Periodic Labs, co-founded by former OpenAI VP of Research Liam Fedus and DeepMind materials science leader Ekin Cubuk, has emerged from stealth mode with a $300 million seed funding round, stunning Silicon Valley. However, OpenAI, their former employer, which had previously offered its blessing, did not participate in this round.
The former vice president of research at OpenAI who created ChatGPT and the leader of materials science and chemistry research at Google DeepMind have teamed up to start a business!
Their new company, Periodic Labs, secured $300 million in seed funding as soon as it launched, with a star-studded investment lineup led by a16z, followed by DST, Nvidia NVentures, Accel, Felicis, and other tech giants like Jeff Bezos, Eric Schmidt, and Jeff Dean.
Such exaggerated financing figures and investor array are extremely rare among start-ups and caused a sensation in the industry.
What exactly does this company do, and why has it attracted so much attention?
Leaving a top laboratory, just to "do real science"
One of the two co-founders is William Liam Fedus, former vice president of research at OpenAI and a core team member who participated in the creation of the groundbreaking ChatGPT;
The other is Ekin Dogus Cubuk (nicknamed "Doge"), who led materials science and chemistry research at Google DeepMind.
He is also one of the leaders of the GNoME project that shocked the academic community. In 2023, the project used AI to discover more than 2 million new crystal materials at once.
Logically speaking, they have already reached the top in their respective fields and have a bright future.
William Liam Fedus
As the head of OpenAI's post-training department, Liam Fedus mainly conducts research and development of ChatGPT, API and AI agent underlying models.
Previously, he worked at Google Brain, focusing on optimizing the efficiency of neural networks through MoE technology.
In 2022, he officially joined OpenAI and initially joined the reinforcement learning team as a core developer. He is one of the co-creators of ChatGPT and is mainly responsible for data processing and model evaluation.
During this period, he led the post-training research and development of several important models (including 4o, o1-mini, o1-preview, etc.).
In October 2024, Fedus replaced Barret Zoph and was promoted to head of the post-training team.
Chief Technology Officer Mira Murati and Chief Research Officer Bob McGrew also left the company at the time.
Fedus received a BS in physics from MIT (where he worked on the Directed Dark Matter Probe Project: DMTPC) and a BA in physics from the University of Cambridge.
In 2016, he received a master's degree in elementary particle physics from the University of California, San Diego, under the supervision of David Meyer and Gary Cottrell.
He then earned a PhD in Computer Science from the University of Montreal, where he studied under Yoshua Bengio and Hugo Larochelle.
Ekin Dogus Cubuk
Another co-founder, Ekin Dogus Cubuk, was previously a research scientist at Google DeepMind.
He joined Google Brain in 2017 and participated in GNoME, the flagship project in the field of materials science discovery. He also built multiple automated synthesis experimental platforms within the company, focusing on how to use AI to find new materials.
He holds a PhD in Condensed Matter and Materials Physics and Computational Science from Harvard University.
However, in March this year, Fedus resolutely resigned from OpenAI, and Cubuk also chose to leave DeepMind and turn to entrepreneurship.
It all started when we flipped tires together at Google
The two first met within Google, where they met during a fun tire flipping incident. But what really brought them together was a clear understanding of the limitations of current AI research paths and a shared pursuit of AI scientists.
At present, AI training mainly relies on Internet text, but although the Internet seems boundless, it is actually limited.
It is estimated that there are about 10 trillion tokens of valuable text data on the Internet (one English word has about 1-2 tokens), and the top large models have almost consumed this data in recent years.
Without fresh data, it is difficult to achieve a qualitative breakthrough simply by increasing the parameter scale without limit.
As Fedus stated in an interview:
The primary goal of AI is not to automate white-collar work. The primary goal of AI is to accelerate science.
In his view, the large-scale model applications that are currently being hyped in Silicon Valley are somewhat "intellectually lazy" and what AI should really focus on is accelerating the speed of scientific discovery.
Cubuk also pointed out that simply relying on large models and reasoning with text for days and nights will not produce groundbreaking scientific discoveries. True scientific breakthroughs require extensive experimentation and countless failures.
What the current AI models lack is precisely the “hands-on experiment” link.
So, the two hit it off at the beginning of this year: instead of being limited by existing data, it would be better to let AI "walk into" the laboratory and create data from scratch.
They want to create an "AI scientist" who can propose hypotheses and conduct repeated experiments in the real world, learning from the experimental results, regardless of whether the results are successful or not.
As Fedus said when communicating with investors:
To allow AI to truly do science, it must be allowed to do real science.
When Peter Deng, a former OpenAI colleague and current Felicis investment partner, heard this for the first time, he even stopped on the hillside in San Francisco and decided to invest immediately.
In his view, large models only grasp the "normal distribution" of training data, which is the knowledge that humans already have.
If we want to achieve original innovation, we must let AI step out of its comfort zone and propose new hypotheses and verify them like scientists.
This concept became the starting point of Periodic Labs.
Autonomous labs make nature an enhanced learning environment
Cubuk concluded that three major technological advances in recent years have made all this possible.
First, robotic arms capable of handling powder synthesis have become reliable, meaning machines can automatically mix raw materials and bake new materials.
Second, machine learning-driven physical simulations are more efficient and accurate, sufficient to simulate complex materials and chemical systems;
Third, the reasoning capabilities of large language models (LLMs) are now far superior to those of the past, enabling more complex planning and analysis.
The leaps in these three areas put together paint a picture: AI can make assumptions and calculations in the virtual world, put them into practice in the real world, and then analyze the experimental results and adjust its thinking.
Now is a good time to build an automated closed-loop laboratory for materials science.
In fact, Cubuk was one of the participants in this pioneering work.
As early as 2023, he and his colleagues published a paper in Nature, describing a fully automated robotic laboratory at Google: the AI language model proposed the experimental plan, and the robot synthesized materials based on it, synthesizing 41 new compounds that had never been recorded before in just 17 days.
This achievement is regarded as a milestone in autonomous AI scientific research, proving the feasibility of the technology.
The core of Periodic Labs is to build such an "autonomous lab".
It is a real physical experimental field, with robotic arms operating test tubes and materials, and sensors characterizing product properties. Every experiment generates massive amounts of first-hand data.
This process is like building a huge reinforcement learning environment for AI, and nature itself becomes its testing ground.
The model reads literature, runs simulations, and makes predictions about the properties of a certain material. The robot then synthesizes the material according to the plan and measures and verifies it. The experimental results either confirm the hypothesis or contradict it.
Regardless of the result, it provides a basis for the next step of improvement and achieves a true "closed loop".
Because each experiment is unique, this system will continuously produce new data that did not exist before, expanding the AI's knowledge base.
And it has an advantage that traditional scientific research cannot match: systematically recording failures.
In normal scientific research, a large number of "negative results" are ignored, and most of the results published in papers are successful cases, which leads to survivor bias in research.
Periodic Labs' independent laboratory regards every failure as valuable wealth, and failure itself becomes nourishment for model learning.
Over time, AI will accumulate a complete experience library covering success and failure cases, helping it explore the unknown more intelligently.
As the Periodic Labs website boasts: "Here, nature itself becomes the reinforcement learning environment."
Targeting superconductors and cutting-edge materials to tackle the multi-billion dollar challenge
Periodic Labs' choice to enter the field of physical science was not a whim.
On the one hand, data in the fields of physics and materials are relatively abundant and objectively verifiable. AI often makes rapid progress in fields with massive amounts of data and verifiable results (such as mathematical theorem proving and protein folding prediction).
On the other hand, the leaps in human technology are largely limited by breakthroughs in materials. Whoever is the first to find the "holy grail" of room-temperature superconductors will completely change the rules of the game.
Fedus and Cubuk knew this, so the company's initial number one goal was to discover new high-temperature superconducting materials. Currently known superconductors require extremely low temperatures or high pressures to work.
If a superconductor that works at near room temperature can be developed, it will have a revolutionary impact: zero-resistance power transmission, almost loss-free power grids, magnetic levitation trains, large-scale nuclear fusion devices, etc. are expected to become a reality.
Finding a room-temperature superconductor could be the next Nobel Prize-level achievement and the trigger for a multi-trillion dollar industry.
Periodic Labs is betting that AI can accelerate the birth of this miracle.
In addition to superconductors, they also turned their attention to practical problems in areas such as semiconductors.
The team is currently working with a chip manufacturer to use specially trained AI agents to optimize heat dissipation materials, helping engineers to iterate faster to solve chip heat dissipation bottlenecks.
In the future, this AI research platform can also be expanded to cutting-edge industries such as aerospace, energy, and national defense.
For example, accelerating the development of new alloys and heat-resistant materials can help humans move towards goals such as deep space exploration and controlled nuclear fusion at a lower cost.
In its investment announcement, a16z described Periodic Labs’ strategy as “landing and expanding in cutting-edge fields”:
We will first target industries like aerospace, defense, and semiconductors, which invest trillions of dollars in research and development each year. We will select key and difficult problems with clear evaluation criteria and huge value, and work with customers to overcome them using the AI laboratory to prove the power of this approach - to show the world how powerful AI can be when it is optimized directly for physical reality, rather than relying on internet text training.
Once we have established a firm foothold in these cutting-edge fields, we will gradually expand our capabilities to the broader scientific landscape.
In other words, win a few "hardcore" battles first, and then replicate and promote the victory experience.
Investors believe that if this step-by-step approach is successful, Periodic Labs will have the opportunity to leverage a massive market with a total output value of approximately $15 trillion, including advanced manufacturing, materials, energy, aerospace, etc.
Now that Moore's Law has hit a bottleneck, perhaps it is time for this new paradigm to take over and write the next chapter.
Frenzy of investment from top talent and capital
Such a grand vision naturally requires a dream team to match it.
Fedus and Cubuk used $300 million in "ammunition" to quickly recruit top global talents. In just a few weeks, more than 20 leaders in the field of AI research resigned from major companies such as Meta, OpenAI, and Google DeepMind to join Periodic Labs.
According to the New York Times, many people gave up millions or even tens of millions of dollars in salary to join this startup.
Their entrepreneurial team's resume is astonishing: not only does it include the co-creator of ChatGPT and the person who led the DeepMind materials project, but it also includes the inventor of the "attention mechanism" of the Transformer neural network, the developer of OpenAI's early intelligent body Operator (Agent), and the creator of Microsoft's large-scale material science model MatterGen.
This combination includes almost all the top experts in the fields of AI algorithms and physical sciences.
To enable experts from different backgrounds to work together, Periodic Labs holds graduate-level cross-disciplinary lectures every week: this week a physicist explains the logic of quantum mechanics, and next week a machine learning expert trains colleagues on cutting-edge AI models, so that all employees have a deep understanding of each other's fields.
The company has also formed a prestigious scientific advisory committee, including Nobel Prize winner Carolyn Bertozzi and many other highly respected professors in the fields of chemistry and physics, to oversee research directions.
It can be said that Periodic Labs established a team that is a rare fusion of academia and industry at the beginning of its establishment.
As investor a16z commented:
Liam and Doge have brought together a unique team: physicists, chemists, simulation experts, and some of the world’s leading machine learning researchers.
Such a team configuration is rare in large model companies, let alone a startup.
On the other hand, the enthusiasm in the investment community is equally high.
When Fedus announced that he was leaving OpenAI to pursue a new venture, Silicon Valley VCs almost went into a frenzy of scrambling.
It is said that one investor wrote a heartfelt "love letter" to Periodic Labs to express his love for the company, while others submitted a multi-page PPT to introduce the value they could provide, hoping to impress the founder.
The first person to knock on their door was Peter Deng, a partner at Felicis Venture Capital. Coincidentally, he had worked with Fedus at OpenAI and also left to become a VC at the beginning of this year.
Deng quickly made an appointment with Fedus to have coffee in the Noe Valley neighborhood of San Francisco, and they excitedly chatted while walking.
When he heard Fedus's vision, he was so excited that he said on the spot: "I will write you a check right now!"
He even forgot at one point that the company had not yet been registered, had not even chosen a name, and had nowhere to send the check.
Ultimately, Felicis successfully secured one of the lead investors in this round.
However, OpenAI itself did not participate in the investment.
Although OpenAI executives gave their blessing when Fedus left, and he even hinted in a tweet that OpenAI might support him, this assumption did not come true.
However, OpenAI's absence doesn't matter.
Thanks to the reputation of the founding team, Periodic Labs soon received a dizzying number of investment offers.
In addition to the aforementioned institutions, the follow-up investment list also includes Silicon Valley's biggest angel investors: Amazon founder Bezos, former Google CEO Schmidt, AI legend Jeff Dean, Silicon Valley investor Elad Gill, etc. Even NVIDIA has come out to support through its fund NVentures.
Almost half of the "god circles" in Silicon Valley have gathered here, which shows how high everyone's expectations are for the AI4S track.
Investors have stated that this could be an opportunity to "compress the scientific research process by decades," and no one wants to miss it.
AI research competition: Giants and startups compete on the same stage
The emergence of Periodic Labs marks an important shift in the landscape of AI exploration: from pursuing "virtual intelligence" such as general artificial intelligence and chatbots to a new arena of deep interaction with the physical world and the goal of creating new scientific knowledge.
They are not alone. Similar ideas are emerging within large technology companies.
Just last month, OpenAI announced the establishment of the "OpenAI for Science" department, attempting to create "AI-driven next-generation scientific instruments" to allow AI platforms to accelerate scientific discoveries.
DeepMind has already set a precedent: its AlphaFold system solved the difficult problem of biological protein folding, changing the landscape of biological research in one fell swoop, and not only won global reputation, but also won the Nobel Prize for its two core developers.
It can be said that technology giants have realized that the next AI breakthrough may very well be in the laboratory and in unknown phenomena in the real world.
In the startup camp, in addition to Periodic Labs, there are also new non-profit organizations like FutureHouse, which have pledged to create independent AI scientists.
Various signs indicate that "using AI for scientific research" is becoming a hot topic in the new round of innovation competition.
In contrast, Periodic Labs has a rare focus and luxurious team in the industry, with ample funds and no burden of a large company. It is regarded as one of the most promising players in this field.
Of course, scientific research is never achieved overnight, let alone using AI to challenge difficult problems in areas unknown to humans.
Periodic Labs itself admits that this is a high-risk gamble.
Even with the smartest AI and the most capable robots, the scientific research process is still full of unpredictable twists and turns, and nine out of ten attempts may fail.
But precisely because it is full of uncertainty, it means huge room for breakthroughs and value. Even if they fail to find the ideal superconductor in the end, the large amount of data and lessons learned from failure accumulated in the exploration itself are of extraordinary significance.
This is in contrast to traditional scientific research, which is overly utilitarian and only rewards the publication of successful papers. Periodic Labs pursues a new paradigm in which "exploration itself is valuable."
It can be foreseen that if this path is successful, our approach to scientific research will also be overturned.
Periodic Labs is reimagining how scientific discovery is conducted.
When AI truly enters the laboratory and when the natural world becomes a training ground for AI, perhaps what humanity will usher in will be a leap in the scientific research paradigm.
With this belief, this group of top scientists and investors are fully committed to this gamble.
We will have to wait and see whether Periodic Labs can deliver a world-changing answer in the next decade.
References:
https://techcrunch.com/2025/10/20/top-openai-google-brain-researchers-set-off-a-300m-vc-frenzy-for-their-startup-periodic-labs/
This article comes from the WeChat public account "Xinzhiyuan" , author: Xinzhiyuan, and is authorized to be published by 36Kr.
