AI could help invent new materials. First, firms must figure out how to make them. Startups like Periodic Labs, Meissner and FL01 combine AI models for physics and chemistry with increasingly automated labs to come up with new substances Periodic Labs CEO Liam Fedus at the All In conference in Montreal in September. The startup is developing AI and automated labs to create new materials and manufacturing processes. Photo: Mélanie Olmstead, All In/Handout Oct 7, 2026 | 5:45 AM ET Listen Now MONTREAL - A new batch of tech firms is applying AI to the physical world, seeking to develop materials that perform better, cost less and fill critical industrial needs, as well as the manufacturing processes to practically produce them. Startups are training AI models on scientific literature and the results of their own experiments to simulate chemical and physical properties. Some firms have built so-called self-driving labs, which can autonomously make and test new formulations. Their early target markets include tech-heavy sectors like data centres, quantum computing and nuclear fusion. Talking Points * Startups are developing AI models and autonomous labs to help come up with and synthesize new materials for applications like data-centre compute and cooling, quantum computing and nuclear fusion * Firms must ensure that the substances their systems suggest can actually be produced in the real world, either by relying on common ingredients and processes or by re-engineering manufacturing Periodic Labs is building "a system that can do the scientific method and has a deeper understanding of different areas of materials engineering," CEO Liam Fedus told The Logic, resulting in a technology that "we think will be very foundational." Launched in September 2025, the San Francisco-based firm aims eventually to generate completely new materials. Still, it can take years to engineer such substances so they can be manufactured and used at scale, Fedus said on the sidelines of the All In conference in Montreal last month. Periodic is also working with clients that want to use its AI tools to improve their products and processes, or conduct joint research and development projects with the startup. The firm is starting out with materials for magnets, semiconductors and superconductors, all widely used in high-tech industries. Periodic has an office in Montreal that leads the development of its machine learning systems. The firm's Canadian outpost has drawn founding researchers from Google DeepMind, Meta and ServiceNow who have previously made significant AI breakthroughs. Fedus, who was raised in Maine, got his doctorate at Université de Montréal under Yoshua Bengio and Hugo Larochelle, the AI pioneer's successor as scientific director at the Mila AI institute. The large language models powering ChatGPT and other generative AI tools were initially trained on the internet's worth of text data. Academic literature alone isn't enough to create similar foundation systems for chemistry and physics, according to Fedus - authors tend to obfuscate key details, leave out failed experiments and reframe accidental discoveries as brilliant predictions. "There's this fictionalized view of the scientific method in papers," he said. So Periodic is using its autonomous labs to fill in the data gaps for its AI model. Together, they create a feedback loop in which digital simulation and analysis bookend real-world experimentation. Periodic charges clients to use its software. It could also split revenues from the intellectual property created during joint R&D. In October 2025, the firm raised a US$300-million seed round from backers including Toronto-based Radical Ventures and prominent Silicon Valley investors Accel, Andreessen Horowitz and Khosla Ventures. Toronto-based FL01 is also betting on its experimental hardware and the data it generates. The startup's technology develops and tests coolants. The current industrial process for evaluating thermal fluids is manual and time-consuming, according to CEO Mohammad Zargartalebi. FL01 has built "a robotic scientist who makes the formulations itself [and] analyzes them in a matter of minutes," he said. Zargartalebi began developing the firm's sensors during a fellowship at the University of Toronto, after being approached by Castrol's parent company, the oil and gas multinational BP. He co-founded FL01 in November 2025. Working with partners like BP, the startup has identified about 30 potential formulations of thermal fluids that can be used in the cooling equipment for data centres. Three passed all the tests for stability, corrosion and other key properties, and are now ready to commercialize. FL01 has now added an AI model that identifies coolant candidates and can direct the self-driving lab to test them. Building the testing hardware first meant the firm had the right data to train its system, said Zargartalebi. "The properties of coolants are not easily predicted." Zargartalebi claimed that gives his four-person firm an advantage over potential competitors like Periodic and Lila Sciences, which are much bigger and better-funded but have a wider focus. FL01 recently graduated from Toronto's Creative Destruction Lab (CDL) accelerator, but has not raised venture capital. The firm sells testing services to clients, and could one day license or manufacture the coolants that it invents. Meissner, which graduated from CDL at the same time as FL01, is similarly applying AI to create new superconductors for quantum computing, nuclear fusion and other disruptive applications. The Toronto-based startup is trying to develop devices that can move electricity without heat or loss, but don't need to be chilled as much as current products. Meissner's area of focus is an advantage, according to CEO Olivia Leng. "If you're building a generalized materials discovery platform, it can't necessarily be applied - effectively, at least - to finding new superconductors," she said. "They're such an exotic, unconventional material class to work with." Meissner is focusing on developing superconductors for which the raw ingredients are easy to procure and that can be made using existing processes. Last month, the startup announced a $3.6-million pre-seed round from BDC Capital and several well-known deep-tech executives. Firms applying AI to materials science must screen for practicality; otherwise, their models might come up with substances that are perfect in theory but impossible to manufacture in the real world, or require rare elements. Periodic generally tries to work with clients' existing manufacturing infrastructure. "People need to deliver on their products and goods," Fedus said. "They don't have time to re-engineer the floor plan." Still, it's trying to ensure promising new materials do get made by applying its AI and labs to solve the physical and engineering challenges that might otherwise prevent the production of tons of the stuff. Fedus previously worked at OpenAI, where he helped develop ChatGPT. These days, he's trying to connect the digital world to physical engineering. "You're not going to think your way to a new high-temp superconductor," he said. To make material improvements, Fedus added, firms like Periodic must recognize that "the sciences are this experimental, iterative field."

Periodic ships the flywheel this blog predicted in July: A trillion-parameter model trained on its own lab's x-ray data. AI Science Infrastructure North America Periodic Labs announced Periodic Neon today: a trillion-parameter model post-trained on data from the company's own high-throughput physical labs in Menlo Park, deployed to interpret X-ray diffraction (XRD) results from real experiments. This blog profiled Periodic in detail in July, when it was essentially the entire disclosed materials-science slice of frontier labs' science-vertical spending, and closed with a prediction: "every frontier lab owns wet-lab capacity by end of 2027... the proprietary-experiment flywheel is too obviously the moat." Today's release is that flywheel's first public turn - a company using its own physical experiments to train a model, then using that model to decide what experiment to run next. Who's behind it, and what they're both already good at. Periodic was founded by Liam Fedus, former head of post-training at OpenAI and a co-creator of ChatGPT, and Ekin Dogus Cubuk, formerly DeepMind's chemistry and physics research lead and a co-author of GNoME - the 2023 Nature paper that used graph neural networks to computationally predict 2.2 million new candidate crystal structures. The pairing isn't incidental: Cubuk already built the digital half of this exact idea at DeepMind (mining computation for new materials candidates), and Periodic's pitch is the physical half - real labs generating real experimental data that a digital model can't get any other way, closing a loop GNoME's own approach couldn't close on its own. The company has raised roughly $300 million from Andreessen Horowitz, Felicis, DST Global, NVIDIA's venture arm NVentures, and Accel, with reporting from earlier this year pointing to a further $500 million round in progress at a $7.5 billion valuation. The Pareto-dominance claim, checked against the chart itself. Periodic's own framing is specific: its infrastructure lets it "train specialized models that Pareto-dominate frontier models, including GPT-6 Astra and Claude Fable 5.1, on our X-ray diffraction evaluations." That's a strong claim worth checking against the published chart rather than taking on faith. Reading the actual points: Periodic Neon's best setting (xhigh) lands around 54% success at roughly $4.30 per problem. GPT-6 Astra's best setting (max) reaches about 53% - essentially matching Neon on success rate, but at roughly $8.40, almost double the cost. Every other GPT-6 Astra and Claude Fable 5.1 point on the chart sits below and to the right of Neon's curve - lower success for higher cost, at every comparable setting. That is, genuinely, a Pareto dominance: no rival point beats Neon on both axes at once. One anomaly worth flagging on its own terms: Claude Opus 5's curve actually drops from 35% success at its "xhigh" setting to 28% at "max" - paying for the highest reasoning effort makes Opus 5 do worse on this specific task, a real and specific result in Periodic's own chart, not an artifact of this post's reading of it. The base model isn't the one you'd assume, and that's the actual story. Here's the detail worth slowing down for. Periodic's infrastructure post states plainly that Neon comes from midtraining and reinforcement learning on top of Kimi K2.6 - Moonshot AI's trillion-parameter open-weight model, released in April 2026. Moonshot's other, larger model, Kimi K3, a 2.8-trillion-parameter model this blog covered at its own launch in July, is not the base - it appears in Periodic's benchmark chart only as a single reference point, tested without the effort-level curve every other model on the chart gets, sitting at the most expensive point on the whole chart (about $8.50) with a success rate around 40%. Read plainly: Moonshot's own newer, larger, more expensive open model, run directly on this task, costs more and performs worse than Periodic's specialized fine-tune of Moonshot's smaller, older, cheaper one. That's a cleaner demonstration of "narrow specialization beats raw scale" than most vendor charts manage to produce, and it lands three days after this blog covered the same underlying argument from TypeSafe's Jev launch - a different company, a different domain, the same structural bet: a small model tuned tightly on a narrow, well-defined task beats a bigger general one running the task cold. The second chart makes the size of that gap concrete. Before any of this post-training, Kimi K2.6's own baseline on Periodic's XRD task sits at roughly 3% success. After RL scaling on Periodic's own lab data, the tuned model climbs past Kimi K3's raw 18% baseline early in the training-compute curve and continues up to roughly 57% at the highest training compute and highest inference-time reasoning effort shown - an improvement of nearly 19x over the untrained starting point. Inference-time compute matters almost as much as training compute here: at every point on the curve, the "high inference compute" setting beats "low" by a wide, fairly consistent margin, confirming that how hard the model is allowed to think at answer-time is doing real, separate work from how much it learned during training. The infrastructure claims are more transparent than most, with one exception. The companion infrastructure post is unusually specific for a vendor blog: 4.1x training throughput against a stated Megatron baseline (with the baseline's own tuning documented in a footnote), 2.5x inference speed against SGLang, and specific pull requests actually upstreamed to the open-source projects it builds on - SGLang PR #24851 for delta router replay, Miles PR #1371 for fast weight resharding. Those are checkable claims in the literal sense: anyone can go look at the linked PRs and see what shipped. The sandboxing comparison keeps that same rigor in its measurement - 4 runs, error bars showing min and max, a specific 100-sandbox test - but drops the naming convention everywhere else in the post: the rival is only ever "a state-of-the-art sandbox provider," never named. Every other comparison in the post names its baseline (Megatron, SGLang, Kimi K2.6, Kimi K3, GPT-6 Astra, Claude Fable 5.1); the one comparison that doesn't is also the one being compared against a commercial product Periodic presumably still has some relationship with or interest in not naming directly. NVIDIA is investor, vendor, and engineering partner in the same post. Worth naming plainly, the way this blog has for every other lab-and-hardware relationship this month: NVIDIA's venture arm, NVentures, is a named investor in Periodic. The entire compute stack described - 1,300 H200 GPUs at peak, the memory and throughput optimizations - runs on NVIDIA hardware. And the infrastructure post credits "the NVIDIA DevTech team" directly for co-developing chunked optimizer offloading, now merged into Megatron-LM's own development branch. None of that makes the technical results false - NVIDIA's own engineers contributing real, upstreamed code is a concrete, checkable claim, not a vague partnership mention - but a funder, hardware supplier, and named engineering collaborator all being the same company, inside one infrastructure announcement, is exactly the kind of overlap worth stating outright rather than leaving implicit. Where this actually lands. Set against the prediction this blog made about Periodic in July - that the proprietary-experiment flywheel would become the default moat for every lab serious about AI-for-science - today's release is closer to confirmation than surprise. What's more interesting than the confirmation itself is the specific shape it took: not a company claiming its trillion-parameter model is smarter in general, but one showing, with a base model it names plainly and a rival model from the same open-weight family it's willing to let underperform in its own chart, that a narrow specialist trained on data nobody else has access to beats general frontier models at the one thing that specialist actually needs to do. That's a smaller, more specific claim than "we built a better AI." It's also the one this week has offered the most direct evidence for.