Google DeepMind's GNoME × Berkeley's A-Lab × a buyer not yet public: after 2.2 million predictions, who states the need

One of the Buyer Chains series. Each piece follows one chain, the AI company, its customer and sometimes the customer's customer, through five roles in the public record: who states the need, who sets the research direction, who runs the experiments, who validates the result, who decides to keep paying. It records only what the public record states and says who said each thing; a missing step is marked not yet public rather than filled in, and a company's own figure is written as the company's. The timeline at the end stays open: when the chain's next step reaches the public record, it goes on as a new line.

This chain runs in the opposite order to the other seven: the result came first and the need is still being looked for. On 29 November 2023 Google DeepMind published GNoME in Nature, reporting 2.2 million candidate crystal structures, 380,000 of them judged most stable; the same day Berkeley's A-Lab published its autonomous lab's results, 353 experiments in 17 days.[1][2] For the next three years the list was validated almost entirely by outsiders: in 2024 Cheetham and Seshadri examined a random sample and found scant evidence of compounds at once new, credible and useful; Leeman, Schoop, Palgrave and colleagues re-examined all 43 A-Lab products and concluded no new material had been made.[1][2] In October 2025 a classifier trained on known crystals estimated that 80 to 84% of GNoME's public structures are likely disordered[3]; in January 2026 Nature published the A-Lab authors' correction, dropping "novel" and confirming 36 of 40 syntheses[4]; in October 2026 a Liverpool audit found 63 to 78% of the predicted structures within 0.05 Å of a known crystal's geometry[5]. Someone also made something from the list: in March 2026 two researchers at Fukuoka Institute of Technology picked MnFeCo4Si2, arc-melted and annealed it and measured its magnetism, getting the predicted structure and a soft magnet with a Curie temperature of 1039 K, one compound, one lab, one preprint.[6] In December 2025 DeepMind announced its first automated lab, for materials, to open in the UK in 2026[7], and the same month created a senior product manager role for Material Intelligence, leading commercial strategy for its AI materials-discovery initiative[8].

What the buyer looks at, concretely

• For a list, a buyer's first question is not a property but whether the thing can be made and is what it says: disorder (80 to 84% likely disordered, itself a prediction)[3], geometric novelty (63 to 78% within 0.05 Å of a known crystal)[5], synthesisability (ChronoSyn orders by year of first report and recovers 52.4% of a held-out set against 48.4% with the years shuffled)[11]. All of these run before anyone touches a furnace.
• For a material, a buyer wants one property and one comparison: the Fukuoka magnet gives a Curie temperature of 1039 K with no incumbent to compare; the NASICON electrolytes give a conductivity, with total conductivity 6 to 11 times below bulk and no equal-budget baseline.[6][10]
• DeepMind's own next steps are written in two places: the first material its UK lab makes and measures, and the first brief its commercial-strategy role brings in. The first is a 2026 item on the company's own date; the second is not yet public.[7][8]

Not yet public · Any buyer; the UK lab's opening date and first material; the commercial role's first customer project; how many entries on the GNoME list anyone has made and tested to a specification.

Decision point · The next decision is DeepMind's own: what the UK lab works on, and which buyer the commercial-strategy role brings the first brief from. No outside buyer has made any decision in this chain yet.

The chain's timeline

• 2023-11GNoME: 2.2 million candidates, 380,000 most stable; A-Lab: 353 experiments in 17 days (Nature)[1][2]
• 2024Outside re-examinations: Cheetham and Seshadri; Leeman, Schoop and Palgrave (the record)[1][2]
• 2025-10Classifier estimates 80 to 84% likely disordered (Advanced Materials)[3]
• 2025-12DeepMind announces a UK automated materials lab for 2026[7]; creates the Material Intelligence product role for commercial strategy[8]
• 2026-01Nature corrects A-Lab: "novel" dropped, 36 of 40 confirmed[4]
• 2026-03Fukuoka makes MnFeCo4Si2 from the list, Curie temperature 1039 K (preprint)[6]
• 2026-09A-Lab's CASS finds two NASICON electrolytes (preprint)[10]
• 2026-10Liverpool audit: 63 to 78% within 0.05 Å of a known geometry (preprint)[5]; ChronoSyn (preprint)[11]
• NextThe first material the UK lab makes and measures (2026, the company's date). Awaiting public data.
• LaterThe first brief from a buyer.

All cases →