Travis Brashears, Cameron Ramos and Serena Grown-Haeberli began collaborating at SpaceX, developing optical communications links that keep thousands of Starlink internet satellites in constant contact. 

Now, the three engineers are co-founders of Mesh Optical Technologies, a Los Angeles startup that announced a $50 million Series A led by Thrive Capital on Tuesday. 

Mesh aims to mass-produce optical transceivers, devices that convert optical signals from fiber or laser into electrical signals for computers. CEO Brashears, President Ramos, and VP of Product Grown-Haeberli realized the opportunity when designing a new generation of compute-hungry SpaceX satellites forced them to assess the optical transceiver market, and they saw its limitations.

Optical transceivers are particularly important for data centers aimed at training and operating large deep learning models, because they allow multiple GPUs to work in concert. One established U.S. supplier, AOI, won a contract worth $4 billion to provide components for AWS data centers last year.

“Someone will brag about a million GPU cluster; you have to multiply by four to five for the number of transceivers in that cluster,” Brashears explained. 

The company’s goal is to manufacture a thousand units per day within the year so they can begin qualifying for bulk orders in 2027 and 2028.   

The optical transceiver market is dominated by Chinese firms and suppliers, and Mesh sees an advantage in building its supply chain outside of that country. While trade restrictions haven’t impacted the market yet, the founders and their backers see themselves as getting in front of a national security dilemma. 

“If AI is the most important technology in several generations (which we believe to be true), to have critical parts of AI data center capex run through misaligned/competitive countries is a problem,” Thrive Partner Philip Clark wrote TechCrunch. “In the immediate term, Mesh is solving our need for better ways to do interconnect if we want to keep scaling AI.”

The challenge for Mesh, the founders say, is executing lights-out, automated manufacturing techniques, which aren’t common in U.S. industry. So much of this expertise is concentrated in China that even European equipment suppliers expect Chinese customers — one German firm’s standard intake form asks for a Chinese company registration number.

By co-locating design and production, the founders hope to realize more efficient and lower-cost components. Their current design removes one commonly used but power-hungry component, which Ramos said could reduce GPU cluster power usage by 3% to 5%, a meaningful amount as hyperscalers seek to wring as much efficiency out of their systems as possible. 

Data centers are just the beginning of Mesh’s aspirations; the company sees optical wavelength communications as the next paradigm in communications.

“The world has primarily focused on [radio frequencies] for a long time,” Brashears told TechCrunch. “We want to be at the precipice of transition from RF to photonics…we want to interconnect everything, and not just computers, but that’s where we’re starting.”

The co-founders of startup Ricursive Intelligence seemed destined to be co-founders.

Anna Goldie, CEO, and Azalia Mirhoseini, CTO, are so well-known in the AI community that they were among those AI engineers who “got those weird emails from Zuckerberg making crazy offers to us,” Goldie told TechCrunch, chuckling. (They didn’t take the offers.) The pair worked at Google Brain together and were early employees at Anthropic.

They earned acclaim at Google by creating the Alpha Chip — an AI tool that could generate solid chip layouts in hours — a process that normally takes human designers a year or more. The tool helped design three generations of Google’s Tensor Processing Units.

That pedigree explains why, just four months after launching Ricursive, they last month announced a $300 million Series A round at a $4 billion valuation led by Lightspeed, just a couple of months after raising a $35 million seed round led by Sequoia.

Ricursive is building AI tools that design chips, not the chips themselves. That makes them fundamentally different from nearly every other AI chip startup: they’re not a wannabe Nvidia competitor. In fact, Nvidia is an investor. The GPU giant, along with AMD, Intel, and every other chip maker, are the startup’s target customers.

“We want to enable any chip, like a custom chip or a more traditional chip, any kind of chip, to be built in an automated and very accelerated way. We’re using AI to do that,” Mirhoseini told TechCrunch. 

Their paths first crossed at Stanford, where Goldie earned her PhD as Mirhoseini taught computer science classes. Since then, their careers have been in lockstep. “We started at Google Brain on the same day. We left Google Brain on the same day. We joined Anthropic on the same day. We left Anthropic on the same day. We rejoined Google on the same day, and then we left Google again on the same day. Then we started this company together on the same day,” Goldie recounted.

During their time at Google, the colleagues were so close they even worked out together, both enjoying circuit training. The pun wasn’t lost on Jeff Dean, the famed Google engineer who was their collaborator. He nicknamed their Alpha Chip project “chip circuit training” — a play on their shared workout routine. Internally, the pair also got a nickname: A&A. 

The Alpha Chip earned them industry notice, but it also attracted controversy. In 2022, one of their colleagues at Google was fired, Wired reported, after he spent years trying to discredit A&A and their chip work, even though that work was used to help produce some of Google’s most important, bet-the-business AI chips.

Their Alpha Chip project at Google Brain proved the concept that would become Ricursive — using AI to dramatically accelerate chip design.

Designing chips is hard

The issue is, computer chips have millions to billions of logic gate components integrated on their silicon wafer. Human designers can spend a year or more placing those components on the chip to ensure performance, good power utilization and any other design needs. Digitally determining the placement of such infinitesimally small components with precision is, as you might expect, hard. 

Alpha Chip “could generate a very high-quality layout in, like, six hours. And the cool thing about this approach was that it actually learns from experience,” Goldie said. 

The premise of their AI chip design work is to use “a reward signal” that rates how good the design is. The agent then takes that rating to “update the parameters of its deep neural network to get better,” Goldie said. After completing thousands of designs, the agent got really good. It also got faster as it learned, the founders say.

Ricursive’s platform will take the concept further. The AI chip designer they are building will “learn across different chips,” Goldie said. So each chip it designs should help it become a better designer for every next chip.

Ricursive’s platform also makes use of LLMs and will handle everything from component placement through design verification. Any company that makes electronics and needs chips is their target customer.

If their platform proves itself, as it seems likely to do, Ricursive could play a role in the moonshot goal of achieving artificial general intelligence (AGI). Indeed, their ultimate vision is designing AI chips, meaning the AI will essentially design its own computer brains. 

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

Mirhoseini adds that the lengthy chip-design process is constraining how quickly AI can advance. “We think we can also enable this fast co-evolution of the models and the chips that basically power them,” she said. So AI can grow smarter faster. 

If the thought of AI designing its own brains at ever increasing speeds brings visions of Skynet and the Terminator to mind, the founders point out that there’s a more positive, immediate and, they think, more likely benefit: hardware efficiency.  

When AI Labs can design far more efficient chips (and, eventually all the underlying hardware), their growth won’t have to consume so much of the world’s resources. 

“We could design a computer architecture that’s uniquely suited to that model, and we could achieve almost a 10x improvement in performance per total cost of ownership,” Goldie said. 

While the young startup won’t name its early customers, the founders say that they’ve heard from every big chip making name you can imagine. Unsurprisingly, they have their pick of their first development partners, too.