Bezos supports this AI startup

Producing the most sought-after semiconductors in the world requires vast amounts of energy and access to rare minerals. CuspAI, a two-year-old British startup, has raised nearly half a billion dollars on a bet that artificial intelligence can improve that process.

On Monday, CuspAI launched the AI ​​Materials Foundry, an alliance of more than 48 technology giants, industrial firms and research facilities. The goal of the body — whose members include Nvidia Corp., Meta Platform Inc., and Hyundai Motor Group — is to pool computing and scientific resources for the purpose of creating software that can help researchers develop new materials for chipmakers and other industries at a faster pace and at a lower cost than today’s methods.

To make this work, the Cambridge, UK-based company has raised $450 million in Series B financing from investors including Jeff Bezos’ fund. CuspAI co-founder and chief executive Chad Edwards said the majority of that funding will be spent supporting labs with Foundry partners in Cambridge, Singapore and the San Francisco Bay Area.

On the research front, he said, 80% of its efforts this year will go into working on semiconductors. One possible focus, he said, is to reduce or eliminate the use in chip manufacturing of rare metals with supply-chain risks, such as ruthenium and iridium.

The new funding – led by venture firm Kleiner Perkins and NEA, with “significant participation” from Bezos Expeditions – gives the startup a valuation of $2.6 billion, up from $520 million last September, according to CuspAI. Bloomberg and the Financial Times previously reported on elements of the deal.

CuspAI initially focused on finding new materials for carbon capture and water purification. But Edwards said the company decided to make the change over the past year in response to intense demand from the chip manufacturing supply chain, whose products are critical to powering artificial intelligence. A severe shortage of semiconductors, particularly memory chips, sent shares of many hardware-dependent tech companies tumbling last week. Chipmakers and semiconductor suppliers are “wildly searching for new materials,” Edwards said in an interview. “We’ve literally been pulled by all four.”

Geetika Gupta, senior director at Nvidia, said the chip maker is seeing strong demand for new materials across its markets, from energy storage systems to data center cooling. Through Foundry, Gupta said Nvidia will likely work on innovative materials in a “three-way collaboration” between CuspAI and a third partner. “All these companies,” he said, “treat this information as the secret sauce.”

CuspAI is one of a growing number of AI developers turning to science to open up new markets. Earlier AI systems have shown that they can learn from biological data such as protein shapes to potentially discover new drugs. The new models could, in theory, sift through vast swathes of molecular data to produce designs for materials that could be useful for manufacturing and renewable energy. Investors have flocked to the effort led by former OpenAI and DeepMind researchers, as well as Prometheus, the new $41 billion company launched by Bezos. Google DeepMind has also indicated that it will soon start work on content search for AI.

CuspAI is trying to set itself apart with talent. Its other co-founder Max Welling is a respected AI researcher. Yann LeCun and Geoffrey Hinton, two pioneers of modern AI technologies, are on the startup’s advisory board. On Monday, the company also named semiconductor giant and Advanced Micro Devices Inc. director Abhi Talwalkar to the board. AMD’s venture unit CuspAI joins the latest round of financing. CuspAI additionally said it has hired former Google and Apple executive John Giannandrea to work part-time setting up the startup’s California outpost.

David Fearn-Jimenez, professor of molecular engineering at the University of Cambridge, said machine learning systems “definitely” speed up the process of developing new materials. But he cautioned that designing new materials that fundamentally improve existing materials is a long and expensive process, and AI can only be of limited help. “It’s very useful,” he said. “I don’t agree that machine learning alone will produce magic.” Fairey-Jimenez has founded two materials science startups, but does not work with or advise CuspAI.

CuspAI is already familiar with how much trial and error goes into discovering AI content. One of the startup’s early projects involved working with researchers at Meta to find new materials capable of capturing planet-warming carbon dioxide and slowing climate change. To do this, CuspAI began its search with 300 trillion possible structures within a class of chemicals called metal organic frameworks, or MOFs, which contain a metal atom surrounded by carbon-based molecular chains and are particularly good at carbon capture. After this the system reduced that list to 10.

Translating that theoretical work into useful results proved more difficult. CuspAI was able to synthesize six of those 10 chemicals, Edwards said, and none performed better than chemicals already commercially available. “They were not state-of-the-art,” he said.

But Edwards emphasized that the company completed it in six months, which is a dramatic improvement compared to typical industry timelines. The startup is now using a similar process to find new materials capable of removing toxic chemicals from water “forever.” Finnish chemicals company Kemira Oy will attempt to synthesize and test 20 of these new structures this year.

Despite its high-profile partnerships and investors, none of CuspAI’s projects have provided enough candidate molecules to leave the lab. This is true for its key rivals as well.

Yet CuspAI co-founder Welling isn’t worried. He cites the lack of reliable data, testing facilities and computing power as problems that companies in the sector are struggling with – and CuspAI hopes to address with its foundries. “People underestimate the friction of actually experimenting,” he said. “It probably sounds simple, but it’s not at all.”

(Update with further context on the industry in the 14th paragraph and additional details from Edwards on the synthesis process. A previous version of this story incorrectly identified CuspAI’s focus area for the year.)