The creation of the world's most sought-after chips is an energy-intensive process that depends on scarce mineral resources. A UK-based startup, CuspAI, is betting that artificial intelligence can revolutionize this field, having secured close to $5 billion in funding just two years after its founding.
This week, the company launched the "AI Materials Foundry," a coalition of over 48 major technology firms, industrial companies, and research institutions. The initiative aims to pool computational resources and scientific expertise to develop software that helps researchers discover new materials for chipmakers and other industries faster and more cost-effectively than current methods allow.
To fuel this ambition, the Cambridge-headquartered firm has completed a $4.5 billion Series B funding round. Investors include a fund backed by Amazon founder Jeff Bezos. CuspAI co-founder and CEO Chad Edwards stated that a significant portion of this capital will support laboratories in Cambridge, Singapore, and the San Francisco Bay Area, established in collaboration with foundry partners.
Regarding research and development, Edwards noted that 80% of the company's R&D expenditure this year will focus on material discovery. He added that one of CuspAI's key objectives is to reduce or eliminate the use of rare metals like ruthenium and iridium in chip manufacturing, materials associated with significant supply chain risks.
The latest funding was co-led by venture capital firms Kleiner Perkins and NEA, with "substantial participation" from Bezos Expeditions. This investment values the startup at $26 billion, a substantial increase from its $5.2 billion valuation last September.
CuspAI initially concentrated on finding novel materials for carbon capture and water purification. However, Edwards explained that over the past year, intense demand from the chip manufacturing supply chain—crucial for supporting AI advancement—prompted a strategic pivot. "Chipmakers and semiconductor suppliers are desperately searching for new materials," Edwards said in an interview. "We were essentially pulled into this area from all sides."
NVIDIA's Senior Director, Geetika Gupta, confirmed the industry's massive demand for new materials across various applications, from energy storage systems to data center cooling technologies. Gupta indicated that through the Materials Foundry, NVIDIA is most likely to engage in "tripartite collaborations" involving CuspAI and third-party partners. "All these companies treat vast amounts of relevant information as their core proprietary secrets," she said.
CuspAI is part of a growing cohort of AI developers leveraging artificial intelligence to explore scientific frontiers and open new markets. Early AI systems have demonstrated an ability to learn from biological data, such as protein structures, with potential applications in drug discovery. Newer AI models can theoretically sift through vast molecular datasets to design novel materials for manufacturing and renewable energy sectors.
Investors are flocking to this domain, backing projects led by former Meta Platforms, Inc. and DeepMind researchers, as well as new ventures like the $41 billion-valued Prometheus, launched by Bezos. Alphabet's Google DeepMind has also announced plans to commence AI-driven material discovery research soon.
CuspAI seeks to differentiate itself through its talent pool. Its other co-founder, Max Welling, is a highly regarded AI researcher. Pioneers of modern AI, Yann LeCun and Geoffrey Hinton, serve on the company's advisory board. This week, CuspAI also announced the appointment of chip industry veteran and Apple board member Abhi Talwalkar to its own board. The venture capital arm of Apple participated in CuspAI's latest funding round.
Additionally, the company has hired former Alphabet and Apple executive John Giannandrea on a part-time basis to establish its California operations.
Professor David Fairen-Jimenez of Cambridge University, a specialist in molecular engineering, acknowledged that machine learning systems "have indeed" accelerated the new material development process. However, he cautioned that designing materials that significantly outperform existing ones remains a lengthy and expensive endeavor where AI offers only limited assistance. "It is very useful. What I disagree with is that machine learning alone will produce some kind of magical breakthrough," said Fairen-Jimenez, who has founded two materials science startups but is not currently collaborating with or advising CuspAI.
CuspAI is already familiar with the extensive trial and error inherent in AI-driven material discovery. One of its early projects, in collaboration with Meta Platforms, Inc. researchers, aimed to find new materials to capture carbon dioxide and mitigate climate change. For this, the company built a database of 300 trillion potential structures from a class of chemicals known as metal-organic frameworks (MOFs), which are particularly promising for carbon capture.
The system narrowed this vast list down to just 10 candidates. However, translating theoretical results into practical applications proved more challenging. Edwards revealed that CuspAI ultimately synthesized fewer than 10 of these compounds, none of which outperformed existing commercial products. "They were not state-of-the-art materials," he said.
Undeterred, CuspAI is applying the same database approach to find materials that can remove toxic "forever chemicals" from water. Finnish chemical company Kemira Oy will attempt to synthesize and test 20 of these new structures later this year.
Despite its impressive roster of partners and investors, none of CuspAI's projects have yet yielded a candidate molecule ready for commercialization and scale-up beyond the lab. Co-founder Welling, however, remains optimistic. He points to the chronic challenges in the field—scarce reliable data, inadequate testing facilities, and insufficient computing power—as issues the Materials Foundry is designed to address. "People underestimate the friction involved in actually doing experiments," Welling noted. "It might look simple, but it's really not."
Comments