Periodic Labs emerges from stealth to build AI scientists for materials discovery
Key Points
- Periodic Labs raises $300 million and emerges from stealth with AI agents designed to run physical experiments in materials science, not just reason about it.
- Co-founders Liam and Dogus, who built ChatGPT and led deep learning work at Google Brain, are targeting superconductors above 200 Kelvin by grounding AI agents in lab feedback loops.
- The startup will commercialize custom materials for semiconductor, space, and defense customers while using research-generated lab data to train its scientific AI models.
Summary
Periodic Labs is a six-month-old startup built on a simple but ambitious premise: AI agents that don't just reason about science but actually run experiments. The company raised $300 million and emerged from stealth with two founders who between them helped build ChatGPT and some of the first trillion-parameter neural networks at Google Brain.
The founders
Liam (William) spent years at Google Brain working on generative models and reinforcement learning before joining OpenAI in late 2022, where his team turned a low-key research preview into ChatGPT. The internal pre-launch poll had people guessing the product might reach a million users; it crossed that number in a week. His co-founder Dogus has a PhD background in applying machine learning to physics and spent years at Google after 2020 leading a team focused specifically on bringing deep learning advances into materials science. They left their jobs six months ago to start Periodic.
The core thesis
The argument is that LLMs are now genuinely capable at math and logic, and the next productive frontier is theoretical physics — specifically quantum mechanics at the level of how atoms bond and what properties emerge. The critical design choice is that Periodic's AI agents are connected to physical labs. The lab is the grading function. Without experimental data in the loop, the agents are just reasoning in the abstract, and as any experimentalist knows, an untested hypothesis is not science.
The agents are being trained to synthesize materials, characterize them, predict their properties, and propose the next experimental step. That lab-grounded feedback loop is what Periodic believes separates meaningful scientific AI from sophisticated speculation.
Superconductors as a near-term goal
The highest-temperature ambient-pressure superconductor currently sits at around 135 Kelvin. Periodic is targeting materials that could push that toward 200 Kelvin. The commercial applications are significant — fusion, low-loss energy transmission, quantum computing, and next-generation chip design all depend on advances here. Dogus notes that their lab characterization tools would have caught the LK-99 false alarm quickly, since they can measure material properties across temperatures directly rather than relying on indirect signals.
Commercial structure
Periodic is positioning itself to serve semiconductor companies, space companies, and defense contractors looking for custom materials — heat shields, novel substrates, anything that requires intentional atomic-level design. The lab data generated during scientific exploration also serves as the training foundation for customer-facing work, so the research and commercial tracks reinforce each other rather than compete for resources.
The company plans to publish selectively when it advances the broader community and has launched an academic grant program to support external researchers working in adjacent areas.
The bigger bet
Liam's framing is that as the cost of intelligence falls, the bottleneck for progress shifts to contact with the physical world. Models that can reason are plentiful; models that can run an experiment, read the result, and design the next one are not. Periodic's argument is that bringing atoms into the AI loop is what converts the current wave of reasoning capability into actual new materials, new devices, and compounding scientific knowledge.