Amazon CEO Andy Jassy revealed in his annual shareholder letter published on April 9 that the company is considering changing its current approach of using custom chips exclusively for internal cloud services, and instead selling entire rack-scale chip units directly to third-party markets. This strategic shift marks Amazon's evolution from a pure compute rental provider to a hardware supplier directly competing with NVIDIA and Intel.
The company's internal chip division currently generates an annual revenue of over $20 billion. Jassy predicted that if it were operated as an independent business for external sales, its potential annual revenue could reach up to $50 billion.
Amid the overflow of AI computing demand, Amazon's custom chips have shown strong market appeal due to their high cost-effectiveness. The second-generation AI training chip, Trainium2, has a cost-performance ratio 30% better than similar GPUs and has been nearly sold out; while the upcoming Trainium3, which is scheduled to ship in early 2026 and will offer a 30% to 40% performance improvement, has already reached near full subscription levels.
Jassy further revealed that a significant portion of the production capacity of Trainium4, which is still in development and about 18 months away from launch, has already been booked. In addition, Amazon's custom CPU Graviton has been adopted by 98% of top EC2 customers, and its successful path is being replicated in the AI accelerator field.
Amazon's move reflects the deep anxiety of cloud computing giants regarding compute bottlenecks. Currently, AWS is facing severe capacity limitations and plans to double its total power capacity by the end of 2027. By offering its underlying hardware capabilities to the outside, Amazon can not only create a second growth curve but also secure a core position in the global AI compute landscape, achieving deep control over the supply side of the era of large models at the infrastructure level.
