AWS will integrate Nvidia’s NVLink Fusion interconnect into its next-generation Trainium4 AI accelerator, deepening a long-running hardware partnership as cloud providers race to build denser AI clusters.
The move, announced at AWS re:Invent in Las Vegas, targets faster chip-to-chip communication for training large models across thousands of servers. AWS did not give a launch date for Trainium4.
Under the expanded deal, AWS is designing its next Trainium4 chips, Graviton processors and Nitro virtualization layer to be compatible with Nvidia’s new NVLink Fusion links and MGX server design. In simple terms, NVLink Fusion and MGX are Nvidia’s “wiring and rack” standards that can allow many chips to act more like one big computer.
Nvidia CEO Jensen Huang said the companies aim to create “the compute fabric for the AI industrial revolution,” referencing the combination of Nvidia’s scale-up fabric with AWS custom silicon. The real-world performance will depend on software, memory models and how AWS implements the racks at scale.
The same partnership also underpins new AWS AI Factories, a service that deploys dedicated AWS AI infrastructure — including Nvidia GPUs, Trainium chips and AWS Bedrock services — inside organizations’ own data centers while AWS operates the stack. The offering targets enterprises and public-sector bodies that need high-end AI while keeping data on premises and meeting sovereignty requirements.
By integrating NVLink Fusion into its custom silicon and data-center infrastructure, AWS is positioning itself to offer high-bandwidth, low-latency compute fabrics for future AI workloads, though the practical impact will become clearer as Trainium4 and the supporting racks are released.