Amazon Web Services (AWS) has unveiled its latest data center servers, powered by its own AI chips, the Trainium2. This move positions Amazon as a formidable contender against Nvidia, the current leader in AI chip sales. The announcement was made at a tech event in Las Vegas, where AWS Chief Executive Matt Garman highlighted the advanced capabilities of the new servers, which are designed for cutting-edge generative AI workloads.
The new Trn2 UltraServers, featuring 64 Trainium2 chips, will form part of a supercomputer comprising hundreds of thousands of chips. AI startup Anthropic will be the first to utilize this technology. Notably, Apple has also joined the ranks of AWS customers, with Apple executive Benoit Dupin confirming their use of Trainium2 chips.
Nvidia, holding over 70% of the AI chip market, faces increasing competition from industry giants like Meta Platforms, Microsoft, and Google’s Alphabet, all of which are developing custom AI chips for internal use and external customers. AWS’s strategy to market its chips aligns with Google’s approach, as both companies aim to offer cost-effective and powerful AI solutions.
AWS plans to launch Trainium3, its next-generation AI chip, next year, further intensifying the competition. Gadi Hutt, head of business development for AWS AI chips, emphasized that Trainium2 offers more computational power than Nvidia’s current offerings, potentially reducing training costs for AI models by up to 40%.
While AWS and Nvidia both rely on Taiwan Semiconductor Manufacturing for chip production, AWS claims to have a robust supply chain, poised to meet the growing demand. However, Nvidia’s shipments have faced delays due to supply chain constraints.
The new AWS servers are set to be operational next year, promising enhanced capabilities for AI training and inference while offering a competitive edge in performance and cost-efficiency. As AWS and Nvidia race to satisfy the surging demand for AI technologies, the landscape of AI chip manufacturing is poised for significant shifts.
