Amazon Unveils New AI Chip, Hints at Nvidia Partnership
▼ Summary
– AWS announced its new Trainium3 AI training chip and UltraServer system, which are over four times faster and have four times more memory than the previous generation.
– The Trainium3 system is also 40% more energy efficient, a key focus as AWS aims to reduce data center power consumption.
– AWS revealed its next-generation chip, Trainium4, is already in development and will feature interoperability with Nvidia’s chips via NVLink technology.
– Early customers like Anthropic have used Trainium3 to significantly cut their AI inference costs, according to AWS.
– The planned Trainium4 systems are designed to attract major AI applications built for Nvidia’s industry-standard CUDA platform to AWS.
Amazon Web Services has officially launched its next-generation artificial intelligence training chip, the Trainium3, alongside a powerful new system called the UltraServer. This announcement, made at the AWS re:Invent 2025 conference, marks a significant step in the cloud giant’s strategy to provide high-performance, cost-effective alternatives for developers building and running AI models. The new hardware promises substantial improvements in speed, memory, and critically, energy efficiency, as the industry grapples with the immense power demands of advanced AI.
The Trainium3 chip, fabricated on an advanced 3-nanometer process, powers the new UltraServer systems. According to AWS, this third-generation technology delivers a performance leap that is more than four times faster than its predecessor, with a corresponding quadrupling of memory capacity. This boost applies not only to the initial training of AI models but also to efficiently serving those models to users during periods of peak demand. The architecture allows for remarkable scalability; thousands of these UltraServers can be interconnected, creating a cluster offering access to up to 1 million Trainium3 chips, a tenfold increase over the previous generation’s maximum capacity.
A major focus of this release is operational sustainability. AWS emphasizes that the new chips and systems are 40% more energy efficient than the last generation. In an era where data center electricity consumption is soaring, this improvement addresses both environmental concerns and the bottom line. By designing systems that consume less power, AWS aims to reduce operational costs, savings it states will be passed on to its cloud customers. Early adopters like Anthropic, Karakuri, SplashMusic, and Decart have reportedly already leveraged the technology to significantly cut their inference costs.
Looking ahead, AWS provided a glimpse into its future roadmap by confirming that Trainium4 is already in development. This next iteration is promised to deliver another substantial performance upgrade and will incorporate a key feature: support for Nvidia’s NVLink Fusion high-speed interconnect technology. This strategic move is particularly noteworthy. It means future AWS systems powered by Trainium4 will be designed to interoperate seamlessly with Nvidia’s industry-leading GPUs, all while utilizing Amazon’s own cost-optimized server infrastructure.
This compatibility could prove crucial for attracting major AI workloads. Since many complex AI applications are built around Nvidia’s CUDA platform, the ability for Trainium4 systems to work directly with Nvidia hardware removes a potential barrier to adoption. It allows customers to leverage the strengths of both chip ecosystems within the AWS cloud. While Amazon did not provide a specific release date for Trainium4, industry observers anticipate more details will emerge at next year’s re:Invent conference, following the company’s typical product cycle.
(Source: TechCrunch)
