Periodic Labs

Periodic Labs is an American artificial intelligence research company that builds AI models and autonomous laboratories aimed at accelerating the discovery of new physical materials. The company emerged from stealth on September 30, 2025 with the announcement of a $300 million seed round, an unusually large first financing that TechCrunch later said had set off a "VC frenzy" among investors.^[1][2][14] Periodic Labs was co-founded by Liam Fedus, the former vice president of research for post-training at OpenAI and a co-creator of ChatGPT, and Ekin Dogus Cubuk, who previously led the materials and chemistry research team at Google Brain and Google DeepMind and whose team at Google DeepMind developed the GNoME materials-discovery system.^[1][2][3]

The company is one of several well-funded "AI for Science" startups attempting to apply large neural networks, robotics, and automated experimentation to problems in chemistry, physics, and biology. Periodic Labs has publicly stated that its initial research goal is the discovery of new superconductors, with adjacent commercial work on heat dissipation in semiconductor chips.^[1][4][13] In September 2026 it introduced Periodic Neon, a model of about one trillion parameters that it post-trained on data from its own laboratories in Menlo Park, California, to analyze X-ray diffraction measurements.^[16][17] In May 2026 Forbes reported that the company was in advanced talks to raise at least $500 million at a $7.5 billion valuation; no close of that round had been announced as of October 2026.^[15][35][31]

Founding

Periodic Labs was founded in 2025; Tech Funding News and Contrary Research both give the founding date as May 2025 and the place as San Francisco.^[11][25] Public reporting first surfaced when Fedus announced his departure from OpenAI on March 17, 2025 to start an AI materials-science company, and the firm formally emerged from stealth approximately six months later.^[6] In a note to OpenAI colleagues that he also posted publicly, Fedus described the move as a difficult decision and said he intended to work closely with OpenAI as a partner; he also said that OpenAI was "planning to invest in and partner with my new company," describing AI for science as one of OpenAI's most strategically important areas.^[6] That investment did not happen. In October 2025 the founders confirmed to TechCrunch that OpenAI was not a backer of Periodic Labs, and Fedus declined to say why.^[1]

According to TechCrunch, the company began with a conversation between Fedus and Cubuk roughly seven months before October 2025. Cubuk told the publication that three developments made the timing right: robotic arms for powder synthesis (the process of mixing and creating new materials) had become reliable, machine-learning simulations had become efficient and accurate enough to model complex physical systems, and large language models had gained strong reasoning abilities.^[1] Felicis partner Peter Deng, a former OpenAI colleague of Fedus, committed to the first check so early that the firm had to wait to sign the term sheet, because the company had not yet been named or incorporated.^[1][10]

Press coverage has generally described Periodic Labs as a San Francisco-based company.^[15][19] Its laboratories are in Menlo Park, California: The Logic reported in October 2025 that it was hiring engineers and researchers for a new lab there, and the company said in September 2026 that it had built high-throughput labs in Menlo Park.^[21][16] CB Insights lists the company's headquarters in Menlo Park, and every role on its job board lists Menlo Park as a location.^[5][32]

Founders

Liam Fedus (also published as William Fedus) is a co-founder of Periodic Labs.^[1][22] Before founding the company, he served as vice president of research in charge of post-training at OpenAI, the stage in which models such as those that power ChatGPT are refined after their initial training.^[6] Fedus joined OpenAI in 2022 after working at Google, and was promoted to vice president in the autumn of 2024.^[6] He is widely credited as one of the co-creators of ChatGPT, and Felicis has said he also drove advances in OpenAI's Operator agent; he earned his PhD at Mila.^[1][3][10] During his earlier time at Google Brain he was the lead author, with Barret Zoph and Noam Shazeer, of the Switch Transformer paper, which simplified mixture-of-experts routing and pre-trained sparse transformer language models of up to a trillion parameters.^[29] TechCrunch has described him as having led the team that created the first trillion-parameter neural network.^[14] Fedus completed his undergraduate studies in physics and has cited that background as a personal motivation for starting an AI-for-science company.^[6][7]

Ekin Dogus Cubuk is Periodic Labs' other co-founder.^[1][22] Before co-founding Periodic Labs, he was a research scientist at Google DeepMind,^[15] and he led the materials and chemistry team at Google Brain and Google DeepMind;^[14] Berkeley Lab described him in 2023 as the lead of Google DeepMind's Materials Discovery team.^[9] He earned his PhD at Harvard University and completed a postdoctoral fellowship at Stanford University.^[10] Cubuk was the last-listed (senior) author of the 2023 Nature paper that introduced GNoME (Graph Networks for Materials Exploration), a graph neural network system used to predict the stability of inorganic crystals; the project reported 2.2 million predicted new crystal structures, of which about 380,000 were classified as most stable, and Google DeepMind said external researchers had independently created 736 of the new structures experimentally in concurrent work.^[8][9][26]

Cubuk was also a co-author of a second Nature paper published the same day, which described the A-Lab, an autonomous laboratory at Lawrence Berkeley National Laboratory that used computations, literature-trained models, machine learning, active learning, and robotics to synthesize inorganic powders. It originally reported 41 novel compounds from 58 targets over 17 days of operation.^[27][9] After concerns were raised about whether the compounds had been unambiguously identified from diffraction data and whether they were new, the authors published a correction in January 2026. A manual re-analysis confirmed 36 of 40 reported successes, with four inconclusive from X-ray diffraction alone; one compound that had mistakenly been included in the training data was removed; and the authors stated that "novel" had been intended to mean new to the prediction platform, not necessarily new to science.^[28]

Mission and approach

Periodic Labs describes its mission as "building AI scientists" and pairing those models with autonomous laboratories that perform real physical experiments.^[4][13] The company's central thesis, as articulated by Fedus and Cubuk in interviews and on the company website, is that further progress in scientific AI is bottlenecked not by model capacity but by data: the text available on the internet, which the company estimates at about 10 trillion tokens, has been exhausted by frontier models and does not cover proprietary experimental measurements, especially negative results.^[1][4][13] Peter Deng has recounted that Fedus told him early on that "everyone talks about doing science, but in order to do science, you actually have to do science"; Deng's own investment post renders the line as "In order to do science, you have to do real science."^[1][10] The company's launch post makes the same point: "though re-reading a textbook may give new insights, they eventually need to try their idea to see if it holds." Andreessen Horowitz's announcement put it as "You can read and re-read the textbook, but eventually you need to run the experiment."^[13][3] The company argues that physics is a verifiable domain and that "nature is the RL environment."^[13]

The Periodic Labs system is intended to close the loop between in-silico prediction and physical execution. According to the company and to reporting on its launch, large models propose candidate compounds and synthesis routes, robotic platforms in Periodic's own laboratory carry out the syntheses, characterisation instruments measure the resulting materials, and the data is fed back to update the next round of model predictions.^[1][4][10] In September 2026 the company described this as a three-phase loop: deciding which target material to make from predictions of stability and properties, predicting how to synthesize it, and determining what was actually made and whether it has the intended properties.^[16] Each experiment is described as producing gigabytes of measurement data, including failures and negative results, which the company argues are particularly valuable training signals not available in published literature.^[4][13]

The first publicly stated scientific target is the discovery of superconductors that work at higher temperatures than today's materials. Felicis has described the goals as including room-temperature superconductivity, next-generation transportation, more efficient power grids, and new supercomputers and computing architectures, and the company has said that automating materials design could also accelerate Moore's Law, space travel, and nuclear fusion.^[4][10][13] Andreessen Horowitz wrote that Periodic was building powder synthesis labs to discover new superconductors, magnets, and heat shields, and that it was already working with customers in space, defense, and semiconductors.^[3] In comments to IEEE Spectrum published in September 2026, Cubuk said the company was also investigating commercial applications such as superconducting power transmission into data center racks.^[19]

The technical bet underlying this approach is that a self-driving laboratory, operated continuously and instrumented to record every experimental input and output, can compress the time required to traverse the design space of inorganic materials. In the company's framing, prior AI-for-materials work has been able to propose candidate compounds at internet scale but has remained data-limited at the synthesis and characterisation stage, with negative results seldom published. By owning the physical experimentation pipeline, Periodic Labs argues, it can capture the high-signal data that downstream models need to make progress beyond what is available from theoretical screening alone.^{[1][4][13][25]}

Laboratories

In October 2025 TechCrunch reported that Periodic Labs had set up its lab and was working with experimental data and simulations, but that its robots were not yet running; Cubuk said they would "take a bit to train."^[1] By September 2026 Cubuk told IEEE Spectrum that the company had built robots that physically operate X-ray diffraction machines: robotic arms load samples, the measurements are sent to a language model, and the model analyzes the diffraction patterns to decide whether an experiment worked and what to try next. He said the company had been conducting 100 experiments a day, was opening a second lab, and expected to ramp up to 1,000 attempts a day.^[19] In its Neon announcement the same month, the company said its labs had moved from first physical experiments to high-throughput operation running 24 hours a day, alongside infrastructure for large-scale physics simulations, in the search for new superconductors and magnets.^[16]

Funding

Seed round, September 2025

Periodic Labs raised a $300 million seed round, announced on September 30, 2025.^[1][2][3] The round was led by Andreessen Horowitz (a16z), with Felicis writing what it and TechCrunch have described as the first check; Felicis partner Peter Deng led the firm's investment.^[1][10] Additional institutional investors in the round included DST Global, NVentures (the venture arm of Nvidia), and Accel.^[1][2][13] The Logic reported that Radical Ventures and Khosla Ventures also took part, and Radical lists Periodic Labs in its portfolio with a 2025 investment year.^[21][22] Lightspeed Venture Partners also lists Periodic Labs as a portfolio company, without stating when it invested.^[36] Wilson Sonsini Goodrich & Rosati served as legal counsel.^[2]

Named individual investors in the round included Jeff Dean, then Google's chief scientist, Jeff Bezos, Eric Schmidt, and Elad Gil.^[1][2][23] In September 2025, before the announcement, Bloomberg reported that the round was expected to value the company at $1 billion before the new money, according to SiliconANGLE, which also cited Bloomberg's sources as saying OpenAI had considered participating.^[23] Later coverage, including Forbes and Tech Funding News, put the seed-round valuation at $1.3 billion.^[15][11] Anjney Midha, who wrote Andreessen Horowitz's announcement of the investment, is listed by the firm as serving on Periodic Labs' board.^[3]

Subsequent valuation discussions

Bloomberg News reported on March 25, 2026 that Periodic Labs was in discussions with investors to raise a new round at a valuation of about $7 billion, which Contrary Research later described as roughly a fivefold step-up over the seed valuation in about six months if a round closed near those terms.^[12][11][25] Tech Funding News, citing Bloomberg, described the talks as early discussions to raise hundreds of millions of dollars, with terms that could still change.^[11]

On May 7, 2026, Forbes reported that Periodic Labs was in advanced talks to raise at a $7.5 billion valuation in a round led by AMP, an investment vehicle founded by Anjney Midha, whom Forbes described as a former Andreessen Horowitz general partner. Two of Forbes' three sources said the round would be at least $500 million, and one said it was "significantly oversubscribed" and that talks had begun for a fast-follow round at an even higher valuation. Periodic Labs did not respond to Forbes' request for comment, and AMP declined to comment.^[15] Contrary Research wrote in June 2026 that the round had not been confirmed as closed.^[25] As of October 2026, the company's news page listed no funding announcement after the seed round, and Debarghya Das's neolab list, updated on October 24, 2026, recorded no close.^[35][31] The 2026 figures are therefore reported deal-talk valuations rather than disclosed transactions.

Timeline

Team and notable hires

Public reporting and Periodic Labs' own materials have identified several senior hires beyond the two co-founders. TechCrunch reported in October 2025 that the founders had hired more than two dozen AI and scientific researchers.^[1] The New York Times reported at the launch that more than 20 researchers had left Meta, OpenAI, Google DeepMind, and other large AI projects in the preceding weeks to join the company, many of them giving up tens or hundreds of millions of dollars to do so; one of them, Rishabh Agarwal, was already a Meta employee when Mark Zuckerberg invited him to join Meta's new AI lab, and turned the offer down.^[24] SiliconANGLE listed Meta, Databricks, and Samsung among the companies the early team came from.^[23]

Alexandre Passos was described by TechCrunch as one of the creators of OpenAI's reasoning models o1 and o3.^[1]

Eric Toberer is a materials scientist whom TechCrunch described as having already made key superconductor discoveries.^[1]

Matt Horton was described by TechCrunch as a creator of two of Microsoft's generative AI tools for materials science, and he is a co-author of the Nature paper describing MatterGen.^[1][14][37] MatterGen is a diffusion-based generative model for proposing stable inorganic crystals, developed at Microsoft.^[37]

As of October 2026, the company's website lists its founding team as Abhijeet Gangan, Costa Huang, Dogus Cubuk, Dzmitry Bahdanau, Elsa Cong, Emirhan Kurtulus, Eric Toberer, Gowoon Cheon, Janosh Riebesell, Joe Checkelsky, Kate Lauterbach, Killian Sheriff, Liam Fedus, Mansheej Paul, Muratahan Aykol, Naveen Menon, Reiichiro Nakano, Rishabh Agarwal, Rohan Pandey, Sam Cross, Wei Chen, Xander Dunn, and Xiang Fu. The site says this group co-created ChatGPT, GNoME, OpenAI's Operator, "the neural attention mechanism," and MatterGen.^[4] Several of those attributions match published papers: Dzmitry Bahdanau was first author of the 2014 paper that introduced the attention-based approach to neural machine translation known as Bahdanau attention; Muratahan Aykol and Gowoon Cheon were co-authors of the GNoME paper; and Xiang Fu was a co-author of the MatterGen paper.^[30][26][37] Separately, Reiichiro Nakano was first author of OpenAI's WebGPT paper.^[34]

Because the team members are experts in different fields, they hold regular cross-disciplinary teaching sessions: TechCrunch reported a weekly graduate-level lecture given by one team member to the others, and Andreessen Horowitz described weekly sessions in which physicists teach language models to reason about quantum mechanics while machine-learning researchers learn the physics.^[1][3] The company website lists a scientific advisory board of five professors: Carolyn Bertozzi (Stanford University, chemistry), Mercouri Kanatzidis (Northwestern University, chemistry), Steve Kivelson (Stanford, physics), Zhi-Xun Shen (Stanford, physics), and Chris Wolverton (Northwestern, materials science and engineering).^[4] Contrary Research has described Shen as a superconductivity experimentalist, Kivelson as a theorist, Kanatzidis as a synthesis expert, and Bertozzi as a Nobel laureate.^[25]

As of October 27, 2026, the company's job board listed 33 openings, grouped under Science, Engineering, Business and Operations, and General Interest, all listing Menlo Park as a location; three research roles also listed Montreal, and the general application was also open to remote candidates. Listings included positions in thin films, powder processing, lab automation, materials characterization, an X-ray diffraction facility manager, midtraining, and scaling reinforcement learning.^[32]

Products and research output

Commercial work

Periodic Labs' commercial activity, as it has described it publicly, consists of engagements with industrial customers. At launch the company said it was helping a semiconductor manufacturer facing heat-dissipation problems on its chips by training custom agents that help the manufacturer's engineers and researchers make sense of experimental data.^[13] Andreessen Horowitz wrote that the company was working with customers in space, defense, and semiconductors, including training agents to automate simulations.^[3] Tech Funding News reported in March 2026 that Periodic had secured customers in the semiconductor industry and was generating revenue; the company has not disclosed revenue figures.^[11][25]

Periodic Neon

On September 15, 2026, Periodic Labs introduced Periodic Neon in three posts covering the announcement, the research, and the infrastructure behind it.^[16][17][18] Neon is a model of about one trillion parameters that the company built by midtraining and reinforcement learning on data from its own labs, starting from Moonshot AI's open-weight Kimi K2.6.^{[16][17][18][38]} Its first job is interpreting powder X-ray diffraction (XRD) patterns: working out which crystalline phases formed in a synthesis experiment and in what proportions, a task the company says can take an expert several hours per complex multiphase pattern. The company said Neon was deployed in its labs, analyzing experiments in its search for better superconductors and magnets.^[17]

Success on FrontierXRD was scored by an LLM-as-a-judge ensemble of Claude Opus 5 and GPT-5.6-Sol, calibrated against expert ratings. The company argued that XRD analysis is "falsifiable, but not easily verifiable," because pattern fitting alone cannot confirm that each identified phase is supported by the data and makes chemical sense; it therefore used this judge to supply training and evaluation signals, after checking its agreement against quality labels that experts with PhDs in materials-related fields assigned to thousands of patterns using a detailed rubric, with three experts annotating each pattern.^[17] Cost comparisons assumed standard API prices with perfect caching for external models, and a price of $2.5 per H200-hour for Neon.^[17]

The infrastructure post described a training and inference stack built on the open-source frameworks Megatron-LM, SGLang, Miles, and Ray. The company reported 4.1 times the training throughput of its Megatron baseline, decoding 2.5 times faster (from 10 to 25 tokens per second per request for a one-trillion-parameter model) after prefill-decode disaggregation and a "delta router replay" change it contributed to SGLang, full-parameter Kimi K2.6 training at 64K context on 64 H200 GPUs, and checkpoint conversion cut from 30 minutes to one minute. It also described pbox, an in-house sandbox that runs model-written scientific code in gVisor sandboxes on the spare CPUs of the GPU nodes already allocated to a reinforcement-learning job, which it said had 4.5 times lower median data-transfer latency and 3.3 times the throughput of a hosted sandbox provider it tested. The company said it shares its cluster with simulation workloads to keep utilization above 95%, and that it contributes its improvements back to Megatron-LM, SGLang, and Miles, including chunked optimizer offloading developed with NVIDIA's DevTech team.^[18]

The company said it planned to extend the approach so that its models direct scientific campaigns, develop synthesis procedures, and decide which experiments to run.^[16] All Neon results were reported by Periodic Labs itself on an internal benchmark; the posts did not announce a public release of the model weights or of the FrontierXRD evaluation set.^[16][17][18] The technology news site Menlo Times summarized the announcement the following day.^[33]

Publications and grants

As of October 2026, Periodic Labs' news page listed four posts: the September 2025 launch announcement and the three September 2026 Neon posts. None announced the discovery of a new superconductor or other new material, and the company had not announced a peer-reviewed paper describing work carried out at Periodic Labs, although members of its team continue to be cited for work performed at prior institutions, including the GNoME and MatterGen papers.^[35][26][37] The company website advertises an academic grant program to support research and lists a dedicated contact address for grant inquiries.^[4]

The company has also stated that the data it generates internally is intended to feed back into the training of its own scientific foundation models rather than into the publicly available corpora that have driven much of the past decade of progress in materials informatics. This position is consistent with the general thesis articulated by Fedus and Cubuk in public interviews, in which they distinguish between data that exists on the open internet, which is largely already used by general-purpose language models, and the experimental data that is generated only when a laboratory physically synthesises and measures a new compound.^{[1][4][7][10]} In September 2026 the company described its midtraining corpus as proprietary.^[17]

Context: the AI-for-science field

Periodic Labs belongs to a broader category of well-funded private companies founded between 2023 and 2025 that aim to apply foundation-model techniques developed for natural language and vision to problems in the natural sciences. Tech Funding News described Periodic Labs in March 2026 as part of a growing wave of companies often called "neolabs," which it characterized as startups combining AI models with physical experimentation.^[11] The term is also used more broadly: Debarghya Das's "The Ultimate Neolab List" defines a neolab as a company that employs elite researchers, spends heavily on compute (usually to train its own model), lacks significant revenue traction, and raises significant money at a high valuation. It classes Periodic Labs as "Mostly" a neolab, noting that it is "Researcher-led and pre-revenue; wet labs make it less purely software."^[31]

Within materials science specifically, Periodic Labs' immediate intellectual genealogy traces to two distinct lines of work. The first is academic and industrial high-throughput computational materials science, including density-functional theory screening efforts such as the Materials Project at Lawrence Berkeley National Laboratory, to which Google DeepMind contributed nearly 400,000 new candidate compounds derived from GNoME in 2023.^[9] The second is the emerging line of generative models for crystal structures, including Microsoft's MatterGen, which uses diffusion-based generation to propose novel inorganic compounds.^[37]

Periodic Labs' approach differs from each of these antecedents in that it explicitly couples model prediction to physical synthesis and characterisation inside a single closed loop within the company. The choice to vertically integrate the experimental stack distinguishes Periodic from companies that focus on simulation alone, and from academic groups that rely on external collaborators for synthesis.^[1][10][25]

Coverage has grouped Periodic Labs with a changing set of companies and groups whose technical scopes overlap only partially. At launch TechCrunch named the startup Tetsuwan Scientific, the nonprofit FutureHouse, and the University of Toronto's Acceleration Consortium as others working on AI scientists.^[14] Contrary Research named Lila Sciences, Radical AI, Google DeepMind, and Microsoft as competitors, and IEEE Spectrum paired Periodic with Quantum Formatics, another startup using AI and robots to search for superconductors.^[25][19]

In December 2025, Periodic Labs was one of 24 organizations that signed memorandums of understanding with the U.S. Department of Energy to collaborate on the Genesis Mission, the federal effort to use AI to accelerate scientific discovery.^[20]

Investors' rationale

Andreessen Horowitz, the lead investor, published a memo on its website at the time of the announcement arguing that "the bottleneck has been the iteration speed of human-led experimentation," and that Periodic removes that constraint. It framed the industries Periodic could affect, including advanced manufacturing, materials science, semiconductors, energy, and aerospace, as representing roughly $15 trillion of global GDP.^[3] Midha wrote that he had come to a similar view while working as a visiting scientist in Stanford's Department of Applied Physics, where he found frontier models to be poor at analyzing condensed-matter experiment data.^[3] Felicis, which has stated that it wrote the first check, titled its investment post "Building a New Way to Do Science," describing a loop between models that form hypotheses and automated labs that execute syntheses and feed the results back.^[10] The company's launch post lists NVentures and Accel among its backers.^[13]

Reception

The launch of Periodic Labs received broad coverage in technology and business press, including TechCrunch, The New York Times, Bloomberg, and SiliconANGLE.^{[1][14][23][24]} Coverage frequently noted that the founding team combined a senior figure from each of OpenAI and Google DeepMind and that the seed round was unusually large.^[1][14][23] Periodic Labs appeared on the Forbes AI 50 "Brink" list in 2026.^[15]

Contrary Research's June 2026 report on the company argued that a step-up to a $7 billion valuation would rest on early commercial traction rather than a demonstrated scientific breakthrough, and compared that private valuation with the much lower public market capitalizations of Recursion Pharmaceuticals and Schrodinger. It identified as the main risks that the lab loop might not produce commercially meaningful discoveries fast enough to justify its costs, given that high-temperature superconductivity has resisted researchers for decades; that progress on superconductors might not generalize to customers' other materials problems; and that well-funded competitors are bidding for the same scarce researchers.^[25]

Periodic Neon's headline results are company-reported comparisons on an evaluation the company built from its own lab data and scored with its own LLM-judge ensemble, and the company noted that its held-out generalization test was easier than FrontierXRD.^[17]

Continue reading

• OpenAIOpenAI is an American artificial intelligence research and deployment organization.Mentioned 18 times above
• Google DeepMindGoogle DeepMind is an artificial intelligence research and model-development organization at Google.Mentioned 10 times above
• Andreessen HorowitzAndreessen Horowitz, commonly styled a16z, is an American venture capital firm founded in 2009 by Marc Andreessen and Ben Horowitz. It invests from seed through growth stage in technology companies.Mentioned 9 times above
• GGNoMEGNoME (Graph Networks for Materials Exploration) is a deep-learning system from Google DeepMind that predicts the thermodynamic stability of inorganic crystals and uses those predictions to search for new…Mentioned 6 times above
• MatterGenMatterGen is a generative artificial intelligence model from Microsoft Research for designing novel inorganic crystalline materials.Mentioned 6 times above
• NNeolabA neolab, also written neo-lab or NeoLab, is an informal term for an AI research company organized around a technical research program.Mentioned 5 times above

More from See also

• List of neolabs
• William "Liam" Fedus
• Lila Sciences
• FutureHouse
• CuspAI
• Genesis Mission
• Kimi K2.6
• Harmonic (AI)
• Math Inc and Gauss
• Jua
• WindBorne Systems
• Project Prometheus (Bezos)
• AI for Science
• AlphaFold
• Switch Transformer

Topics:AI CompaniesAI for Science

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28Julie Bort, "Top OpenAI, Google Brain researchers set off a $300M VC frenzy for their startup Periodic Labs," TechCrunch, October 20, 2025. techcrunch.com/...-for-their-startup-periodic-labs Accessed 2026-10-27.
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7"Wilson Sonsini Advises Periodic Labs on $300 Million Seed Round," Wilson Sonsini Goodrich & Rosati, October 1, 2025 (archived copy; the original page is no longer available). web.archive.org/...on-dollar300-million-seed-round Accessed 2026-10-27.
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Anjney Midha, "Investing in Periodic Labs," Andreessen Horowitz, September 30, 2025. a16z.com/...investing-in-periodic-labs Accessed 2026-10-27.
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Periodic Labs, official company website. periodic.com Accessed 2026-10-27.
5. ^"Periodic Labs," CB Insights company profile. cbinsights.com/...periodic-labs Accessed 2026-10-27.
6. ^1 ^2 ^3 ^4 ^5 ^6Chris McKay, "OpenAI's Post-Training Head Liam Fedus Departs to Launch AI-Powered Materials Science Startup," Maginative, March 17, 2025. maginative.com/...owered-materials-science-startup Accessed 2026-10-27.
7. ^1 ^2"Liam Fedus on AI for Scientific Discovery," StartupHub.ai, April 4, 2026. startuphub.ai/...us-on-ai-for-scientific-discovery Accessed 2026-10-27.
8. ^Amil Merchant and Ekin Dogus Cubuk, "Millions of new materials discovered with deep learning," Google DeepMind blog, November 29, 2023, summarising Merchant et al., "Scaling deep learning for materials discovery," Nature (2023). deepmind.google/...s-discovered-with-deep-learning Accessed 2026-10-27.
9. ^1 ^2 ^3 ^4"Google DeepMind Adds Nearly 400,000 New Compounds to Berkeley Lab's Materials Project," Lawrence Berkeley National Laboratory News Center, November 29, 2023 (archived copy; the original page is no longer available). web.archive.org/...new-compounds-materials-project Accessed 2026-10-27.
10. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Peter Deng, "Building a New Way to Do Science: Our Investment in Periodic Labs," Felicis, September 30, 2025. felicis.com/...periodic-labs-investment Accessed 2026-10-27.
11. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8"Former OpenAI and DeepMind researchers seek $7B valuation to build 'AI scientists'," Tech Funding News, March 26, 2026. techfundingnews.com/...or-ai-startup-periodic-labs Accessed 2026-10-27.
12. ^1 ^2 ^3"AI Science Startup Periodic Labs Is in Deal Talks at About $7 Billion Valuation," Bloomberg, March 25, 2026. bloomberg.com/...alks-at-about-7-billion-valuation Accessed 2026-05-19.
13. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Periodic Labs, "Introducing Periodic Labs," September 30, 2025. periodic.com/...introducing-periodic-labs Accessed 2026-10-27.
14. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Julie Bort, "Former OpenAI and DeepMind researchers raise whopping $300M seed to automate science," TechCrunch, September 30, 2025. techcrunch.com/...ng-300m-seed-to-automate-science Accessed 2026-10-27.
15. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Iain Martin and Anna Tong, "Former OpenAI Researcher To Raise $500 Million For AI Science Startup," Forbes, May 7, 2026, republished by Forbes Africa on May 25, 2026. forbesafrica.com/...million-for-ai-science-startup (original: forbes.com/...e-500-million-for-ai-science-startup) Accessed 2026-10-27.
16. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Periodic Labs, "Building Labs that Learn," September 15, 2026. periodic.com/...building-labs-that-learn Accessed 2026-10-27.
17. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18Periodic Labs, "Nature Is Our Learning Environment," September 15, 2026. periodic.com/...nature-is-our-learning-environment Accessed 2026-10-27.
18. ^1 ^2 ^3 ^4 ^5 ^6 ^7Periodic Labs, "AI Infrastructure at Periodic," September 15, 2026. periodic.com/...ai-infrastructure-at-periodic Accessed 2026-10-27.
19. ^1 ^2 ^3 ^4 ^5Alan Patterson, "Betting on AI and Robots to Automate Superconductor Discovery," IEEE Spectrum, September 2, 2026. spectrum.ieee.org/...re-superconductor-ai-research Accessed 2026-10-27.
20. ^1 ^2U.S. Department of Energy, "Energy Department Announces Collaboration Agreements with 24 Organizations to Advance the Genesis Mission," December 18, 2025. energy.gov/...nts-24-organizations-advance-genesis Accessed 2026-10-27.
21. ^1 ^2 ^3"Radical Ventures backs Periodic Labs in US$300M round to build AI scientists," The Logic, October 1, 2025. thelogic.co/...us300m-round-to-build-ai-scientists Accessed 2026-10-27.
22. ^1 ^2 ^3Radical Ventures, "Periodic Labs," portfolio page. radical.vc/...periodic-labs Accessed 2026-10-27.
23. ^1 ^2 ^3 ^4 ^5 ^6Maria Deutscher, "Periodic Labs raises $300M to accelerate scientific research with AI," SiliconANGLE, October 1, 2025. siliconangle.com/...elerate-scientific-research-ai Accessed 2026-10-27.
24. ^1 ^2"Top A.I. Researchers Leave OpenAI, Google and Meta for New Start-Up," The New York Times, September 30, 2025, as excerpted by OODA Loop. oodaloop.com/...i-google-and-meta-for-new-start-up (original: nytimes.com/...ai-meta-google-openai-periodic) Accessed 2026-10-27.
25. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Contrary Research, "Periodic Labs: Business Breakdown & Founding Story," June 2026. research.contrary.com/...periodic-labs Accessed 2026-10-27.
26. ^1 ^2 ^3Amil Merchant, Simon Batzner, Samuel S. Schoenholz, Muratahan Aykol, Gowoon Cheon, and Ekin Dogus Cubuk, "Scaling deep learning for materials discovery," Nature 624, 80-85, published online November 29, 2023. nature.com/...s41586-023-06735-9 Accessed 2026-10-27.
27. ^Nathan J. Szymanski et al. (including Ekin Dogus Cubuk), "An autonomous laboratory for the accelerated synthesis of inorganic materials," Nature 624, 86-91, published online November 29, 2023. nature.com/...s41586-023-06734-w Accessed 2026-10-27.
28. ^Nathan J. Szymanski et al., "Author Correction: An autonomous laboratory for the accelerated synthesis of inorganic materials," Nature 650, E1, published online January 19, 2026. nature.com/...s41586-025-09992-y Accessed 2026-10-27.
29. ^William Fedus, Barret Zoph, and Noam Shazeer, "Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity," arXiv:2101.03961, January 11, 2021. arxiv.org/...2101.03961 Accessed 2026-10-27.
30. ^Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio, "Neural Machine Translation by Jointly Learning to Align and Translate," arXiv:1409.0473, September 1, 2014. arxiv.org/...1409.0473 Accessed 2026-10-27.
31. ^1 ^2 ^3Debarghya Das, "The Ultimate Neolab List," updated October 24, 2026. debarghyadas.com/neolabs Accessed 2026-10-27.
32. ^1 ^2Periodic Labs, job board on Ashby. jobs.ashbyhq.com/periodic-labs Accessed 2026-10-27.
33. ^Karan Bhatia, "Periodic Labs Introduces Periodic Neon," Menlo Times, September 16, 2026. menlotimes.com/...ic-labs-introduces-periodic-neon Accessed 2026-10-27.
34. ^Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, et al., "WebGPT: Browser-assisted question-answering with human feedback," arXiv:2112.09332, 2021. arxiv.org/...2112.09332 Accessed 2026-10-27.
35. ^1 ^2 ^3 ^4Periodic Labs, "News." periodic.com/news Accessed 2026-10-27.
36. ^Lightspeed Venture Partners, "Periodic Labs," portfolio page. lsvp.com/...periodic-labs Accessed 2026-10-27.
37. ^1 ^2 ^3 ^4 ^5Claudio Zeni et al. (including Matthew Horton and Xiang Fu), "A generative model for inorganic materials design," Nature 639, 624-632, published online January 16, 2025. nature.com/...s41586-025-08628-5 Accessed 2026-10-27.
38. ^Moonshot AI, "Kimi-K2.6," model card on Hugging Face. huggingface.co/...Kimi-K2.6 Accessed 2026-10-27.

Show all 38 references