Shares of Amazon (AMZN 1.30%) plunged on Friday after the company disclosed it would ramp up capital expenditures to $200 billion this year. Amazon is betting big on AI infrastructure: "...we are monetizing capacity as fast as we can install it," said Amazon CEO Andrew Jassy in the fourth-quarter earnings call.
While the jury is still out on whether Amazon's prolific AI spending will pay off in the long run, some of that spending will go toward a business that isn't getting much attention from investors. Amazon has installed 1.4 million of its Tranium2 AI chips in its data centers, and the results have been spectacular. Revenue from Amazon's custom chips, which include Tranium and Graviton CPUs, has reached an annual run rate of $10 billion and is growing by more than 100% annually.
Amazon is still a prolific buyer of Nvidia's (NVDA 1.21%) GPUs, but the success of Tranium puts a crack in Nvidia's long-term growth story.
Image source: Getty Images.
Lowering costs for AWS customers
Nvidia dominates the AI accelerator market, and its chips are wildly expensive. Many AI infrastructure providers have little choice but to pay up for Nvidia's chips.
For AI companies training advanced AI models, the high initial cost and high operating costs of Nvidia's GPUs for AI infrastructure providers translate into higher training costs. Amazon claims that its Tranium chips offer 30% to 40% improved performance-per-dollar than comparable GPUs.
The proof is in the pudding. AI start-up Anthropic, which recently made headlines with its groundbreaking Cowork product that sent software stocks tumbling, is using Amazon's Tranium2 chips to train its next-generation AI models. AWS's Project Rainier, which currently features 500,000 Tranium2 chips and will eventually scale to 1 million chips, is used by Anthropic to both train and run its Claude family of AI models.
Amazon's next-gen Tranium3 will be deployed soon, offering a 40% improvement in performance-per-dollar over Tranium2. The company's fleet of Tranium2 chips is fully sold out, and it expects to sell out Tranium3 capacity by mid-2026. The company is also working on Tranium4.
Amazon has also installed its custom Graviton CPU into its data centers, which is itself a multi-billion-dollar business. The company claims that Graviton delivers up to 40% more performance per dollar than leading x86 CPUs and that 90% of AWS's top 1,000 customers use the custom chip.
Nvidia's dominance is slowly eroding
Amazon doesn't sell its custom chips directly, but every Tranium chip installed in its data centers represents an Nvidia GPU that it's not buying. The company isn't alone. Alphabet has been developing its TPU AI accelerators for years, and Microsoft is now on the second generation of its Maia AI chips.
These custom AI chips are particularly efficient at AI inference, or the act of running a trained AI model. As AI usage expands, particularly in multi-step agentic AI workloads, AI inference is likely to overtake AI training in terms of overall compute capacity. With custom AI chips proliferating, Nvidia isn't the only game in town for AI infrastructure providers with the resources to design their own AI chips.
NASDAQ: AMZN
Key Data Points
Increased AI adoption will depend on costs coming down, and custom AI chips designed with efficiency in mind could help achieve that. More powerful AI models are enabling new use cases, like capable AI coding assistants that can write most of a developer's code, but broader adoption may be hindered by the high cost of running those models.
Amazon and the other cloud giants will continue to buy Nvidia GPUs in gargantuan quantities as they rapidly expand their AI computing capacity, but having alternatives gives them some leverage. Nvidia may see its profit margins come under pressure as the AI accelerator market becomes more competitive, and as companies like Amazon lean more heavily on their homegrown AI chips.
