Public cloud infrastructure giant Amazon Web Services Inc. may become the first major customer to adopt new artificial intelligence chips made by Advanced Micro Devices Inc., according to a report from Reuters.

However, the company has decided not to buy entire systems from AMD’s rival, Nvidia Corp.

AWS, the world’s largest cloud computing company, is reportedly considering AMD’s new Instinct MI300X accelerator, which was announced at a special event yesterday. The MI300X chip is designed specifically for generative AI workloads, and is meant to provide companies with an alternative to Nvidia’s graphics processing units, which currently dominate the AI industry.

In an interview today with Reuters, Dave Brown, vice president of elastic compute cloud at AWS, said the company has not yet made a decision on whether to use AMD’s new chips.

AMD’s shares rose about 1% following the report. On Tuesday, they declined almost 4% after the company announced the MI300X chip but failed to disclose any major customers that had agreed to use it.

AMD Chief Executive Lisa Su told Reuters after the event that her company is particularly interested in winning over cloud providers such as AWS. To do so, it’s offering a menu of all of the pieces required to build the systems required to power services such as ChatGPT. Customers would be able to choose which components they want to use, and slot them together with industry standard connections.

According to Su, “a lot of people are going to want choice, and they’re going to want the ability to customize what they need in their data center.”

Brown told Reuters that AWS and AMD are currently working together to see how it could take advantage of the new MI300X chips. “That’s where we’ve benefited from some of the work that they’ve done around the design that plugs into existing systems,” he said.

AMD’s case might be helped by the fact that AWS likes to build its own servers from the ground up. As such, it would make sense for AWS to diversify the AI hardware it uses to power those servers, said Holger Mueller of Constellation Research Inc.

“The arms race among cloud computing vendors has opened a new front with the rise of generative AI, and they’re all scrambling for the chips they need to run those workloads,” Mueller said. “The cloud vendors know that some customers may have proprietary workloads that run better on a different AI chipset, so they want to give them more options. But having conversations about new chips is one thing, and building the actual mission-critical hardware is quite another, so only the future will tell if this leads to anything.”

AWS’s need to diversify its offerings also explains why it declined to offer an entire system designed by Nvidia called the DGX Cloud. That system is an “AI supercomputer” that bundles hundreds of GPUs, ready made to handle some of the most intensive AI workloads. It’s currently offered by Oracle Corp. and Nvidia reportedly asked AWS if it would also be willing to offer the DGX platform to its customers. But according to Brown, AWS declined, saying it prefers to buy Nvidia’s GPUs piecemeal.

AWS does offer customers access to Nvidia’s most advanced H100 GPU, but only in server systems AWS has designed.

Photo: Tony Webster/Flickr