Amazon’s Trainium3 accelerator chip — the company’s first AI silicon built on 3nm technology — is now in general release. It comes to market with the Trainium3 UltraServer, a high-density integrated system purpose-built for large-scale GenAI model training. Trn3 UltraServers can scale to 144 Trainium3 chips, delivering up to 362 FP8 PFLOPs. The Trn3 chips are viewed as challengers to Nvidia’s AI GPUs and Google’s Tensor TPUs. But Amazon also announced it will provide enterprises with turnkey AWS AI Factories that utilize key Nvidia components and support its AI chips, allowing customers to choose or combine brands.

An AWS AI Factory allows customers to use their own data center buildings and power sourcing with AWS handling equipment deployment and management.

“Large-scale AI requires a full-stack approach,” Nvidia VP and GM of Hyperscale and HPC Ian Buck said in Amazon’s Factory announcement. By offering Nvidia’s Grace Blackwell and Vera Rubin chips, AI software and apps alongside Amazon’s Trn3 and ecosystem of AI products, customers can apply the best solution to individual applications.

“AWS will support Nvidia NVLink Fusion high-speed chip interconnect technology in next-generation Trainium4 and Graviton chips and in the Nitro System,” Amazon explains. AWS is also designing Trainium4 to integrate with NVLink 6 and the Nvidia MGX rack architecture, Nvidia adds in its own blog post.

“The idea is to cater to companies and governments concerned with data sovereignty, or absolute control over their data so it can’t wind up in a competitor’s or foreign adversary’s hands,” reports TechCrunch, adding that an on-premises AI Factory “means not sending their data to a model maker and not even sharing the hardware.”

The Wall Street Journal calls the Trn3 chips “the latest broadside against Nvidia,” providing a use case in which San Francisco-based AI video startup Decart is using Amazon’s new silicon to render AI-generative footage “in real-time, without bugs or hiccups.”

Trn3 “can reduce the cost of training and operating AI models by up to 50 percent” compared with systems that use equivalent GPUs, WSJ writes.

In an interview with Bloomberg, AWS VP Dave Brown said the company plans to “scale out very, very quickly” with the new chips in early 2026. Many will be used in Trainium3 UltraServers in the Amazon EC2 (Elastic Compute Cloud) as well as AI Factory facilities, improving performance by up to 4.4x, per the Trn3 announcement.

“The chip push is a key element of Amazon’s strategy to stand out in AI,” writes Bloomberg, noting that “AWS is the largest seller of rented computing power and data storage, but it has struggled to replicate that dominance among leading developers of AI tools.”