Lab to Company · LAB TO COMPANY 006 · Researchers leave a frontier lab
Periodic Labs: two researchers leave OpenAI and Google DeepMind, carrying experience, not a licence
The question · When researchers leave a frontier lab to found a company, where do its research, its lab and its money come from?
At a glance · A $300M founding round led by a16z; public IP agreements with OpenAI or Google: none; 23 people and 5 scientific advisers on the company's site.
One of the Lab to Company series. Each piece reads the public record through five roles: who supplies the research (the results, methods or team it came from), who holds the IP (patents, licences, what a parent kept), who gives the lab and equipment, who puts in the first money and on what terms, and who joins the company. It records only what the public record states and says who said each thing; a missing step is marked not yet public, and a company's own figure is written as the company's.
Periodic Labs was founded by Liam Fedus and Ekin Dogus Cubuk. Fedus was OpenAI's VP of research for post-training and helped create ChatGPT; Cubuk led the materials and chemistry team at Google Brain and Google DeepMind and is a GNoME author.[1][2] On 30 September 2025 the company launched with $300M led by a16z, to build "AI scientists" that synthesise and measure materials in its own labs and learn from the data.[3][4][5][13] Unlike a team spun out of a large company, nothing was handed over by a parent: what the record shows is the founders' experience, open-source tools and public data, plus investors' money.
Who supplies the research · The founders' earlier work: Fedus on ChatGPT's post-training at OpenAI, Cubuk on GNoME and materials and chemistry research at Google.[1][2] The company says its founding team co-created ChatGPT, GNoME, OpenAI's Operator, the neural attention mechanism and MatterGen.[3] Its first paper after founding (arXiv 2508.19781) screened Google's public GNoME database for superconducting hydrides by computation, with Cubuk listed at Periodic Labs; the company's own results since (the September 2026 Neon model and others) are scored in-house.[6][7]
Who holds the IP · Any licence or assignment from OpenAI or Google: not yet public. Leaving OpenAI in March 2025, Fedus wrote that OpenAI was "planning to invest in and partner with my new company"; TechCrunch reported in October that the investment did not happen, which the founders confirmed without saying why.[1][2] GNoME's data is public under a non-commercial licence (CC BY-NC 4.0).[8] Fedus says the company is weighing outcome-based pricing in which it keeps the IP in what its systems produce.[9]
Who gives the lab and equipment · The company builds its own. It says it has high-throughput labs running around the clock in Menlo Park; its job board lists a powder synthesis lab, an X-ray diffraction facility and a thin-film lab being set up, with thin-film work done meanwhile at unnamed outside partner facilities.[5][11] In October 2025 TechCrunch reported the lab set up but the robots not yet running.[2] Cubuk's target is about a thousand experiments a day within roughly a year; no achieved throughput is public.[9]
Who puts in the first money, and on what terms · The first cheque came from Peter Deng of Felicis, who says he committed before the company had a name or was incorporated.[12] The $300M founding round was led by a16z, with Felicis, DST Global, NVIDIA's NVentures, Accel and individuals including Jeff Bezos, Eric Schmidt and Jeff Dean; Fedus and a16z each state the amount and the lead themselves.[4][13] The $1.3B valuation is reported by Forbes alone; Lightspeed's own page records its 2025 investment as Series A, which does not match the "seed" label elsewhere.[14][15] Preferences and board composition are not public, except that a16z's Anjney Midha sat on the board at the time.[13]
Who joins the company · The company's site lists 23 people, without titles, and five scientific advisers: Carolyn Bertozzi, Mercouri Kanatzidis, Steve Kivelson, Zhi-Xun Shen and Chris Wolverton.[3] In interview the founders name MIT professor Joe Checkelsky, on leave to run the lab full time (not confirmed with MIT), and Dzmitry Bahdanau, who proposed the neural attention mechanism, leading AI and language models.[10] TechCrunch named early hires Alexandre Passos, a creator of o1 and o3, and Matt Horton, a creator of Microsoft's generative materials tools; neither is on today's list.[2]
What each side put in
- The founders put in experience and reputation. Nothing on the record moved over from OpenAI or Google; the GNoME data and open-source tools it uses are open to anyone.[3][8]
- The investors bet before the company had a name, and what they bought was a lab to be built from nothing: from TechCrunch's "robots not yet running" to the company's "running 24/7" took about a year.[2][5][12]
- The company put forward a pricing claim: keep the IP in the materials it finds, rather than only sell a tool.[9]
Not yet public · Any agreement with OpenAI or Google; the incorporation date; the founding round's valuation (only a media figure), preferences and board seats; whether the 2026 round closed; lab size and achieved throughput; a named customer; any new material measured outside the company.
Decision point · The next decision belongs to the investors in a new round. Forbes reported in May 2026 that talks were advanced, at least $500M at a $7.5B valuation led by AMP; neither the company nor an investor has announced an outcome.[14] Until then, every result the company has published is scored in-house.
The chain's timeline
- 2025-03Fedus announces he is leaving OpenAI and says OpenAI plans to invest (his own post)[1]
- Spring 2025Felicis commits the first cheque before the company has a name (investor's blog)[12]
- 2025-08First paper, Cubuk listed at Periodic Labs (arXiv)[6]
- 2025-09-30Launch; $300M founding round led by a16z (company, founder, a16z)[3][4][13]
- 2025-10TechCrunch: OpenAI did not invest; lab built, robots not yet running[2]
- 2026-05Forbes: talks on a new round, at least $500M at $7.5B[14]
- 2026-09The company publishes its Neon model and says its Menlo Park labs run 24/7[5][7]
- NextWhether the new round closes; the first material result measured outside the company. Awaiting public data.
- LaterA named customer and revenue; whether the keep-the-IP pricing takes hold.
Sources
- Liam Fedus, X post on leaving OpenAI, 2025-03-17. x.com ↗ (read in full through fxtwitter)
- TechCrunch, 2025-10-20. techcrunch.com ↗ (read in full)
- Periodic Labs, "Introducing Periodic Labs" and the homepage team list, 2025-09-30. periodic.com ↗ (read in full)
- Liam Fedus, X post on the round, 2025-09-30. x.com ↗ (read in full through fxtwitter)
- Periodic Labs, "Building labs that learn", 2026-09-15. periodic.com ↗ (read in full)
- Sanna et al., arXiv 2508.19781, 2025-08-27. arxiv.org ↗ (first page read)
- Periodic Labs, "Nature is our learning environment", 2026-09-15. periodic.com ↗ (read in full)
- Google DeepMind, GNoME repository README (licence section). github.com ↗
- The Generalist interview transcript, 2026-09-29. generalist.com ↗ (relevant passages found by keyword)
- Latent Space interview transcript, 2026-10-08. latent.space ↗ (relevant passages found by keyword)
- Periodic Labs job board. jobs.ashbyhq.com ↗ (lab-related postings partly read)
- Felicis, Peter Deng's post. felicis.com ↗ (read in full)
- a16z, "Investing in Periodic Labs", 2025-09-30. a16z.com ↗ (read in full)
- Forbes, 2026-05-13 (Forbes Australia's copy). forbes.com.au ↗ (read in full)
- Lightspeed portfolio page. lsvp.com ↗ (read in full)