Amazon Adopts NVIDIA’s NVLink, Unveils New Servers in AI Push

Amazon Web Services (AWS) plans to use key technology from Nvidia in its homegrown processors and has also announced servers that use AWS Trainium3 chips, all as part of a larger, more intense focus on AI. While AWS is undoubtedly the world’s largest cloud-service provider, it faces increasing competition from rivals such as Google and Microsoft, both of which have either developed or announced chips designed to meet the new soaring demand for AI computing. 

NVLink and UltraServers

Announcing the plans at the AWS re: Invent event in Las Vegas on December 2, AWS signalled that while Nvidia remains central to AI, hyperscalers are under increasing pressure to offer cheaper and readily available alternatives. AWS disclosed that it would be incorporating Nvidia’s NVLink high-speed interconnect technology in multiple forthcoming chips.

The new servers built around Trainium3 chips are being called UltraServers and come with 144 processors in a single unit. They also consume 40% less power while offering more than four times the performance of the previous generation of machines, according to AWS. These efficiency claims hold a lot of weight, especially as AI is quickly becoming the most power-hungry task in the world today.

Creating a viable Nvidia alternative

It’s understandable that some hyperscalers are looking to diversify their supply of AI accelerators, given current issues with supply and high costs. Dave Brown, a vice president at AWS, said the strategy is focused on delivering the performance customers need at the right price point, while expanding the range of alternatives available for running AI workloads.

In other words, AWS isn’t looking to the deployment of Trainum as a hard pivot away from Nvidia, but wants customers to see it as an option that sits alongside the Nvidia GPU-based instances. In fact, the Nvidia partnership on NVLink is an interesting indicator of where AWS sees the upcoming infrastructure bottlenecks. As customers are training more frontier models, they’ll want more than just raw compute; they’re going to be looking for faster interconnectivity between chips.

NVLink is designed to address that. AWS says we can expect Trainum 4 and future AWS chips to use Nvidia tech in order to strengthen that interconnect layer further. 

Balancing custom silicon with Nvidia’s dominance

AWS has been investing pretty heavily in building its own chips for a while (like Graviton), mainly to cut costs in its cloud datacenters. With generative AI and higher expectations from customers, however, the AI market is moving simply too quickly.  Many enterprises are already deeply invested in Nvidia-based software ecosystems, while also looking for cheaper, more available capacity as they expand AI experiments into production.

The rollout underscores how the AI cloud market is evolving. AWS is not abandoning Nvidia; it is building a parallel path. By pairing Nvidia’s interconnect with its own chips and packaging them into large server systems, AWS is trying to offer both scale and efficiency at a time when AI demand is rewriting the economics of cloud computing.