Amazon has formalized its goal to shed the image of a "chaser" in the generative artificial intelligence (AI) market.
Peter Desantis, Amazon's head of AI, said in a CNBC interview on the 17th (local time) that while it is hard to say the company's AI models have risen to the very top of the industry, Amazon is narrowing the gap by refining its data, model architecture, and computing infrastructure. The message is that Amazon will fully join the race for high-performance AI models—led by OpenAI, Anthropic, and Google—within a year.
Amazon's strategy is to run its AWS AI platform Bedrock in parallel with its own foundation model family Nova. Bedrock serves as a gateway for customers to choose and use AI models from multiple companies, while Nova is the set of models whose performance Amazon directly drives up. Amazon unveiled Nova 2 in Dec. last year, and Desantis said about 50,000 customers are using Nova 2. However, he noted that Nova 2 is not yet at the stage of being regarded as one of the very best models.
Semiconductors are also a key pillar. Amazon aims to reduce its dependence on Nvidia by pushing its in-house AI chip Trainium and its server processor Graviton. Chief Executive Officer Andy Jassy mentioned in an April shareholder letter the possibility of selling Trainium-based server racks externally in the future. AWS also unveiled Nova 2, Nova Forge, and the Trainium 3 Ultra Server at last year's re:Invent event.
Nova Forge is a service that allows corporations to combine their own data to build customized models. On top of that, businesses such as installing dedicated AI infrastructure inside customer data centers—like the AWS AI Factory—are being added. As AI training and inference expense has emerged as a key variable in big tech competitiveness, some say Amazon has mounted a counteroffensive with a vertical integration strategy that bundles models, chips, and data centers.