Amazon to pilot AI-designed material for carbon removal

By Jeffrey Dastin

(Reuters) - Amazon.com Inc plans to pilot a new carbon-removal material for data centers, which are at risk of worsening emissions from artificial intelligence systems they power, a startup behind the deal said on Monday.

In a twist, AI itself, from the startup Orbital Materials, is what designed the carbon-filtering substance, its Chief Executive Jonathan Godwin said.

"It's like a sponge at the atomic level," Godwin told Reuters. "Each cavity in that sponge has a specific size opening that interacts well with CO2, that doesn’t interact with other things."

Potential cost-savings are partly the draw. The new material adds up to an estimated 10% of the hourly charge to rent a GPU chip for training powerful AI -- a fraction of carbon offsets' price, Godwin said.

At the same time, data centers are requiring more energy to sustain AI's development and more water to keep them cool. That poses a challenge to companies like Amazon, which has committed to have net-zero carbon emissions by 2040.

Its unit, Amazon Web Services (AWS), is the world's largest cloud-computing provider by revenue. It is piloting the novel material in one data center to start in 2025 as part of its three-year partnership with Orbital, Godwin said. The agreement also provides for Orbital to use AWS technology and to make its open-source AI available to AWS customers.

Howard Gefen, general manager of AWS Energy & Utilities, in a statement said the partnership would encourage sustainable innovation. Godwin declined to state the financial terms.

Orbital, which has operations in Princeton, New Jersey and London, set up a lab about a year ago to synthesize substances that had been simulated by its AI, Godwin said. The startup aims to work with AWS to test still-more AI-generated materials to address water use and chip cooling in data centers.

Godwin co-founded the 20-person company, backed by Radical Ventures and Nvidia's venture arm among others, after helping lead materials science work for Alphabet's DeepMind until 2022.

(Reporting By Jeffrey Dastin in San Francisco; editing by Diane Craft)