Former OpenAI vice president teams up with DeepMind scientists to start a business: Over 20 elite scientists and $300 million bet on "AI for science"

“Our goal is to create an AI scientist. The way science works is to speculate about what the world might be like, conduct experiments, and learn from the results.” — Quote from the official blog of Periodic Labs.

In the spring of 2025, a shocking resignation announcement spread: Liam Fedus, who formerly served as the vice president of research at OpenAI and was in charge of post - training work, decided to leave. In a tweet, he wrote that he had “the most strategic interest” in “the application of AI in science” — that was the direction of his new journey to explore.

Meanwhile, another heavyweight figure, Ekin Dogus Cubuk — who once led the chemistry and materials science team at DeepMind and participated in projects generating over two million crystal structures — also resolutely chose to leave DeepMind and turned to entrepreneurship.

Left: Ekin Doğuş Cubuk, Right: Liam Fedus. Image source: TechCrunch

“The valuable data of 10 trillion tokens on the Internet is nearly exhausted, and expanding parameters can't bring about qualitative leaps,” Fedus said straightforwardly in a sharing session. Cubuk's addition was even more incisive: “Merely deducing in the literature with large language models will never lead to disruptive discoveries like room - temperature superconductors.”

So, the two hit it off at the beginning of this year. Instead of “involution” in the existing data pool, they decided to let AI enter the laboratory and create data from scratch.

Founding Purpose: The Inspirational Collision between AI and Physics

The founding of Periodic Labs originated from an inspirational collision. Seven months ago, Fedus and Cubuk discussed in a conversation in San Francisco how generative artificial intelligence could reshape the process of scientific discovery. Both of them had witnessed the power of AI in different laboratories but also felt its boundaries at the same time.

“We realized that generative AI can already write papers, program, and even paint, but it hasn't really helped humans discover new knowledge,” Fedus recalled. “The experimental speed in the scientific community is too slow, and AI is ready to change all this.”

Cubuk approached from the perspective of physics. He saw that the technological curves of robot automation, material simulation, and AI reasoning converged at the same time point. “This is an unprecedented opportunity,” he explained. “Robot automation, simulation accuracy, and the reasoning ability of large language models can finally be integrated into one system.”

That day's conversation became the starting point of Periodic Labs. A few weeks later, the two officially left their respective companies, gathered a group of like - minded scientists, and founded a research company that uses AI to drive experimental science.

Subverting Traditional Concepts: Building an AI - Driven Science Platform

Periodic Labs claims to be building an “AI - driven science platform”. Its vision is to enable artificial intelligence not only to analyze data but also to design experiments, drive physical instruments, and discover new materials.

In other words, it attempts to integrate “intelligence” and “experimental operations” into a closed - loop system — from algorithms to reagent bottles, from large models to robot arms.

This is not a new topic. In the past decade, AI has made breakthroughs in areas such as drug design, protein folding, and material simulation — DeepMind's AlphaFold, Microsoft and Meta's molecular generation models, and Chemify's automated chemical system have all proven that AI can participate in scientific discovery.

But Periodic Labs' goal is even greater. Fedus and Cubuk want to create a “universal experimental entity” — to enable AI not only to understand science but also to conduct experiments in a real laboratory.

In the concept of Periodic Labs, there is a highly subversive view: Failed data is equally valuable.

Traditional scientific research tends to pursue the publication of “successful experiments” and ignores thousands of “negative results”. In the view of Fedus and Cubuk, these “failures” are precisely the key fuel for training AI scientists. “Every deviation in an experiment and every error feedback is an opportunity for the model to understand the physical world,” Cubuk said. “AI is not afraid of failure; it's only afraid of having no data.”

Therefore, Periodic Labs is not in a hurry to release results but focuses more on accumulating experimental data to build an unprecedented “scientific experience database” to lay the foundation for the next - generation scientific research AI.

Technology Stack: Synchronizing AI, Simulation, and Robots

In the laboratory of Periodic Labs, robotic arms are precisely mixing metal powders, high - temperature furnaces are heating up according to pre - set programs, and spectrometers are capturing real - time data on material properties — this is not a scene from a science - fiction movie but the future daily operation of its “autonomous laboratory”. The inspiration for this system comes from Cubuk's breakthrough research published in Nature in 2023: That year, the A - Lab platform he led synthesized 41 new compounds in 17 days, proving the feasibility of AI - driven experiments.

Paper link: https://www.nature.com/articles/s41586-023-06734-w

Now, Periodic Labs is taking this logic to the extreme. Its core innovation lies in its “trinity” science stack:

* Autonomous Robotic Lab: It can perform powder synthesis, substance mixing, and material preparation in a fully automated environment, precisely execute experimental instructions, and greatly improve the speed and repeatability of scientific research.

* High - Fidelity Simulation: Through AI - driven simulation technology, it can quickly evaluate physical and chemical reactions in a virtual environment and provide a high - precision hypothesis verification platform for experimental screening.

* LLM Research Assistant: The language model no longer just generates text but can analyze experimental data, propose correction suggestions, and design the next round of experiments, truly becoming the “cognitive center” of the scientific research process.

The three form a closed - loop system. First, an AI system that combines LLM and physical simulation analyzes the literature and generates experimental hypotheses; then, automated equipment performs synthesis and characterization, and each round of experiments generates several gigabytes of high - dimensional physical data; finally, the AI analyzes the results (whether successful or not) and optimizes the next - round plan. This cycle of “virtual deduction - physical verification - data feedback” completely subverts the traditional scientific research model and exponentially increases the speed of scientific discovery.

“Our real innovation is the way of data production,” Cubuk emphasized. Different from traditional AI that relies on Internet text, the unique data generated by its laboratory every day contains a large number of “negative results” that are ignored in traditional scientific research. In the field of materials science, failed experiments account for over 90%, and this valuable information not recorded in the literature has become the unique nourishment for Periodic Labs' AI models. As stated in the company's official website declaration: “Here, nature itself becomes the reinforcement learning environment.”

Behind the technical feasibility is the simultaneous maturity of three major fields: The precision of industrial - grade robotic arms has reached the 0.1 - millimeter level, which is sufficient for complex synthesis operations; AI - driven physical simulators can control the prediction error of material properties within 5%; and the reasoning ability of models such as o1 can handle interdisciplinary complex tasks like “designing superconducting crystal structures”. The combination of the three has turned Fedus' vision of “AI doing science hands - on” into a reality.

Capital Frenzy: The Silicon Valley Consensus behind $300 Million

In September 2025, the news that Periodic Labs had completed a $300 - million seed - round financing shocked the industry. This figure not only set a record for the seed - round financing of AI startups but also subverted the rules of the venture - capital industry — in addition to Andreessen Horowitz leading the investment, top institutions such as a16z, DST, and Nvidia NVentures all followed suit, and the list of angel investors was also star - studded: Amazon founder Jeff Bezos, former Google CEO Eric Schmidt, and Jeffrey Adgate, the key figure of DeepMind, were all on the list.

The prelude to this capital feast was full of drama. When Fedus announced his departure from OpenAI at the beginning of 2025, the Silicon Valley VC circle went into a collective frenzy: Some investors submitted dozens of pages of PPTs to promote themselves, some wrote “love - letter - style” investment intention letters, and some institutions promised to provide comprehensive support from computing power to the supply chain. But the first call they actually received was from Peter Deng, who was once Fedus' colleague at OpenAI and later became an investor at the top seed company Felicis. After listening to Fedus' vision, Deng even forgot that the company hadn't been registered yet and wanted to write a check.

Felicis investment blog. Image source: Felicis official website

The investors' frenzy was not blind. a16z said straightforwardly in its investment announcement: “This is an opportunity to compress decades of scientific research progress.” In multi - billion - dollar industries such as semiconductor heat dissipation and new - energy materials, the traditional R & D cycle often takes more than 10 years, while Periodic Labs' technical route is expected to shorten it to several years.

a16z investment blog. Image source: a16z official website

It's intriguing that its former employer, OpenAI, was absent. Although Fedus received blessings from the management when he left and was even hinted at possible support, even Sam Altman sent his blessings when the company was founded, it didn't appear on the list of investors in the end. Some industry analysts speculated that this might be due to fundamental differences in technical routes: OpenAI focuses on general artificial intelligence, while Periodic Labs' “AI for Science” vertical route is closer to Google DeepMind's strategic direction.

Dream Team: The Collective Migration of Half of Silicon Valley's Elites

After the $300 - million financing arrived, Periodic Labs launched one of the most astonishing talent recruitments in Silicon Valley history. In just a few weeks, more than 20 top researchers from Meta, OpenAI, and DeepMind joined collectively, including the inventor of the Transformer attention mechanism, the developer of the OpenAI Operator agent, and the creator of Microsoft's MatterGen large model. Many of them gave up millions of dollars in equity incentives just to participate in this “scientific research revolution”.

This team's cross - border characteristics are extremely rare: Half of the members come from the AI field, and the other half are experts in physics, chemistry, and materials science.

Team list. Image source: Periodic Labs official website

The luxurious advisory group further strengthens this cross - advantage. In the academic committee led by Nobel laureate Carolyn Bertozzi, there are both authorities in superconducting physics from Stanford University and masters in materials science from the Massachusetts Institute of Technology, providing new ideas for search algorithms to AI experts.

Scientific advisor list. Image source: Periodic Labs official website

Based on this powerful talent matrix, the company focuses its initial goal on discovering new high - temperature superconducting materials. Since all currently known superconductors need extremely low temperatures or high pressures to work, if a superconductor that can work at near - room temperature can be developed, its potential is huge. Periodic Labs bets that AI can accelerate the birth of this miracle.

In addition to superconductors, they are also looking at real - world problems in fields such as semiconductors. Currently, the team is collaborating with a chip manufacturer, using specially trained AI agents to optimize heat - dissipation materials and helping engineers iterate faster to solve the chip heat - dissipation bottleneck.

Conclusion

From the resolute decision to leave their jobs to the sensational $300 - million financing, from the conceptual collision of two scientists to the formation of a dream team spanning AI and physics, Periodic Labs has completed the evolution that traditional scientific research institutions take several years to achieve in less than a year. Its model of “AI scientist + automated laboratory” may not only lead to disruptive discoveries like room - temperature superconductors but also reshape the underlying logic of human exploration of nature.

As Sonal Chokshi, a partner at a16z, said: “Bell Labs once changed the world with transistors, IBM Research Institute opened up the future with laser technology, and Periodic Labs is reshaping science itself with AI