AWS’ custom chip strategy is showing results, and cutting into Nvidia’s AI dominance

Amazon Web Services is set to announce an update to its Graviton4 chip that includes 600 gigabits per second of network bandwidth, what the company calls the highest offering in the public cloud.

Ali Saidi, a distinguished engineer at AWS, likened the speed to a machine reading 100 music CDs a second.

Graviton4, a central processing unit, or CPU, is one of many chip products that come from Amazon’s Annapurna Labs in Austin, Texas. The chip is a win for the company’s custom strategy and putting it up against traditional semiconductor players like Intel and AMD.

But the real battle is with Nvidia in the artificial intelligence infrastructure space.

At AWS’s re:Invent 2024 conference last December, the company announced Project Rainier – an AI supercomputer built for startup Anthropic. AWS has put $8 billion into backing Anthropic.

AWS Senior Director for Customer and Product Engineering Gadi Hutt said Amazon is looking to reduce AI training costs and provide an alternative to Nvidia’s expensive graphics processing units, or GPUs.

Anthropic’s Claude Opus 4 AI model launched on Trainium2 GPUs, according to AWS, and Project Rainier is powered by over half a million of the chips – an order that would have traditionally gone to Nvidia.

Hutt said that while Nvidia’s Blackwell is a higher-performing chip than Trainium2, the AWS chip offers better cost performance.

“Trainium3 is coming up this year, and it’s doubling the performance of Trainium2, and it’s going to save energy by an additional 50%,” he said.

The demand for these chips is already outpacing supply, according to Rami Sinno, director of engineering at AWS’ Annapurna Labs.

“Our supply is very, very large, but every single service that we build has a customer attached to it,” he said.

With Graviton4’s upgrade on the horizon and Project Rainier’s Trainium chips, Amazon is demonstrating its broader ambition to control the entire AI infrastructure stack, from networking to training to inference.

And as more major AI models like Claude 4 prove they can train successfully on non-Nvidia hardware, the question isn’t whether AWS can compete with the chip giant — it’s how much market share it can take.

The release schedule for the Graviton4 update will be provided by the end of June, according to an AWS spokesperson.

CORRECTION: A previous version of this story misstated the nature of the bandwidth for the updated Graviton4 chip.