Amazon is preparing a sharper push into the AI hardware market, with its cloud unit now in early talks to sell custom Trainium chips for use inside other companies’ data centers. The move would take Amazon beyond its current model, where its in-house chips are mainly offered through Amazon Web Services, and put it closer to direct competition with Nvidia, the company that still dominates AI accelerators.
The talks remain preliminary. Amazon has not named potential customers, and the company is not walking away from Nvidia hardware inside AWS. But the direction is important: Amazon no longer wants custom silicon to be only a behind-the-scenes cost advantage for its cloud business. It wants Trainium to become a larger platform for AI infrastructure.
From Cloud Feature to Chip Business
Amazon has spent years building its own chips to control cloud costs and improve performance. Graviton handles general-purpose cloud workloads, Inferentia was built for AI inference, and Trainium is aimed at training and running large AI models. Until now, those chips have mostly been part of the AWS service package rather than products sold into outside data centers.
That may now change. Amazon’s AI chief Peter DeSantis said AWS has begun discussions with companies interested in using Trainium directly. A company spokesperson described the talks as “exploratory early conversations,” which means no major public sales deal has been announced yet.
The timing makes sense. Demand for AI compute has outpaced supply across the industry. Companies training large models or running AI products at scale are looking for lower-cost alternatives to Nvidia GPUs, especially as infrastructure spending becomes one of the biggest pressure points in the AI economy.
Jassy Signals Bigger Ambition
Amazon CEO Andy Jassy has already made clear that the company sees chips as one of AWS’s next large businesses. In his 2025 shareholder letter, Jassy said Amazon’s chips business, including Graviton, Trainium, and Nitro, had crossed a $20 billion annual revenue run rate and was growing at triple-digit percentages year over year.
He also argued that the number understates the size of the opportunity because Amazon currently monetizes most of that silicon through EC2. If the chip operation were treated as a standalone business selling chips to AWS and third parties, Jassy said its annual run rate would be around $50 billion.
His most direct signal was brief but telling: “There’s so much demand for our chips.” He added that Amazon could eventually sell racks of them to third parties. The new Trainium talks suggest that idea is moving from shareholder-letter possibility to business discussion.
Why Trainium Matters
Trainium matters because AI infrastructure economics are changing. The most expensive part of many AI systems is no longer the app interface, the chatbot layer, or even the model itself. It is the compute required to train, fine-tune, and serve models at scale.
AWS says Trainium3 UltraServers can scale up to 144 Trainium chips, delivering up to 362 MXFP8 PFLOPs, 20.7 TB of HBM3e memory, and 706 TB per second of aggregate memory bandwidth. The company also says Trainium3 improves price-performance over Trainium2, while its Neuron software stack supports frameworks such as PyTorch, Hugging Face Transformers, vLLM, Ray, Amazon EKS, and AWS Batch.
Those details matter because Amazon is not simply trying to sell a cheaper chip. It must prove that developers can actually move serious workloads onto Trainium without losing too much time rewriting, debugging, or optimizing code. Nvidia’s advantage is not only its silicon. It is the years of software support and developer familiarity around CUDA.
Anthropic Gives Amazon Proof
Amazon’s strongest public proof point is Anthropic. The Claude maker has worked with Amazon since 2023 and has committed more than $100 billion over the next decade to AWS technologies. Anthropic said it has secured up to 5 gigawatts of capacity for training and deploying Claude, spanning current and future generations of Trainium.
The scale is already large. Anthropic says it currently uses more than one million Trainium2 chips to train and serve Claude. It also works with AWS on Project Rainier, one of the world’s largest AI compute clusters.
Jassy framed the partnership as validation of Amazon’s custom silicon strategy, saying: “Our custom AI silicon offers high performance at significantly lower cost for customers, which is why it’s in such hot demand.”
For Amazon, Anthropic serves two roles. It is a major AI customer that absorbs huge amounts of compute, and it gives AWS a real-world example to show other companies considering alternatives to Nvidia-heavy infrastructure.
Nvidia Still Sets Benchmark
Even with Amazon’s push, Nvidia remains the company to beat. Its GPUs power much of the modern AI boom, and its ecosystem extends far beyond chips into networking, developer tools, libraries, and production workflows. For many AI teams, Nvidia remains the safest and fastest route to deployment because the software path is familiar.
That is why Amazon’s challenge will be difficult. Selling Trainium inside AWS is one thing. Selling chips for outside data centers is another. Customers will expect hardware availability, software maturity, integration support, long-term road maps, and predictable performance across training and inference.
Amazon also has to avoid turning this into a simple anti-Nvidia story. AWS still offers Nvidia GPUs and continues to benefit when customers choose Nvidia infrastructure in the cloud. Jassy has said Amazon has a strong partnership with Nvidia and will keep making AWS a strong place to run Nvidia workloads.
A Wider Infrastructure Fight
The broader market is moving toward more chip diversity. Google continues to push TPUs. Microsoft is developing Maia. AMD is competing with Instinct accelerators. Amazon’s Trainium effort adds another large platform to a market that customers increasingly want to diversify.
The reason is simple: AI demand is still rising, and the industry cannot rely on one supplier for every major workload. DeSantis recently said AI needs “a couple more orders of magnitude” of improvement before it becomes truly transformative, adding that the industry is “just at the beginning of innovation at all layers of the stack.”
For Amazon, that stack now clearly includes chips. If Trainium becomes available outside AWS, Amazon would not just be renting AI infrastructure through the cloud. It would be trying to sell a core piece of the AI buildout itself.
That would not immediately weaken Nvidia’s lead, but it would make the next phase of AI competition more interesting. Amazon is betting that lower costs, large-scale cloud experience, and tight chip-model integration can turn Trainium from an internal AWS advantage into a serious AI hardware business.
Comments