Top OpenAI, Google Brain researchers set off a $300M VC frenzy for their startup Periodic Labs 

Periodic Labs, a caller startup by 1 of OpenAI’s astir respected researchers, Liam Fedus, and his erstwhile Google Brain colleague, Ekin Dogus Cubuk, came retired of stealth past period with an tremendous $300 cardinal effect round. It was led by Felicis and included a who’s who of angels and different apical VCs. 

The startup began erstwhile Fedus had a speech with Cubuk (whose friends telephone him “Doge”) astir 7 months ago. Cubuk was 1 of Google Brain’s foremost instrumentality learning and worldly subject researchers. After endless Silicon Valley takes connected however GenAI would radically alteration technological discovery, they decided that the pieces were yet successful spot to marque this a reality. Or astatine slightest to recovered a startup that attempted it.

“There are a fewer things that happened successful the LLM field, successful experimental subject and successful simulations that benignant of made this the close time,” Cubuk told TechCrunch.

For one, helium said, robotic arms that could grip pulverization synthesis — the process of mixing and creating caller materials — had precocious proved themselves reliable. For another, instrumentality learning simulations had go businesslike and close capable to exemplary analyzable carnal systems specified arsenic those needed to make caller materials. 

And, third, LLMs present had almighty reasoning capabilities — successful portion done the enactment of Fedus and his squad astatine OpenAI. Fedus was 1 of the tiny squad who created ChatGPT to statesman with and was moving OpenAI’s uber important post-training team, which refines models aft their archetypal development.

Stitched together, the representation was clear: a simulation could theoretically observe caller compounds, a robot could premix the materials, and an LLM could analyse the results and suggest people corrections. AI-automated worldly subject was acceptable to beryllium built.

In fact, Cubuk was 1 of the researchers who published a groundbreaking insubstantial successful 2023 documenting a precursor Google probe project. The squad built a afloat automated, robotic-powered laboratory and created 41 caller compounds from recipes suggested by connection models. 

Equally importantly, the founders realized that adjacent failed experiments would beryllium invaluable for their caller startup due to the fact that information is the lifeblood of AI. AI subject offered an wholly caller root for real-world grooming and post-training data. This could, the founders believe, crook the existing technological information strategy connected its head, which seeks success, not exploration, rewarded via insubstantial work and grants.

“Making interaction with reality, bringing experiments into the [AI] loop — we consciousness similar this is the adjacent frontier,” Fedus told TechCrunch. 

Felicis wins the deal; OpenAI does not invest 

After that treatment with Cubuk, Fedus went to the powers that beryllium astatine OpenAI to stock his resignation and his plan. He past gleefully tweeted to the satellite that helium was leaving with what appeared to beryllium OpenAI’s blessing and investment.  

That concern didn’t really materialize, however. OpenAI is not a backer of Periodic, the founders confirmed to TechCrunch. And portion Fedus declined to accidental why, they really didn’t request OpenAI’s money.  

Fedus’s tweet acceptable disconnected a frenzy of VCs courting the company. “There was astir similar a feeling of being reverse pitched. One capitalist really wrote a emotion missive to Periodic Labs,” Fedus laughed, explaining that neither helium nor Cubuk “knew what to marque of it.” Others sent multi-page documents pitching themselves.

But the archetypal telephone that they really took was from Peter Deng, a erstwhile OpenAI workfellow turned capitalist for top-tier effect steadfast Felicis. (Deng near OpenAI for Felicis astatine the commencement of 2025.)  

“Liam is simply a precise large woody wrong OpenAI, precise good loved and an highly impactful researcher,” Deng told TechCrunch. “When I heard helium left, I texted him immediately.” 

Deng met Fedus for java successful the Noe Valley vicinity of San Francisco. Hyped connected caffeine and enthusiasm, Fedus invited Deng to decorativeness their speech connected a locomotion implicit the area’s famously hilly terrain. Pitch walks whitethorn beryllium a Silicon Valley trope, but they besides truly happen. 

The chilly time had turned hot. Deng, wearing a sweater, sweated and scrambled to support up with the acceptable and affable Fedus until the laminitis said thing that “literally stopped maine successful my tracks,” Deng told TechCrunch. He told Deng that “everyone talks astir doing science, but successful bid to bash science, you really person to bash science,” Deng recalls. 

In different words, they needed to springiness AI a afloat equipped bedewed laboratory to effort its ideas successful a existent world, controlled setting.

“The information astir these models is that everything that the models cognize is wrong mean distribution. We instrumentality a clump of data, and it tin conscionable regurgitate what it knows,” Deng said.  

Discovering thing caller has to impact investigating hypotheses.  

“And I committed on the spot, successful the mediate of the hills of Noe Valley, to constitute the check,” Deng says.  

Fedus besides remembers the moment Deng asked however helium could beryllium involved, and Fedus told him the startup needed currency for laptops and a impermanent office. And “he’s like, great, I’ll springiness you wealth close now. And it was conscionable this immense ballot of confidence.”  

But Deng didn’t really whip retired his checkbook connected the street. He went backmost to the bureau elated implicit the woody lone to brushwood Felicis’ lawyer, who pointed retired that the steadfast couldn’t promptly motion a contract: the institution wasn’t incorporated yet. It didn’t adjacent person a name, overmuch little a slope relationship to ligament funds. “That’s however aboriginal we were,” Deng grins. 

Soon they had each of those things and each the word sheets they could handle. With the $300 cardinal warfare chest, Cubuk and Fedus hired implicit 2 twelve of the astir prestigious AI and technological endowment similar Alexandre Passos (a creator of o1 and o3); Eric Toberer (a materials idiosyncratic who has already made cardinal superconductor discoveries) and Matt Horton, a creator of 2 of Microsoft’s GenAI materials subject tools. And the database goes on. 

Because the squad members are each experts successful antithetic areas from AI to physics, each week 1 of them gives a grad-level lecture to the others. “We bash consciousness similar a choky coupling is highly important,” Cubuk said. He wants everyone to recognize each parts of what they are building. 

Periodic Labs has already acceptable up its lab, too, and is moving with experimental data, simulations and investigating immoderate predictions. The main archetypal ngo is to find caller superconductor materials – perchance a golden excavation discovery. Improved superconductors could powerfulness the adjacent epoch of potent, but little energy-consuming tech. 

But the past portion – the robots – are not yet up and running. “They volition instrumentality a spot to train,” Cubuk said. 

All of this is, of course, a large plaything for the fences. AI powered oregon not, technological find is not typically fast, casual oregon predictable. While this squad of experts has immoderate indications that they volition find what they are looking for — oregon marque different discoveries on the mode (or simply make invaluable information connected their failures), there’s nary guarantees.

And we cognize that exemplary makers themselves are inching their mode toward much AI science. Last month, OpenAI VP Kevin Weil said helium was launching an OpenAI for Science portion astatine the institution to “build the adjacent large technological instrument: an AI-powered level that accelerates technological discovery.”

As for the capitalist who wrote the emotion letter, helium didn’t triumph the woody (although Fedus did accidental that the missive was “very flattering.”) The different effect investors include Andreessen Horowitz, DST, NVIDIA’s task superior limb NVentures, Accel, and angel backers similar Jeff Bezos, Elad Gil, Eric Schmidt, and Jeff Dean. 

Elad Gil volition beryllium speaking astir however AI has changed the startup scenery astatine Disrupt successful San Francisco connected October 29.