Amazon is exploring the possibility of making its own AI processors available directly for use in customers’ data centers. With this move, the company is taking the next step in its ambition to offer an alternative to Nvidia’s dominant position in the AI infrastructure market.
According to Peter DeSantis, senior vice president in charge of AWS Utility Computing, Amazon is currently in talks with potential customers. According to Bloomberg, the company no longer intends to offer its Trainium chips exclusively through Amazon Web Services but is also looking into supplying hardware for use in environments outside its own cloud.
Trainium Is Growing Into a Multi-Billion-Dollar Business
Amazon introduced the first Trainium chip in 2020 as an accelerator for AI training and inference. Through AWS, organizations such as OpenAI, Anthropic, and Uber are already using the hardware for a variety of AI workloads.
CEO Andy Jassy hinted earlier this year at broader availability of the technology. In his annual shareholder letter, he wrote that Amazon could eventually supply complete racks of Trainium systems to external parties. In doing so, he provided a striking indication of the potential scale of the business. If the chip division were to operate as an independent company serving both AWS and external customers, Jassy estimated it would generate annual revenue of approximately $50 billion.
That is still considerably smaller than Nvidia, which is currently on track for annual revenue of more than $300 billion, but it would immediately make Amazon one of the largest players in the semiconductor sector.
Why Amazon Has Been Reluctant Until Now
The question arises as to why Amazon hasn’t sold its chips to third parties sooner. According to additional information from TechCrunch, AWS earns much more from a Trainium chip within its own cloud environment than from a standalone hardware sale.
After all, in addition to computing power, customers also purchase additional services, such as storage, networking, security, and monitoring. As a result, an AI workload within AWS generates a significantly larger revenue stream than the sale of a processor alone. This explains why Amazon has so far primarily used the chips as a strategic tool to enhance the appeal of AWS.
Production Remains a Challenge
Another factor is chip availability. DeSantis noted that the current generation, Trainium3, is largely sold out and that there is already significant interest in the next generation, scheduled for release next year.
TechCrunch adds that Amazon may face the same constraint as other AI chip developers: production capacity. The chips are manufactured by TSMC, which also handles a large portion of Nvidia’s production. If Amazon wants to sell Trainium on a large scale outside of AWS, it will need to secure additional production capacity in a market where demand for advanced chip manufacturing is at an all-time high.
Sovereign AI as an Additional Growth Market
A key driver behind these plans appears to be the growing demand for locally managed AI infrastructure. In Europe in particular, organizations are seeking ways to host AI capacity within national or regional borders without being entirely dependent on U.S. cloud platforms.
By making Trainium available outside of AWS as well, Amazon can capitalize on that market. Organizations would then be able to deploy the chips in their own data centers or with regional cloud providers.
With the potential sale of Trainium systems, Amazon is taking a new step in the competition surrounding AI infrastructure. While Nvidia has been virtually unchallenged for years, hyperscalers are increasingly striving to build their own ecosystems of AI hardware. The question here is not only who develops the best chips, but also who can secure sufficient production capacity to meet the explosively growing demand.