Amazon to Challenge Nvidia with Its Own Graviton4 and Trainium Chips

Amazon Web Services is preparing an update to its Graviton4 chip that will include network bandwidth of 600 gigabits per second. The company describes this speed as the highest available in the public cloud. Ali Saidi, a distinguished engineer at AWS, compared this speed to a machine capable of reading 100 music CDs per second. Graviton4 is a central processing unit (CPU) developed at Amazon's Annapurna Labs in Austin, Texas, and represents a victory for the company's custom chip strategy in its competition with traditional semiconductor players such as Intel and AMD.

Main Competitor Nvidia

However, the real battle is taking place against Nvidia in the field of artificial intelligence infrastructure. At the AWS re:Invent 2024 conference in December, the company announced Project Rainier—an AI supercomputer built for the startup Anthropic, in which AWS has invested $8 billion. Gadi Hutt, senior director of customer and product engineering at AWS, said that Amazon is seeking to reduce the cost of AI training and provide an alternative to Nvidia's expensive graphics processing units (GPUs). Anthropic's Claude Opus 4 AI model was launched on AWS's Trainium2 GPUs, and Project Rainier is powered by more than half a million of these chips—an order that would traditionally have gone to Nvidia.

Hutt acknowledged that while Nvidia Blackwell is a more powerful chip than Trainium2, the AWS chip offers a better price-to-performance ratio. "Trainium3 will arrive this year and double the performance of Trainium2 while saving an additional 50% in energy," Hutt said. Demand for these chips already exceeds supply, according to Rami Sinno, director of engineering at AWS Annapurna Labs. "Our supply is very, very large, but every service we build has a customer assigned to it," Sinno said.

Amazon's Big Ambitions

With the upcoming Graviton4 update and Trainium chips as part of Project Rainier, Amazon is seeking to control the entire AI infrastructure stack, from networking and training to inference. As major AI models such as Claude 4 prove that they can be successfully trained on non-Nvidia hardware, the question is not whether AWS can compete with the chip giant, but how much market share it can capture. According to an AWS spokesperson, the release schedule for the Graviton4 update will be provided by the end of June.