Periodic Labs
Last reviewed
May 19, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 ยท 2,757 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 19, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 ยท 2,757 words
Add missing citations, update stale details, or suggest a clearer explanation.
Periodic Labs is an American artificial intelligence research startup that builds AI systems 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, one of the largest first institutional financings ever recorded for a private technology company.[^1][^2] 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 was a senior contributor to the GNoME materials-discovery project.[^1][^2][^3]
The company is one of the most prominent entrants in a wave of 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 in semiconductor thermal management.[^1][^2][^4]
Periodic Labs was founded in 2025 in San Francisco, California.[^4][^5] 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 an internal note that Fedus also shared publicly, he described the move as a difficult decision and said he intended to work closely with OpenAI as a partner; OpenAI confirmed at the time that it planned to invest in and partner with the new company, framing AI for science as a strategic priority.[^6]
The company was incorporated and represented during its seed round by the law firm Wilson Sonsini Goodrich & Rosati.[^2] Periodic Labs is registered at a Market Street address in San Francisco and has been described in industry databases as headquartered in the San Francisco Bay Area.[^4][^5]
Liam Fedus (also published as William Fedus) is the chief executive of Periodic Labs.[^1][^6] Before founding the company, he served as vice president of research in charge of post-training at OpenAI, where he oversaw fine-tuning, reinforcement learning from human feedback, and other techniques used to refine models such as those that power ChatGPT.[^6][^7] 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.[^1][^3] During his earlier time at Google Brain he co-authored the Switch Transformer paper, an influential demonstration of how mixture-of-experts routing could scale transformer models to a trillion parameters at modest compute cost.[^1][^7] 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' co-founder and chief scientist for materials and physics.[^1][^3] Before joining the company, Cubuk was a staff research scientist at Google, where he led the materials and chemistry research efforts that became part of Google DeepMind after Google's 2023 reorganisation of its AI research groups.[^3][^8] He earned his PhD in applied physics at Harvard University and held a postdoctoral fellowship at Stanford University, where he worked on machine learning for disordered solids.[^3] Cubuk was a senior author of the 2023 Google DeepMind paper that introduced GNoME (Graph Networks for Materials Exploration), a graph neural network system used to predict the stability of inorganic crystals; the GNoME project reported 2.2 million predicted novel crystalline materials, of which approximately 380,000 were classified as most likely to be stable, and partnered with external laboratories that synthesised 736 of the candidates.[^8][^9] The GNoME work was widely regarded as a milestone for machine-learning-accelerated materials science and is often compared, in scope if not in modality, to AlphaFold's impact on biology.[^8][^9]
Periodic Labs describes its mission as "building AI scientists" and pairing those models with autonomous laboratories that perform real physical experiments.[^4][^10] 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 textual and simulated corpora available on the internet do not adequately cover proprietary experimental measurements, especially negative results.[^1][^4][^7] Fedus has summarised the position succinctly: "everyone talks about doing science, but in order to do science, you actually have to do science."[^1] In a separate quote attributed to the company's launch materials, he said: "You can read and re-read the textbook, but eventually you need to run the experiment."[^10]
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] Each experiment is described as producing gigabytes of structured measurement data, including failures and negative results, which the company argues are particularly valuable training signals not available in published literature.[^4][^11]
The first publicly stated scientific target is the discovery of new high-temperature superconductors, with the longer-term aspiration of identifying materials that can superconduct at or near room temperature. The company has framed such a discovery as enabling more efficient power grids, lossless long-distance transmission, more powerful magnets, and new computing architectures.[^1][^4][^10] Periodic Labs has also publicly indicated work in adjacent problem domains, including heat shielding, thermal management for semiconductor packaging, and materials for space and defense applications.[^1][^4][^10]
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 most published experimental literature reporting only successful outcomes. By owning the physical experimentation pipeline, Periodic Labs argues, it can capture the high-resolution, high-signal data that downstream models need to make progress beyond what is available from theoretical screening alone.[^1][^4][^10][^11]
The company's internal organisation has been publicly described as comprising two divisions, informally referred to as "Bits" and "Atoms," reflecting the separation between machine-learning research and physical laboratory operations.[^4]
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 Ventures writing what the firm has publicly described as the first check; Felicis partner Peter Deng led the firm's investment.[^1][^3] Additional institutional investors in the round included DST Global, NVentures (the venture arm of Nvidia), Accel, and General Catalyst.[^1][^2][^11] Wilson Sonsini Goodrich & Rosati served as legal counsel.[^2]
Named individual investors in the round included Jeff Dean (chief scientist of Google), Jeff Bezos, Eric Schmidt, and Elad Gil.[^1][^2][^3] Reporting at the time of the announcement noted that the round was unusually large by historical seed-round standards and that the company became a unicorn at the seed stage; subsequent reporting placed the post-money valuation at approximately $1.3 billion.[^1][^11][^12] Andreessen Horowitz partner Anjney Midha joined Periodic Labs' board as part of the round.[^3]
Bloomberg News reported on March 25, 2026 that Periodic Labs was in discussions with investors to raise a new round at a valuation of approximately $7 billion, roughly five times the seed-round valuation reached six months earlier.[^12] Subsequent reporting in spring 2026 indicated that the discussions involved a target raise of several hundred million dollars and characterised the round as still in negotiation rather than closed.[^11][^12] As of the date of this article, the company has not made a public announcement confirming closure of any post-seed round, and the figures should be treated as reported deal-talk valuations rather than disclosed transactions.[^11][^12]
Public reporting and Periodic Labs' own materials have identified several senior hires beyond the two co-founders.
Alexandre Passos is a member of the founding team and was previously a researcher at OpenAI, where he is credited as one of the creators of OpenAI's reasoning models o1 and o3.[^1] He had earlier worked on natural-language and reinforcement-learning systems at Google before joining OpenAI.[^1]
Eric Toberer joined Periodic Labs from the Colorado School of Mines, where he was a materials-science professor with a research focus on thermoelectric and superconducting materials.[^1] Reporting at the time of the company's launch credited Toberer with prior superconductor-related discoveries that informed Periodic Labs' choice of an initial scientific target.[^1]
Matt Horton joined Periodic Labs from Microsoft, where he worked on the company's generative AI tools for materials science, including contributions to the MatterGen project.[^1][^10] MatterGen is a diffusion-model-based generative system for proposing novel inorganic crystals, developed within Microsoft's AI for Science group.[^1]
In addition to these named hires, the company has publicly stated that its team includes physicists, chemists, simulation researchers, and machine-learning engineers, and that it holds regular cross-disciplinary technical seminars in which team members deliver graduate-level lectures to one another on their areas of expertise.[^3][^10] The company website describes a scientific advisory board comprising professors from Stanford University and Northwestern University in the fields of chemistry, physics, and materials science, although the website does not publicly name them in a single canonical roster.[^4]
The company posts open job listings, organised under its "Bits" and "Atoms" divisions, on a hosted hiring platform.[^4]
As of mid-2026, Periodic Labs has not released a public product in the conventional sense of a downloadable model, dataset, or hosted API. The company's commercial activity, as it has described it publicly, consists of bespoke engagements with industrial customers, including semiconductor manufacturers seeking to improve heat dissipation in chip packaging, and partners in the space and defense sectors.[^1][^4][^11] One press report has described the company as already generating revenue from such engagements at the time of its seed-round announcement.[^11]
Periodic Labs has not, as of the date of this article, published a flagship peer-reviewed scientific paper attributable to work carried out at the company itself, although several members of its team have continued to be cited in the literature for work performed at prior institutions, including the GNoME paper from Google DeepMind.[^8][^9] The company has launched an academic grants programme, described on its website, 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 structured, multimodal experimental data that is generated only when a laboratory physically synthesises and measures a new compound.[^1][^4][^7][^10]
Periodic Labs occupies a position within 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. The category is sometimes referred to in venture-capital writing as "neolabs," reflecting its emphasis on coupling computational models with physical experimentation in company-operated laboratories.[^11]
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.[^1]
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][^11]
Other companies pursuing related autonomous-laboratory and AI-for-science theses, frequently mentioned alongside Periodic Labs in industry coverage, include private firms in the broader category of automated chemistry and biology platforms. Periodic Labs has not, to the author's knowledge, publicly disclosed direct competitive comparisons with named peers, and the venture press routinely groups it with a fluid set of companies whose technical scopes overlap only partially.[^11]
Andreessen Horowitz, the lead investor, published a public memo on its website at the time of the announcement framing the investment as an effort to remove what it called the "iteration speed of human-led experimentation" as the bottleneck in scientific progress.[^3] Felicis Ventures, which has stated that it wrote the first check, framed its investment as a bet on a "new way to do science" combining computational models with physical laboratory automation.[^10] Nvidia's NVentures and Accel publicly confirmed their participation through the company's launch materials.[^1][^2]
Press coverage of the round emphasised that the seed-stage cheque size, the unicorn-status valuation, and the participation of unusually senior individual investors reflected investor appetite for AI-for-science theses with credentialled founding teams more than they reflected disclosed traction at the time of the round.[^1][^11]
The pattern of senior individual investors is also notable in the context of the broader AI ecosystem: Jeff Dean remained chief scientist at Google at the time of the investment, and Periodic Labs' technical lineage on the materials-science side is rooted in his organisation. Eric Schmidt, the former chief executive of Google, has separately funded a wide array of philanthropic and venture initiatives in AI for science; his participation in the Periodic Labs round is consistent with that pattern. Jeff Bezos has invested in several other AI infrastructure and applications companies in the same period. The combined presence of these investors, along with Felicis, Andreessen Horowitz, DST Global, Accel, NVentures, and General Catalyst, has been described in coverage of the company as one of the most credentialled investor groups assembled for a seed-stage technology investment.[^1][^2][^3][^11]
The launch of Periodic Labs received broad coverage in technology and business press, including TechCrunch, Bloomberg, and trade outlets covering the AI and venture-capital industries.[^1][^11][^12] Coverage frequently noted three points: that the founding team combined a senior figure from each of OpenAI and Google DeepMind; that the seed round was unusually large; and that the company's stated objective of room-temperature superconductivity is a long-standing open problem in condensed-matter physics whose resolution would have considerable practical impact.[^1][^11][^12]
The size of the seed financing also drew analytical commentary about valuation discipline in the AI-for-science segment, particularly given that the company had not, at the time of the round, publicly demonstrated a flagship scientific result of its own.[^11]