Periodic Labs

San Francisco "AI scientist" startup — autonomous robotic labs to automate physics and chemistry discovery; $300M seed, talks for a $500M round [1][2].

1. Core Product / Service

Periodic Labs (San Francisco) is building autonomous "AI scientists" — robotic labs driven by large language models that design, run, and interpret experiments in a closed loop, automating scientific discovery [1]. Its initial target is higher-temperature superconductors (running thousands of experiments in search of new materials), plus work with the semiconductor industry [1].

The founding team is elite: Liam Fedus (former VP of Research at OpenAI, a ChatGPT co-creator) and Ekin Dogus Cubuk (former Google Brain/DeepMind materials scientist, creator of the GNoME discovery tool) [1]. The company has hired 20+ researchers from Meta, OpenAI, and DeepMind [2].

Beyond lab-as-a-service, Periodic has shipped a concrete model: Neon, a ~1-trillion-parameter scientific model post-trained from the open-weight Kimi K2.6 base (not a from-scratch pre-train), specialized for X-ray diffraction (XRD) phase identification [3][4]. Periodic ran scientific mid-training over academic literature, code and proprietary lab data, then long-context multimodal RL on its own experimental results [3]. Neon reads diffraction patterns together with synthesis conditions, thermodynamic reasoning, crystallographic databases and prior experiments, proposes candidate crystalline phases, and iteratively tests its interpretation — automating phase identification in multi-phase, overlapping-peak samples [3]. It is deployed in Periodic's Menlo Park high-throughput materials lab [3].

2. Target Users & Pain Points

• Industrial R&D (semiconductors, energy, materials) seeking faster discovery than manual experimentation.
• Frontier AI labs collaborating on science automation — local research notes Periodic Labs "works with multiple AI Labs" [local].

Pain solved: the bottleneck of physical-experiment throughput — replacing human-run, one-at-a-time experiments with robots running thousands of trials and feeding results back to the model [1].

3. Competitive Landscape

Periodic Labs' differentiation is combining frontier-LLM research pedigree (Fedus) with materials-science automation (Cubuk) — a bet that an "AI scientist" needs both world-class model people and domain experimentalists [1].

4. Unique Observations

• The valuation arithmetic is the headline. A $300M seed at $1.3B (Sept/Oct 2025), then talks for ≥$500M at ~$7.5B within eight months — a nearly 6× repricing with, by all accounts, no commercial product yet [1][2]. It is the clearest sign of how hard investors are bidding the autonomous-science thesis, mirroring lila-sciences' ~$8.5B mark.
• Founder talent as the moat. Fedus (OpenAI VP Research) and Cubuk (GNoME) signal that the company can actually build the loop — hired researchers reportedly gave up substantial equity to join [1][2].
• Local research flags Periodic as a US benchmark for AI4S automated labs (alongside lila-sciences and radical-ai), against which Chinese and European players are being compared [local].
• Post-training on an open Chinese base model is the tell. Neon starts from Kimi K2.6 rather than a proprietary base — a US frontier lab renting capability from an open-weight Chinese model and spending its compute on scientific mid-training and RL rather than pre-training. Jimmy's late-September research tracked this as the "Periodic Labs × Kimi" post-training case [local: 2026-09-21-summary.md][local: 2026-09-23-summary.md].
• FrontierXRD: 2.7% → 55.3%, with caveats. On Periodic's internal FrontierXRD benchmark (134 hard real samples, ~5 phases each), success rose from 2.7% for untuned Kimi K2.6 to 55.3% for Neon — a ~20× gain, and Periodic claims it beats GPT-6 Astra and Claude Fable 5.1 under the same scientific harness [3][4]. Both the benchmark and the "beats GPT-6" claim are first-party and unvalidated, and ~45% of the hardest samples remain unsolved [3]. Training peaked at ~1,300 H200 GPUs with 4.1× Megatron-baseline throughput and 95%+ cluster utilization [4].

5. Financials / Funding

• Seed (Sept/Oct 2025): $300M at a $1.3B valuation, from a16z, DST, Nvidia (NVentures), Accel, and Felicis; angels Jeff Bezos, Eric Schmidt, Jeff Dean, and Elad Gil [1].
• Next round (May 2026, in talks): ≥$500M at a ~$7.5B valuation, led by AMP (Anjney Midha's vehicle); reported significantly oversubscribed, with a fast-follow round at a higher valuation in discussion [2].

6. People & Relationships

• Co-founders: Liam Fedus (ex-OpenAI VP Research), Ekin Dogus Cubuk (ex-Google Brain/DeepMind, GNoME) [1].
• Team: 20+ researchers from Meta, OpenAI, and DeepMind [2].
• Investors: a16z, DST, Nvidia (NVentures), Accel, Felicis; angels Jeff Bezos, Eric Schmidt, Jeff Dean, Elad Gil [1].
• Peers: lila-sciences, radical-ai, cuspai, orbital-materials.