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

Amazon Web Services (AWS) is on the verge of revealing an upgrade to its Graviton4 chip, boasting an impressive network bandwidth of 600 gigabytes per second, which AWS claims to be the highest available in the public cloud sector. Ali Saidi, a prominent engineer at AWS, compared this remarkable speed to a machine processing 100 music CDs each second. The Graviton4 chip, a central processing unit (CPU), is one of several innovations emerging from Amazon's Annapurna Labs located in Austin, Texas. This development underscores the company's commitment to its custom chip strategy as it aims to compete with established semiconductor giants like Intel and AMD. However, the most significant rivalry lies with Nvidia in the realm of artificial intelligence infrastructure. During the AWS re:Invent 2024 conference last December, the company unveiled Project Rainier, an AI supercomputer designed specifically for the startup Anthropic. AWS has committed a substantial $8 billion to support Anthropic’s endeavors. Gadi Hutt, AWS Senior Director for Customer and Project Engineering, emphasized the company’s goal to lower AI training costs while providing a viable alternative to Nvidia’s costly graphics processing units (GPUs). Anthropic's Claude Opus 4 AI model, as per AWS, is being trained on Trainium2 GPUs, while Project Rainier is powered by over half a million of these chips, representing a significant shift in orders that traditionally favored Nvidia. Hutt noted that although Nvidia's Blackwell chip outperforms Trainium2, AWS’s chip offers superior cost efficiency. "Trainium3 is set to launch this year, doubling the performance of Trainium2 and achieving a further 50% reduction in energy consumption," he stated. The demand for these chips is already surpassing supply, according to Rami Sinno, director of engineering at AWS's Annapurna Labs. "Our supply is substantial, but every service we develop has a customer linked to it," he clarified. With the Graviton4 upgrade imminent and Project Rainier’s Trainium chips making headway, Amazon is showcasing its ambition to dominate the entire AI infrastructure landscape, from networking to training and inference. As more significant AI models like Claude 4 demonstrate their training capabilities on non-Nvidia hardware, the real question shifts from whether AWS can rival the chip titan to how much market share it can capture. An AWS spokesperson confirmed that the release timeline for the Graviton4 update will be announced by the end of June.