COMPANY PROFILE / MATERIALS INFORMATICS
Periodic Labs
A research and deployment company founded in 2025 that says "nature is our learning environment" and trains AI on its own lab data
Periodic Labs is a privately held U.S. company that says it is building AI models and automated laboratories (autonomous labs) together to speed up discovery in the physical sciences. It was announced on 30 September 2025, and in September 2026 it presented "Periodic Neon", a one-trillion-parameter AI further trained on its own lab data. But the results published so far centre on how well an AI reads X-ray diffraction data, and official material makes no claim that new materials such as superconductors have been found. This profile uses only the company's own official announcements, and keeps the company's claims apart from what could not be confirmed.
- Periodic Labs in 30 seconds, and where it stands in materials informatics
- The logo and how the name is written
- Where the company sits in the materials informatics field
- Scale
- Resources committed to materials informatics
- What it has done so far
- What comes next
- What the company is aiming for in materials informatics
- Map of materials informatics partnerships
- Technologies, products and services
- The risks this company carries
- Glossary, references and claim-to-source audit
1. Periodic Labs in 30 seconds, and where it stands in materials informatics
California, U.S.The headquarters address is not given in official material
Foundation models & compute (secondary)Our own classification Our calculation
(1 trillion parameters)Announced 15 September 2026. Weights appear not to be public
Materials informatics (MI) is a way of searching for new materials by gathering data, using AI to predict and design candidates, and testing them by experiment. Periodic Labs aims to run that whole flow by owning labs and training the AI itself on those labs' data Sourced. The company's thinking is that text on the internet is being used up, and that experiments on nature produce new, high-quality data. But this is a description of company strategy. What has actually been published is, so far, limited to how well an AI reads X-ray diffraction results measured in its labs. This profile reads the conditions of that evaluation carefully and keeps the company's large goals (such as high-temperature superconductors) apart from its track record.
2. The logo and how the name is written
A note on names: the name is Periodic Labs. The logo on the website is a grid of short strokes beside the lower-case words "periodic labs". The domain is periodic.com, and the older periodiclabs.ai redirects to it (as of the checking date). The origin of the name, such as any link to the periodic table, is not stated in official material.
Official site: https://periodic.com/3. Where the company sits in the materials informatics field
First, the figure sorts MI companies into four types and shows where Periodic Labs sits.
Next, the three-phase loop for materials discovery that Periodic Labs describes, and where its published result (Periodic Neon) falls. "X-ray diffraction (XRD)" is a measurement that shines X-rays on a powder to tell which crystals are present and in what amounts.
4. Scale
The company is privately held, so there are no statutory financial filings. Only what can be confirmed is shown Sourced.
| Item | Detail | Source and type |
|---|---|---|
| Announced | "Introducing Periodic Labs" was published on 30 September 2025 | Company announcement. The date of incorporation is not given |
| Location | A high-throughput lab in Menlo Park (U.S.), and infrastructure for large-scale physical simulation, as described | Announcement of 15 September 2026. The headquarters address is not given |
| Employees | Not disclosed. The website lists 23 names as the founding team | The list is on the website's team section and is not the total headcount |
| Revenue | Not disclosed | Not stated in any official material. This article does not estimate it |
| Funding raised | The amount is not disclosed. The website names backers (chapter 5) | A search of U.S. SEC Form D filings found none by the company itself (our check) |
| Listing | Privately held | No share information on the website |
The funding amount is not written in any official company material. Press reports mention figures, but this article does not use press-only numbers. A search of SEC filings also turned up a filing by an investment fund whose name contains the company's name; that is a vehicle on the investor side and does not show the company's own fundraising Not yet confirmed.
5. Resources committed to materials informatics
Periodic Labs specialises in AI for the physical sciences, so company-wide resources are almost entirely MI resources. The four questions are people, money, outside investments, acquisitions and joint ventures, and ties to universities and public bodies.
| Item | What could be confirmed | Status |
|---|---|---|
| People | The website lists 23 founding-team members. The company says the team helped create ChatGPT, DeepMind's GNoME, OpenAI's Operator and MatterGen (the company's self-description). The total headcount is not disclosed | Sourced (the background is the company's claim) |
| Money (raised) | The company lists a16z, Felicis, DST, NVentures (NVIDIA's venture arm) and Accel as backers, plus individuals (Jeff Bezos, Elad Gil, Eric Schmidt, Jeff Dean). The amount is not disclosed | Names sourced; amount could not be confirmed |
| Money (computing) | The company says the final training run peaked at 1,300 NVIDIA H200 GPUs and that cluster utilisation exceeds 95%. The cost of buying or renting GPUs is not disclosed | Sourced (figures are the company's) |
| Outside investments, acquisitions, joint ventures, divestments | None stated in official material | Could not be confirmed |
| Universities, research institutes and public funding | Five professors serve as scientific advisors (chemistry and physics at Stanford University; chemistry and materials science at Northwestern University). The company says it is launching a grant programme for academic research. No government funding or joint work with national laboratories is stated | Sourced (public funding could not be confirmed) |
Status of the main models and facilities
The table lists name, type, specification, form of release, stage and date, as far as official material allows. Numbers measured under different conditions are not ranked against each other.
| Name | Type | Main specifications (including company claims) | Release | Stage and date |
|---|---|---|---|---|
| Periodic Neon | Large language model specialised for XRD analysis (further-trained) | Starts from an open-weight model with 1 trillion parameters (Kimi K2.6), with "midtraining" and reinforcement learning on material that includes lab data. The company claims a 55.3% success rate on its hardest internal evaluation (FrontierXRD), up from 2.7% | Weights and code: could not be confirmed as released (no match in a Hugging Face search) | Deployed in its own labs. 15 September 2026 |
| Periodic harness | Framework that gives the model tools (databases, analysis software) | Equipped with lab context, in-house crystal-structure and simulation databases, and XRD analysis software. The company claims a success rate 3.8x that of a standard Claude Code-based setup | Not public | In use. September 2026 |
| Training and inference infrastructure | Infrastructure for large-scale training, inference and sandboxing | Modified open-source tools (Megatron, SGLang, Miles, Ray). The company claims 4.1x the training throughput of a baseline and 2.5x faster inference, and built its own sandbox, "pbox" | The company says some improvements were contributed back to open source | In use. September 2026 |
| High-throughput labs | Automated synthesis and measurement labs | Described as running 24/7. Types and numbers of instruments and daily throughput are not disclosed | Operated in house | In operation (as described in September 2026) |
In powder X-ray diffraction, each crystal that is present gives a "fingerprint" of peaks, but when several crystals are mixed the peaks overlap and are hard to tell apart. The company says the samples in its evaluation are hard ones, with five phases on average among those that have an accepted solution. The right answer also depends on context such as starting materials, temperature, atmosphere and handling history (an oxide makes sense for a sample heated in air, but a hydroxide would be odd unless it could have met moisture) Sourced. In other words, even when the fitting software gives a good-looking number, a person has had to judge whether the answer makes chemical sense. The company wants an AI that can make that judgement so it can keep up with the stream of measurements from its labs.
6. What it has done so far
| When | What happened | Meaning |
|---|---|---|
| 30 September 2025 | Published "Introducing Periodic Labs", announcing it would build "AI scientists" and autonomous labs | Goals include higher-temperature superconductors. As an industry example, it says it works on a semiconductor maker's heat-dissipation problem (the customer is not named) |
| 30 September 2025 to September 2026 | Built high-throughput labs in Menlo Park and infrastructure for large-scale physical simulation | No dates or scale given. The company says it moved "from its first physical experiments to scaled-up labs that run 24/7" |
| 15 September 2026 | Announced Periodic Neon (one trillion parameters, deployed in the labs) | It handles XRD analysis |
| 15 September 2026 | Published the research post "Nature Is Our Learning Environment" | Explains the evaluation method and results. The evaluation data are internal and not published |
| 15 September 2026 | Published the infrastructure post "AI Infrastructure at Periodic" | Explains GPU counts, speed-ups and contributions to open source |
7. What comes next
New magnetsCompany goals. Nothing achieved yet
None of the following is stated in official material Not yet confirmed.
Whether and when a superconductor or magnet was found; the lab's instrument count, throughput and investment; release of Neon's weights; release of the evaluation data (FrontierXRD and others); the funding amount; customer contracts and revenue; the size and timing of the academic grants. The company writes that its next AI will plan experiments, develop synthesis procedures and choose which experiments to run, but that is a goal.
8. What the company is aiming for in materials informatics
This chapter includes things the company has not said directly. They are inferences from primary sources, set out with the evidence.
| What can be read from it | Primary evidence | Type |
|---|---|---|
| The core product is likely to be science-focused AI trained on lab data rather than the materials themselves | The company writes that lab experiments provide the data to train increasingly capable scientific AI, which in turn guides better experiments. What has been published is AI performance only, with no discovered materials | Inference |
| Near-term revenue may come from industry deployments such as AI agents that read a client's experimental data | The launch announcement says it is training custom agents for a semiconductor maker's engineers and researchers to make sense of their experimental data | Inference |
| New material discoveries are probably not yet at a stage it can announce | The September 2026 posts say only that it is searching for superconductors and magnets, with no candidate materials or measurements | Inference |
| It treats AI infrastructure (large-scale training and inference) as a pillar alongside the labs | Training on 1,300 GPUs, a dedicated infrastructure post and contributions back to open source | Sourced |
| It builds on external open-weight models and appears not to pre-train its own | Neon starts from the publicly released Kimi K2.6, and the company writes that Kimi K3 could be a stronger starting point. No training from scratch is mentioned | Inference |
The reading that "new discoveries are not yet ready to announce" rests on which results the company chose to publish. The company's central goal is a high-temperature superconductor, but what the three September 2026 posts describe in detail is the success rate of an AI that reads X-ray diffraction (55.3%) and the speed of its training infrastructure Sourced. There is nothing showing that a new substance was made (composition, measured values or a paper). From the order in which it presents lab results, first as the performance of an analysis AI, this article reads the company as not yet at the stage of publishing the labs' own output Not yet confirmed.
9. Map of materials informatics partnerships
10. Technologies, products and services
Prices and contract terms for anything offered to outside customers are not disclosed Not yet confirmed. What has been published is the AI used in its own labs (Periodic Neon) and the infrastructure that trains and runs it. For background, see our explainers on large language models and text mining for materials, AI agents for materials research, self-driving laboratories, and AI analysis of spectra and images.
How Periodic Neon was built
How to read the evaluation results
| Item | What the company writes | Type |
|---|---|---|
| Name and content of the evaluation | FrontierXRD: 134 complex patterns from the company's labs that take experts hours to resolve. Among those with an accepted solution, the fits contain five phases on average | Company's claim. Internal and not published |
| Result | Neon succeeds 55.3% of the time, against 2.7% for the starting model, Kimi K2.6, about a 20-fold gain. The company claims it beats other high-end models at lower cost | Company's claim. The comparison models are those the company names; this article has not verified their performance |
| Scoring | Success is judged by an ensemble of two judge AIs (Opus 5 and GPT-5.6-Sol, in the company's wording), calibrated against labels from PhD-level experts, three per pattern. Experts agreed with each other 77.2% of the time, the judge ensemble with experts 74.6%, and with the expert consensus 84% | Company's claim |
| Cost comparison | For external models, standard API prices assuming "perfect caching". For Neon, an assumed US$2.5 per H200-hour | Estimate based on assumptions (company) |
| Generalisation check | On 198 measurements from chemical systems excluded from training, the company says Neon outperformed other models. It describes this as easier than FrontierXRD | Company's claim |
| Effect of tooling | With the same model (Claude Opus 5), the Periodic harness scores 3.8x higher than Claude Code with open databases and analysis software | Company's claim |
This is a company announcement based on an internal evaluation set, scoring by judge AIs, and comparison models the company chose. No evaluation data that outsiders can reproduce, and no peer-reviewed paper, could be found. The company says the judge AI's agreement with experts (74.6%) is close to experts' agreement with each other (77.2%), but errors in scoring still affect the evaluation values. The names, performance and costs of other companies' models are as the company wrote them and were not checked here. This article reads the result as a claim that, on its own problems, a specialised AI trained on lab data scored better than general-purpose AIs.
Materials engineers can take three points from this announcement Sourced.
(1) Failures and ambiguous results become training material. The company says the "negative results seldom published" that labs generate are valuable data. A company with many instruments producing data is in a position to learn from them.
(2) When the right answer cannot be checked cheaply, build a judge. In XRD analysis a good-looking fit is not enough; whether it makes chemical sense has to be judged. The company turned expert judgement into a scoring rubric for a judge AI and checks its reliability by agreement with experts. That is a useful design idea for anyone who wants to automate a similar analysis.
(3) Check separately whether it works on unseen systems. Evaluating on chemical systems held out from training is a way to avoid the familiar problem of materials AI that is strong only on systems it has seen. However, that evaluation uses internal data, and the company itself calls it easier.
11. The risks this company carries
| Risk | Detail | Type |
|---|---|---|
| Results cannot be verified outside | The evaluation data and model weights are not public and there is no peer-reviewed paper. Performance rests on the company's announcement alone | Sourced |
| No materials discovery shown | It aims at superconductors and magnets, but no result about a new substance appears in official material. The gap between goal and track record is large | Sourced |
| Lab scale unknown | Instrument counts, throughput, investment and funding are all undisclosed, leaving few ways to check progress from outside | Not yet confirmed |
| Dependence on an external base model | Neon starts from a model released by another party. The terms of use (licence) of that model were not checked in this article and are not described in official material | Inference |
| Fast-moving comparison targets | The AI models the company compared against change quickly through updates and renaming, so any advantage holds for a point in time only | Inference |
| Gap in compute scale | The company itself writes that raising training compute to the level of today's frontier models would unlock more capability, which would take further funding | Inference |
| Caution about grand language | Phrases such as "AI scientist" and "nature is our learning environment" express the company's philosophy, not a record of results. This article does not use them to rank anyone | Sourced |
12. Glossary
- Materials informatics (MI)
- Searching for new materials efficiently with data and AI, in the flow of data, then prediction and design, then testing by experiment.
- Autonomous lab
- A lab in which AI chooses the next experiment, robots and automated tools make and measure the material, and the results go back to the AI.
- X-ray diffraction (XRD)
- A measurement that shines X-rays on a sample to tell which crystals are present and in what proportions. In mixtures the peaks overlap and analysis is hard.
- Midtraining
- A stage after basic training in which a large amount of data from a specific field is fed in to instil knowledge, before specialised further training.
- Reinforcement learning (RL)
- A way of learning in which an AI tries things and improves from scores (rewards) for the results.
- Open-weight model
- A model whose weights (the result of training) are public so that users can run it themselves. Terms of use are set individually.
- Harness
- The surrounding setup that gives an AI model tools (databases, analysis software) and working procedures.
- H200
- A model of NVIDIA GPU for AI computing. For its cost estimate the company assumes US$2.5 per GPU-hour.
- Judge AI (LLM-judge)
- A method in which one AI scores another AI's output, with reliability checked by agreement with expert ratings.
- Superconductor
- A material whose electrical resistance drops to zero below a certain temperature. For practical use, ones that work at higher temperatures are wanted.
13. References
- Periodic Labs Official website (home page): the company's description, the lists of founding team and backers, scientific advisors and the academic grant programme. Checked 22 October 2026. https://periodic.com/
- Periodic Labs "Introducing Periodic Labs" (30 September 2025): the launch announcement, with goals, approach and an example of industry deployment. https://periodic.com/news/introducing-periodic-labs
- Periodic Labs "Building Labs that Learn" (15 September 2026): the Periodic Neon announcement and the three-phase loop for materials discovery. https://periodic.com/news/building-labs-that-learn
- Periodic Labs "Nature Is Our Learning Environment" (15 September 2026): how Periodic Neon was trained and how it scored on X-ray diffraction analysis. https://periodic.com/news/nature-is-our-learning-environment
- Periodic Labs "AI Infrastructure at Periodic" (15 September 2026): the training, inference and sandbox infrastructure, and the number of GPUs used. https://periodic.com/news/ai-infrastructure-at-periodic
- GitHub Pull request #24851 in sgl-project/sglang (one of the changes the company says it contributed upstream; we confirmed that the pull request exists and its title, but not the author's link to the company). https://github.com/sgl-project/sglang/pull/24851
- GitHub Pull request #1371 in radixark/miles (same; a change that parallelises Megatron checkpoint conversion on Ray). https://github.com/radixark/miles/pull/1371
14. Claim-to-source audit
| Claim in the text | Type | Source |
|---|---|---|
| The name Periodic Labs, the announcement of 30 September 2025, and that the company is privately held | Sourced | Reference 1, 2 https://periodic.com/ https://periodic.com/news/introducing-periodic-labs |
| The company says its labs in Menlo Park run 24/7 | Sourced | Reference 3 https://periodic.com/news/building-labs-that-learn |
| The 23-person founding team and five scientific advisors are listed on the website | Sourced | Reference 1 https://periodic.com/ |
| Backers (a16z, Felicis, DST, NVentures, Accel, individual investors) are named on the website | Sourced | Reference 1, 2 https://periodic.com/ https://periodic.com/news/introducing-periodic-labs |
| The founding team's background (ChatGPT, GNoME, Operator, MatterGen) | Sourced | Reference 2 The company's self-description (a claim) https://periodic.com/news/introducing-periodic-labs |
| The goals (high-temperature superconductors and so on) and the semiconductor-maker example | Sourced | Reference 2 Company goals https://periodic.com/news/introducing-periodic-labs |
| The announcement of Periodic Neon (one trillion parameters, deployed in the labs) | Sourced | Reference 3 https://periodic.com/news/building-labs-that-learn |
| FrontierXRD: 134 items, 55.3% success (2.7% at the start), five phases on average | Sourced | Reference 4 Company's claim; internal evaluation, not published https://periodic.com/news/nature-is-our-learning-environment |
| Judge-AI agreement (experts with each other 77.2%, judge with experts 74.6%, with consensus 84%) | Sourced | Reference 4 Company's claim https://periodic.com/news/nature-is-our-learning-environment |
| The Periodic harness scores 3.8x higher; the midtraining data double every month | Sourced | Reference 4 Company's claim https://periodic.com/news/nature-is-our-learning-environment |
| The generalisation evaluation of 198 items; cost assumed at US$2.5 per H200-hour | Sourced | Reference 4 Company's statement https://periodic.com/news/nature-is-our-learning-environment |
| A peak of 1,300 H200 GPUs in the final run; cluster utilisation above 95% | Sourced | Reference 5, 4 https://periodic.com/news/ai-infrastructure-at-periodic https://periodic.com/news/nature-is-our-learning-environment |
| Infrastructure figures such as 4.1x training and 2.5x inference, and the self-built pbox | Sourced | Reference 5 Company's claim https://periodic.com/news/ai-infrastructure-at-periodic |
| Contributions back to open source (SGLang, Miles, Megatron-LM) and cooperation with NVIDIA DevTech | Sourced | Reference 5 https://periodic.com/news/ai-infrastructure-at-periodic |
| Pull request #24851 in SGLang and #1371 in Miles exist | Sourced | Reference 6, 7 Only existence and title were checked; the author's link to the company was not https://github.com/sgl-project/sglang/pull/24851 https://github.com/radixark/miles/pull/1371 |
| The idea of discovery as a three-phase loop | Sourced | Reference 3 https://periodic.com/news/building-labs-that-learn |
| The plan to extend the AI to planning experiments, synthesis steps and choosing experiments | Not yet confirmed | Reference 3 Company plan https://periodic.com/news/building-labs-that-learn |
| Funding amount, revenue, headcount and lab scale | Not yet confirmed | Reference 1 Not stated in official material. This article does not estimate them https://periodic.com/ |
| Whether a superconductor or magnet was found; release of Neon's weights and the evaluation data | Not yet confirmed | Reference 3, 4 Not stated. This article does not assert that none exists https://periodic.com/news/building-labs-that-learn https://periodic.com/news/nature-is-our-learning-environment |
| A SEC Form D search found no filing by the company itself | Not yet confirmed | Our check (EDGAR full-text search) in this article. A filing by an investment fund is a different matter |
| Main type autonomous experiments & robotics, secondary type foundation models & compute | Our calculation | Our own classification, not the company's self-description (this article) |
| The three-phase loop with Neon's position, and the illustration of how Neon was built | Our calculation | Our own framing (illustrations in this article) |
| The reading that science-focused AI will be the core product | Inference | Our interpretation of the lab-data-to-AI approach and of the published results being AI performance only (this article) |
| The reading that AI agents reading clients' experimental data may be near-term revenue | Inference | Our interpretation of the semiconductor-maker example in the launch announcement (this article) |
| The reading that new discoveries are not yet ready to announce | Inference | Our interpretation of which results were published (this article) |
| The reading that it does not appear to pre-train its own model | Inference | Our interpretation of Neon's starting point being an external model (this article) |
| The risks of dependence on a base model's licence, shifting comparison targets and the compute-scale gap | Inference | Our interpretation of what official material says and omits (this article) |
Last updated: 22 October 2026 / Troy Technical
The content of this article is based on Periodic Labs' website and official announcements and on public GitHub pages. No market-research estimates or press-only figures are used. The performance figures (55.3%, 3.8x, 4.1x and so on) are the company's own claims and have not been confirmed by independent verification. Passages marked "Inference" are this article's reading of primary sources and are not statements by the company.