Key Takeaways
- • Andreessen Horowitz is leading a $300 million Series A funding round for Periodic Labs at a $1.2 billion valuation, positioning the startup as a major player in AI-driven scientific discovery.
- • Periodic Labs was founded by Liam Fedus (former head of post-training at OpenAI and co-creator of ChatGPT) and Dogus Cubuk (from Google DeepMind's compound discovery initiative), bringing elite AI researcher pedigree to the venture.
- • The startup addresses a critical gap in frontier AI: current models excel at math, code, and logic but fail substantially at physics, chemistry, and applied physical sciences—fields essential for drug discovery, semiconductors, and advanced manufacturing.
- • Periodic Labs is building "AI scientists" paired with autonomous laboratories, moving beyond text-trained models to incorporate real-world experimental data—a fundamental limitation acknowledged by the founders after exhausting internet-scale training datasets (approximately 10 trillion tokens).
- • The funding syndicate includes Felicis, DST, and Accel alongside a16z, indicating institutional confidence in the commercial potential of AI-augmented scientific research in high-value industries.
Summary
Andreessen Horowitz has led a $300 million Series A funding round for Periodic Labs, a newly formed startup tackling a structural weakness in contemporary AI systems: their inability to perform meaningful scientific reasoning in the physical sciences. The investment values the company at $1.2 billion, marking a significant institutional bet on AI-driven materials science and experimental discovery.
Periodic Labs was founded by two veteran AI researchers with exceptional track records. Liam Fedus served as head of post-training at OpenAI and was a co-creator of ChatGPT, while Dogus Cubuk led the team at Google DeepMind that pioneered the GNOME approach to scalable compound discovery using AI. Both founders independently arrived at the same conclusion: frontier AI models, despite remarkable progress in mathematical reasoning and code generation, are objectively worse than human investigators at analyzing scientific experimental data, particularly in condensed matter physics—a field underpinning semiconductors and advanced manufacturing.
The core insight driving the company's strategy centers on a fundamental limitation in current AI training methodologies. State-of-the-art models have been trained on approximately 10 trillion tokens of internet text, effectively exhausting publicly available linguistic datasets. This training regime produces systems that excel at abstract reasoning but lack the grounding in real-world physical phenomena necessary for scientific discovery. Periodic Labs' founders determined that the missing ingredient was empirical experimental data—observations that cannot be adequately synthesized from text alone.
The company's approach diverges from conventional AI applications: rather than deploying pre-trained language models to analyze scientific text, Periodic Labs is building autonomous laboratory systems paired with purpose-built AI scientists capable of interpreting and learning from actual experimental results. This architecture addresses the bottleneck in modern science, where the pace of discovery is constrained by the time and cost required to conduct physical experiments and interpret their outcomes.
The funding syndicate—including Felicis, DST, and Accel—reflects confidence in the commercial scale of the opportunity. AI-augmented scientific discovery has immediate applications in pharmaceuticals, materials science, and industrial chemistry, sectors where accelerating the experimental cycle translates directly to competitive advantage and revenue impact. As of September 2025, Periodic Labs remains a private company without public market access.
Periodic Labs was founded by two veteran AI researchers with exceptional track records. Liam Fedus served as head of post-training at OpenAI and was a co-creator of ChatGPT, while Dogus Cubuk led the team at Google DeepMind that pioneered the GNOME approach to scalable compound discovery using AI. Both founders independently arrived at the same conclusion: frontier AI models, despite remarkable progress in mathematical reasoning and code generation, are objectively worse than human investigators at analyzing scientific experimental data, particularly in condensed matter physics—a field underpinning semiconductors and advanced manufacturing.
The core insight driving the company's strategy centers on a fundamental limitation in current AI training methodologies. State-of-the-art models have been trained on approximately 10 trillion tokens of internet text, effectively exhausting publicly available linguistic datasets. This training regime produces systems that excel at abstract reasoning but lack the grounding in real-world physical phenomena necessary for scientific discovery. Periodic Labs' founders determined that the missing ingredient was empirical experimental data—observations that cannot be adequately synthesized from text alone.
The company's approach diverges from conventional AI applications: rather than deploying pre-trained language models to analyze scientific text, Periodic Labs is building autonomous laboratory systems paired with purpose-built AI scientists capable of interpreting and learning from actual experimental results. This architecture addresses the bottleneck in modern science, where the pace of discovery is constrained by the time and cost required to conduct physical experiments and interpret their outcomes.
The funding syndicate—including Felicis, DST, and Accel—reflects confidence in the commercial scale of the opportunity. AI-augmented scientific discovery has immediate applications in pharmaceuticals, materials science, and industrial chemistry, sectors where accelerating the experimental cycle translates directly to competitive advantage and revenue impact. As of September 2025, Periodic Labs remains a private company without public market access.
Go deeper than the headline
Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.