Alibaba's DAMO Academy built an AI agent that screened 2.4 million crystal structures in 28 GPU hours and walked away with four real, lab-confirmed superconductors. That speed is the actual story.
On July 3, 2026, Alibaba's DAMO Academy unveiled Elements Claw, an AI agent built with Renmin University of China and the University of Chinese Academy of Sciences to hunt for new superconducting materials. According to a report from the South China Morning Post, the system screened 2.4 million candidate crystal structures in just 28 GPU hours and narrowed them down to roughly 68,000 promising candidates. Researchers then synthesized and experimentally confirmed four materials nobody had identified before: Hf21Re25, Zr4VRe7, HfZrRe4 and Zr3ScRe8, with critical temperatures reaching up to 6.5 Kelvin.
That's not a simulation claiming a discovery. That's atoms in a lab, cooled down, behaving the way the model said they would.
Elements Claw runs on a 1-billion-parameter foundation model trained on 125 million molecular and crystal structures, according to Pandaily's reporting on the project. It doesn't just rank candidates, it reads scientific literature, reasons about whether a proposed structure can actually be synthesized, and plans the discovery workflow itself. That's the part worth sitting with: the agent isn't a faster search engine bolted onto old chemistry software. It's doing the judgment calls a senior materials scientist would normally make about which of 68,000 options are worth a chemist's time.
Elements Claw isn't the first AI system to turn theoretical screening into physical results, and it won't be the last. Google DeepMind's GNoME project, published in late 2023, predicted 2.2 million potentially stable inorganic crystal structures, with about 380,000 judged stable enough to be worth pursuing. The number that actually matters came later: independent labs have since synthesized and experimentally validated more than 736 of those predicted materials, confirming GNoME wasn't just generating plausible-looking chemistry on paper.
Apple Hands China's AI Brain to Alibaba to Get Apple Intelligence Approved China has approved Apple Intelligence, but the AI running it isn't Apple's own. Alibaba's Qwen will power Siri and writing tools for Chinese users, with Baidu handling search, after nearly two years of stalled talks that included a rejected DeepSeek partnership. Alibaba and Baidu shares jumped on the Hong Kong exchange the same day the approval... - apple intelligence china approval requirements - alibaba qwen replacing apple ai model
Berkeley Lab pushed that further with its A-Lab, an autonomous robotic system that ran for 17 straight days testing GNoME's predictions without a human in the loop. It successfully synthesized 36 of 57 targeted compounds, a roughly 63% hit rate that would be remarkable even for a team of PhD chemists working by hand for months.
Microsoft took a different approach with MatterGen, a diffusion model that skips screening altogether and generates a candidate material directly from the properties you want, the same way an image generator turns a text prompt into a picture. Microsoft Research says that when asked to design a material with a bulk modulus of 200 gigapascals, MatterGen predicted a compound called TaCr2O6, which researchers then synthesized. Microsoft open-sourced the model's code, which means any lab with the compute can now run the same kind of generative search Alibaba and DeepMind built proprietary systems for.
None of this is abstract for the battery and chip industries watching it. Researchers at Politecnico di Torino recently mined the Energy-GNoME database, built on DeepMind's original crystal predictions, using machine-learning force fields called MACE to flag a shortlist of candidate cathode materials for sodium, potassium, magnesium and calcium-ion batteries, chemistries that don't depend on lithium at all. Lithium supply is a real geopolitical chokepoint for EV makers, and a credible lithium-free cathode candidate is worth more to a battery company than almost anything else AI could hand them this year.
Here's the thing that should unsettle anyone still thinking of AI as a chatbot category: superconductors, battery cathodes and semiconductor materials are the physical inputs the entire next hardware cycle runs on. A chip, a battery pack or a quantum computer is only as good as the material inside it, and until recently finding that material meant years of trial-and-error synthesis in a lab. Elements Claw collapsed a search problem that used to take a research team years into 28 GPU hours. That's not an incremental tool upgrade. It changes who can afford to go looking.
The bottleneck has also shifted. DeepMind's GNoME can propose hundreds of thousands of candidates, but only a few hundred have actually been made and tested so far, which tells you the limiting factor isn't ideas anymore, it's lab capacity. Whoever builds the fastest synthesis and testing pipeline, not the biggest model, is going to capture the value here first. Berkeley's A-Lab is one answer to that problem. Expect more of them.
Also read: Princeton's 4 Billion Parameter Queen Model Reaches 2697 Elo In Chess • Amazon Bedrock Adds Zhipu's GLM-5.3 Under a Revenue Sharing Deal • Nvidia Backs Reactor as World Model Startups Draw Big Money
Guangdong Taps Alibaba to Power Its AI and Semiconductor Push Guangdong's provincial government signed a strategic cooperation framework with Alibaba on August 13, covering AI, semiconductors, computing infrastructure and public services. The deal builds on Alibaba's Zhenwu AI chip data center already running in Shaoguan and signals Guangdong's bid to anchor China's chip and AI buildout. - alibaba semiconductor manufacturing Guangdong - provincial AI infrastructure investment China
This article is posted in AI News, check it out for more related stories.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.