Periodic Labs was built to discover new materials with AI. Liam Fedus, who built the company with Ekin Doğuş Çubuk, says it has also been its own first customer. On an episode of the Latent Space podcast published October 8, 2026, Fedus told hosts swyx and Brandon that Periodic is now taking those internal tools into other industries. He said its biggest current focus is the semiconductor industry.
According to the publisher's episode notes, the company launched in September 2025. Much of the conversation covered its science. One part answered a simpler question: how does a lab like this make money?
Customer zero
"We've been building our own products, our own tools, our own AI and computational for ourselves. So we've been customer zero," Fedus said. In this case, "computational" means the simulation software the company uses to predict how materials behave. He said the tools and know-how from Periodic's own materials discovery and engineering are now going to industrial partners.
Periodic's website gives one example. It describes a semiconductor manufacturer working on how chips get rid of heat. The company says it is training custom AI agents to help that manufacturer's engineers and researchers interpret experimental data and try new solutions faster. An AI agent is a system that can carry out multistep tasks rather than only answer questions.
Engineers who go onsite
Fedus said the partners are very private companies, so Periodic has to work inside "some of the most secure environments." That shapes the business model. Periodic does not just sell access to a model over the internet. Fedus said its forward-deployed engineers and researchers will work at the partner's site. Forward-deployed engineers are technical staff who work inside a customer's operations instead of at the vendor's office. He said they integrate Periodic's AI and computational systems so partners reach their goals more quickly, instead of "just like throwing some technology over the wall and telling them to, to figure it out."
These engineers also run the AI locally. In technical terms, they do inference on the partner's side, meaning the trained model produces its answers there. They can also train the model further on the partner's data. Fedus said the aim is to make Periodic's system expert in each partner's data "so that they can own their own intelligence." He contrasted this with "just like hitting some API or some untrained model." An API is the standard way software calls a model hosted by another company.
Fedus said this makes the job much broader than a typical forward-deployed engineering role. It needs deep machine-learning expertise and precision on data training, infrastructure and physics. The engineers also have to understand the chemistry, materials science and devices in highly technical industries. A remark in the same exchange said the role includes a willingness to spend extended periods onsite in Taiwan. The remark also described this onsite work as Periodic's main way of earning money for now, a model that might change.
From copilots to paying for outcomes
Fedus compared the path ahead with software engineering. At first, developers used tools such as GitHub Copilot and early versions of ChatGPT as copilots: assistants that helped them reach their own solutions. "As the automation improved, now we have things like codex and very few of our engineers are writing code the way they used to," he said. Codex is OpenAI's coding agent.
Fedus thinks something similar "could play out" at Periodic. Its first systems speed up the work of human materials scientists, materials engineers and process engineers. As the systems' autonomy and capability grow, he said, Periodic could "begin to, price outcomes." A customer would pay for a request like "help me get to this kind of goal state" rather than for help along the way. Fedus called this "a really interesting area" for the company. He presented it as a possibility, not current practice.
Why commercial success matters to the science
The conversation turned to Periodic's frequent comparison with Bell Labs, AT&T's research arm, and then to money. One remark suggested that Periodic's best contribution to solid-state physics might be making the field's tools profitable, in the way ChatGPT made computer science and language-model studies more popular in universities. Commercial success would draw attention and might bring more young people into physics. The same remark noted Bell Labs' mixed record. It failed to commercialize some major advances, and some of its research, such as the cosmic background, could never be sold. But it benefited from the vacuum tube that connected East and West Coast by phone lines.
That history fits the lab's best-known invention. A Nokia Bell Labs history describes the transistor as an answer to a practical telephone problem: replacing bulky, power-hungry vacuum tubes that amplified phone signals. John Bardeen and Walter Brattain, working under William Shockley, demonstrated a germanium point-contact transistor on December 16, 1947.
Fedus made the economic argument directly. "Technology and capital are incredibly intertwined," he said. He asked listeners to compare chatbot progress from 2010 to 2015, when it would be hard to say how much they improved, with 2021 to 2026, which he called "night and day." The difference, he said, was that ChatGPT and similar systems found product-market fit, meaning a product that many people want to use and pay for. That "changed the capital landscape entirely," which in turn changed hiring, computing resources and data. The discussion described this as a self-reinforcing cycle in which commercial success brings more resources, and Fedus agreed. "We want to achieve the same thing in the physical world," he said.
Open-source code and academic grants
A question in the conversation noted that science funding is shifting from government grants toward venture capital and private money, and asked whether Periodic would release anything, such as datasets or models. The answer focused on code. The discussion listed contributions to several open-source tools:
- pymatgen, the Python Materials Genomics library for representing and analyzing materials
- Custodian, which supervises large simulation jobs, detects common failures, applies fixes and restarts them
- TorchSim, a PyTorch-based engine for simulating atoms, including with machine-learning models
- JAX MD, a molecular-dynamics library built on the JAX framework
According to the discussion, Periodic researcher Abhijeet Gangan created TorchSim and still maintains it, and he also maintains JAX MD. Gangan is listed on the company's team roster. "We've been very active contributors back into open source," Fedus said.
The discussion also described an academic grant program, with gift grants to university groups whose work advances what Periodic calls "synthesis superintelligence," its goal of AI that can design and make new materials. According to the discussion, the first paper from that funding was about to come out.
Experts with their hands on the work
The Bell Labs comparison also shapes hiring. The conversation described a guiding principle: experienced people should keep doing hands-on work. Academia often loads strong researchers with grant writing and teaching until they no longer have time for research. The discussion pointed instead to Bell Labs and IBM, where Bardeen was in the lab every day although he was a theorist, and Alex Müller ran experiments even as lab lead. The discussion said Periodic tries to follow the same principle, with leading figures in different fields who still do hands-on work.