thing about this process is that it learns from experience,” Goldie said.

The purpose of their AI chip design is to use a “reward token” that shows how good the design is. The therapist takes the dose to “change parts of his neural network to be better,” Goldie said. After making thousands, the agent became the best. It also got there as quickly as it learned, the founders say.

The Recursive platform takes that idea a step further. The AI ​​chip maker that’s building it “will study different chips,” Goldie said. So every chip it makes must make it better for every subsequent chip.

The Recursive platform also uses LLMs and handles everything from product placement through design validation. Any company that makes electronics and needs chips is their customer.

If their platform proves itself, as it appears, Recursive could play a role in the moon’s goal to achieve Artificial General Intelligence (AGI). Indeed, their ultimate vision is to create AI chips, meaning that AI will create its own computer brain.

“Chips are the fuel for AI,” Goldie said. “I think by making more powerful chips, that’s the best way to push the boundaries.”

Mirhoseini adds that the long process of making a chip is hindering AI progress. “We think that we can also enable the evolution of different types of chips that power them,” he said. So AI can grow smarter faster.

If the idea of ​​AI developing its brain at an ever-increasing speed brings visions of Skynet and Terminator to mind, the founders point out that there is something better, more recent and, they think, more profitable: hardware performance.

As AI Labs can produce more chips (and, eventually, all basic devices), their growth shouldn’t cost much. of the country’s economy.

“We can create a computer design that’s right for the brand, and we can achieve about 10x the efficiency at total cost of ownership,” Goldie said.

While the young startup won’t name its first customers, the founders say they’ve heard every chip maker name you can think of. Unsurprisingly, they have their choice of development partners, too.