A Conversation with Kaiwuji's 陆子恒: Inventing New Materials with AI
Summary
Kaiwuji has raised hundreds of millions of yuan in angel-plus financing, with investors betting not on current revenue but on the possibility that, once model capabilities cross an inflection point, AI will discover new materials capable of moving an entire industry materially forward. The company was registered last September and has a team of about 10; revenue and profit remain “zero.” The investment is a bet on original IP discovery and subsequent commercialization, not on expanding a mature business.
The essence of materials discovery is “placing atoms on a blank sheet of paper,” but an investable result must clear 3 gates at once: it must be synthesizable, possess the target properties, and be capable of mass production and commercialization. A lithium-germanium-phosphorus-sulfur material discovered in 2011 illustrates the leverage of a single compound: it lets lithium ions move through a solid at nearly the speed they move through liquid, helping create the solid-state battery category; it still sold for roughly RMB700K–800K per ton last year.
Kaiwuji does not plan to sell models or tools first; it wants to take a material from original IP through kilogram-scale production, a customer’s production line, and eventually mass production. 陆子恒’s rationale is blunt: “If you have a shovel that can dig up gold, you certainly wouldn’t let someone else use it first.” The team needs one successful end-to-end pipeline before platformization and asset-light licensing can become credible; some employees may even have to “go become factory managers.”
What truly triggered the startup was MatterSim’s zero-shot prediction of properties such as phonons and specific heat without training on any specific material or property, outperforming nearly all models specifically trained on phonons at the time. The result brought the team’s GPT-3.5-inspired scaling insight into materials: use scalable models and data to achieve generalization instead of building a separate model for every domain. After a quick test “over a meal,” skeptics realized the model had genuinely crossed the threshold.
AI’s highest-value role across the materials chain is finding the original compound that could upend an industry from an enormous chemical space, not replacing every experiment and process step. In practice, generative models, known-material databases, and predictive models screen candidates in parallel; veteran materials scientists then review them one by one, make gram-scale samples, put them into devices, advance promising candidates to kilogram scale, and send them to customers for production-line trials. Each screening run can take anywhere from tens of seconds to tens of minutes: “I only need one that can make money.”
The clearest near-term technical milestone is to try, over the next 1–2 years, to solve material free energy well enough to determine whether any real or hypothetical material can be synthesized thermodynamically. 陆子恒 believes success could replace a substantial share of experiments, with solid-state experiments largely replaceable. He is explicit that the timeline for finding a material capable of changing the commercial ecosystem is “not very certain”; even if foundation models stopped improving today, Kaiwuji might still succeed—stronger models would simply raise the odds of “winning the lottery.”
The main uses of the financing are training compute and AI talent, not early-stage materials experiments. Compute costs tens of millions of yuan a year, while the lab costs several million yuan; before kilogram scale, AI and people account for most of the spending. Reaching production capacity could require more than RMB100M. When 陆子恒 sent out 3 offers this year, “his hands were shaking,” but costs should fall rapidly if the model workload eventually shifts toward inference.
The long-term moat is more likely to come from coupling model intuition, materials intuition, and commercial judgment in the same room than from exclusive control of a foundation model. Kaiwuji wants people who are “extremely deep in one domain while having an exceptionally broad view,” and ranks taste, initiative, and the ability to choose hard problems above memorized knowledge. 陆子恒 even expects coding and model design to lose value over time, while “tinkering” and expanding one’s surface area for luck remain important skills.
Deep dive
1. The hundreds of millions in financing are buying a materials-discovery inflection point, not current financial performance
Kaiwuji was registered last September, has a team of about 10, and has both revenue and profit at “zero.” It operates in AI for Materials: scaling AI models while also handling materials discovery, experimental validation, production scale-up, and commercialization.
陆子恒’s summary of the investment thesis is deliberately restrained: model capabilities have improved sharply over the past few years, and the next few could produce materials that “change human civilization.” The more realistic payoff is discovering a material capable of moving an industry or industrial sector materially forward.
He accepts that the capital is also betting on the team, but says the financing brings pressure: “We have to work extremely hard to make this happen; otherwise, we’re being irresponsible to the industry.” The reason is the exceptionally high expectations surrounding AI for Materials.
2. A new material is not a formula name but a constrained game of placing atoms
陆子恒 explains materials from first principles: “You place atoms on a blank sheet of paper, and once you’re done, you have a material.” The structure first cannot be too high in energy and must be genuinely synthesizable; it then needs the desired physical and chemical properties; finally, it must be manufacturable and enter a commercial system.
Single-crystal silicon is the most extreme historical example: one material supported the computer and AI industries, with Silicon Valley itself named after silicon. Its example shows that the value of original materials IP does not depend on whether the compound sounds complicated when described.
A more recent example is the lithium-germanium-phosphorus-sulfur material discovered in 2011. Pressed into a pellet, its powder becomes as hard as a tabletop, yet lithium ions pass through it at nearly the speed they move through water. In 陆子恒’s view, “this one material brought the entire solid-state battery industry into being.”
The industry built around the material is not yet fully commercialized, but it sold for roughly RMB700K–800K per ton last year. The scientists involved were at Tokyo Institute of Technology and initially received funding from Toyota; at least the research itself was free. 陆子恒 is unsure how the individuals might share in the value created, which is precisely the commercial question Kaiwuji cares about.
3. Materials giants excel at scale-up, but breakthrough foundational IP still depends on a few people
陆子恒 divides materials R&D into 2 stages: discovering the original compound or foundational IP, then scaling the process and selling it. Commercial companies such as Dow and BASF are extremely strong in the second stage, but few organizations focus on finding entirely new foundational IP; discovery still happens mainly in universities and research institutes.
The talent distribution is even more uneven where AI could help most: breakthrough IP is often concentrated in a very small number of scientists. 陆子恒 estimates that Goodenough “should have” produced 4 or 5 distinct materials used as core lithium-battery materials. His self-deprecating formulation is that he is not that smart, but he wants AI to be that smart—to cultivate a group of people who are “good enough.”
He also cites Dacron, a polyester fiber that created enormous value for DuPont. Today’s materials giants were essentially single-product companies on day 1, starting with products such as polyimide and Nylon 66 and expanding into multi-category platforms only after raising capital.
4. The pipeline splits between commercial value and proof of capability
Kaiwuji divides its targets into 2 groups: high-value materials with clear technical bottlenecks and a path to making money; and materials that may not be profitable in the short term but could change civilization. In 陆子恒’s words, the choice is between “I want to get rich” and “I want to demonstrate our capabilities and prove to the world that we can do this.”
Energy is an obvious area because AI itself is ultimately constrained by energy. Superconductors and first-wall materials for nuclear fusion are ideal targets that could combine commercial value with something more important: representing the frontier of human materials capability.
Other opportunities sit in industrial materials that sound far less glamorous. Spherical high-pressure tanks for storing methane and natural gas on ocean-going vessels require specialized adhesives that can operate under extremely low temperatures and high pressure; the materials carry high value and face chokepoint problems. He stressed that this illustrates the selection logic, not a disclosure of a specific pipeline.
5. Kaiwuji wants to replicate Flagship, but first it will dig for gold with its own shovel
The company analogy 陆子恒 uses is Flagship Pioneering: incubating drug pipelines from lead compounds, with most projects failing but successful ones becoming independent companies comparable to Moderna rather than merely another drug. Kaiwuji wants to incubate several standalone, world-class companies in materials.
Koji asked whether the ultimate deliverable would be a patent, an R&D pipeline, or a material. 陆子恒’s “honest answer” is “we’ll see as we go,” but the current preference is clear: first take at least one material all the way from start to finish rather than immediately shipping a general-purpose tool.
His commercial judgment is direct: “If you have a shovel that can dig up gold, you certainly wouldn’t let someone else use it first. I want to dig first.” The industry also needs an end-to-end example proving that AI materials can move beyond attractive scientific metrics and enter a real commercial chain. The team may therefore sell materials itself in the future, with “some people having to go become factory managers.”
6. The moment MatterSim predicted phonons turned scaling from analogy into conviction
Before starting the company, 陆子恒 worked on AI for Science at Microsoft Research and helped build the materials program at the Zhongguancun Academy. For the past 2–3 years, most of the team’s time went into model development, while it insisted on making 2 or 3 materials for real. 陆子恒 does not fully trust virtual screening: “I have to see and hold the thing before I believe it.” The combination of model capability and physical validation triggered the startup.
MatterSim initially aimed to be a general-purpose model: given an atomic structure, it would infer basic physical and chemical properties across material categories in a zero-shot setting. A condensed-matter physicist whom 陆子恒 deeply respects was skeptical: “Forget predicting everything. First predict the simplest physical property.” That property was phonon; at the macroscopic level, it corresponds to specific heat.
Around 2023–2024, the team ran an informal test and found that the model, despite not being trained on that property or on any particular class of material, outperformed nearly every model at the time that had been trained specifically for phonons. “We tested it over a meal, and afterward everyone was in a different state of mind.” It was the key evidence that the model had crossed the generalization threshold.
Koji asked whether the 2025 Nature paper was also about scaling laws. 陆子恒 said that may be a misreporting: the MatterGen paper did not directly discuss scaling laws. The real shift was from an early VAE architecture that was not scalable enough to diffusion, inspired by DALL-E 2 and DALL-E 3; he believes diffusion can scale and absorb large datasets, producing a major improvement in generation quality.
7. A materials pipeline moves through model queues, expert judgment, and real production-line feedback
陆子恒 uses a hypothetical lithium-battery anode requirement to break down the process. Suppose large-scale energy storage needs a new material with higher capacity, higher power density, and lower voltage. The team first works backward from the industry context to define specific requirements, then asks models to generate or screen original compounds that meet them.
Generative models, known-material databases, and predictive models run in parallel and continuously build a candidate queue. The team is not fixated on searching the entire space: “I’m not trying to search 10,000 materials. I only need one that can make money.” The ideal outcome is a single-shot hit; in reality, the team may need to rapidly eliminate hundreds of candidates.
Once the models produce their outputs, the process does not continue with endless AI optimization. Veteran materials scientists meet and inspect the candidates one by one, identify which can enter the lab the next day, make gram-scale samples, and put them into batteries. If the feedback is positive, the team advances to kilogram scale and hands the material to customers who actually use it for production-line trials, then decides whether to scale volume itself or through partners.
Current generative models are primarily graph-based diffusion models, not text-style sequence Transformers. Training data includes materials humans have synthesized and structures the team believes could plausibly be synthesized. The model learns the manifold of synthesizable materials in a high-dimensional space; screening takes anywhere from tens of seconds to tens of minutes and can be done on a single GPU.
8. The moat is not who can call a foundation model, but who knows where its limits are
Koji’s challenge was straightforward: if teams A, B, and C can all use foundation models, where does differentiation come from? 陆子恒 first rejects the idea that the industry is already in a competitive phase. The industry urgently needs any company using any model to produce and commercialize a breakthrough material; “creating the pie together” matters more than fighting over share.
The real cognitive differences have 2 layers: knowing what model capabilities can find expensive, transformative materials; and knowing how much commercial value the same model can actually deliver in different domains. Misjudging the model’s boundaries can send a team’s resources into the wrong category or the wrong stage of R&D.
He insists that the veterans of modeling, statistical-physics simulation, and experiments must work together every day. If materials and AI teams communicate only through project-based coordination, they will develop separate cultures. The ability to understand the boundaries on both sides may directly determine whether a team finds breakthrough IP.
The industry is still far from converging. Commercial value may ultimately concentrate in original compounds or in downstream process optimization, and the technical path is unsettled. Periodic Labs may rely more on the inference capabilities of language models and mid-training to develop reasoning-oriented models. 陆子恒 does not claim to know Project Prometheus’s direction precisely and is not even sure about reports that it plans to build rockets. CuspAI and Orbital Materials are also still exploring their approaches.
9. Free energy is a 1–2 year tool milestone; the timeline for breakthrough materials remains unknown
Kaiwuji wants to “break through material free energy” over the next 1–2 years: accurately infer, across a broad materials space, whether real or hypothetical structures can be synthesized thermodynamically. 陆子恒 believes that achieving this could replace a substantial share of experiments, with solid-state experiments largely replaceable.
The second milestone is producing a material that genuinely changes the world or the commercial ecosystem, but his answer contains no packaging: “To be honest, we’re not very certain ourselves.” The team can confirm that its broad direction and methodology are gaining more evidence, but it cannot set a reliable date for pipeline success.
Koji then posed a counterfactual in which foundation models stopped improving from this point onward. 陆子恒 believes Kaiwuji could still succeed; he simply does not know when it would hit. If foundation models continue to improve, “the odds of winning the lottery go up.” This is a probability-improvement argument, not a promise of certain delivery.
10. Early on, training and AI talent cost the most; later, production capacity does
陆子恒 breaks down the financing needs in concrete terms: compute currently costs tens of millions of yuan a year, and AI talent costs “a lot more” than other roles. When he sent out 3 offers this year, “his hands were shaking,” because model capabilities had not yet reached the expected inflection point and the team had to keep training and iterating rather than simply buying inference from external models.
The materials experiments themselves are not especially expensive early on. A lab costs several million yuan, and advancing to kilogram scale is not yet large-scale manufacturing; before kilogram scale, people and AI account for most of the cost. If the feedback is positive and the company or a partner begins scaling volume, downstream production investment could exceed RMB100M.
Training is expensive while inference is relatively cheap, creating a staged shift in the business model. If the team eventually focuses mainly on inference, costs will fall immediately; until then, when asked how much more it needs to burn, 陆子恒 says, “I have a rough number, but it’s still quite expensive.”
He wants commercial feedback to feed back into fundamental research. Examples such as DeepSeek and Gemini show him that concentrated capital and strong commercial incentives can accelerate technology; nuclear fusion, superconductivity, and quantum computing could be pushed by the same mechanism. The benefit is faster technological progress, while the cost is that research organizations face more direct pressure to deliver.
11. The AI era still rewards hard-core depth, taste, and relentless tinkering
Kaiwuji is not looking for generalists who know a little about every field. It wants people who are “extremely deep in one domain, highly specialized, and broad in perspective.” Modeling, simulation, and experiments can be different areas of expertise, but each person needs an anchor point and must solve the hardest problem that cannot be decomposed further.
陆子恒 observes that PhD students at the Zhongguancun Academy enter projects on day 1 and develop around a goal like startup employees. When he first worked on generative models, he “hadn’t read a single paper,” relying mainly on persistent questioning of GPT and DeepSeek to reconstruct the mathematical framework step by step, from cross-entropy loss to Bayesian inference. AI has changed the path to learning, but it does not automatically provide initiative.
Knowledge questions in interviews are becoming harder to use as a measure of candidates because GPT may know more than the interviewer. He still asks hard technical questions to verify a candidate’s past learning ability, then tests vision and taste with a prompt: “You have RMB1B, a 100-person team, and no restrictions or oversight. What do you want to do?” The choice of problem is itself a capability.
陆子恒 admits this may be “a reckless thing to say,” but believes coding and model design could become less valuable and eventually be phased out. Models such as Claude and Opus, along with agentic systems such as Open Cloud, are already changing how companies produce; Kaiwuji equips every employee with Claude and currently imposes no token limit. He still advises young people to pursue a PhD: “Someone gives you money, you don’t have to pay it back, and you get to explore whatever you find interesting.”
His own ambitions moved from wanting, as an undergraduate, “a RMB3K monthly salary, a wife, and a child” to becoming a professor, commercializing research, entering AI research, and founding a startup. Around 2018–2020, he researched batteries and worked on some commercialization at the Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences; in 2021, he moved to Cambridge despite the pandemic, then turned down a slate of faculty offers to join Microsoft Research. Each turn expanded his so-called “surface area for luck”: “As long as you keep tinkering, there’s still hope.”
Asked about himself 10 years from now, he only wants to know: “Is this fun? Do I still have the energy to keep tinkering?” 陆子恒 attributes part of his willingness to keep moving to the strong sense of security provided by his parents. Both were university professors, and although he spent years moving between Hong Kong, Shenzhen, the UK, and the US, he rarely felt homesick because they remained a powerful source of support.