For years, Amazon built AI chips exclusively for itself. Trainium and Inferentia powered AWS workloads only — customers could rent capacity through the cloud, but buying the hardware outright was never an option. According to Bloomberg, that's now under discussion.
The company is in early talks to sell Trainium3 processors directly to outside data centers. Peter DeSantis, Amazon's AI chief, describes the chip as delivering four times the performance of its predecessor at roughly half the cost of conventional GPUs. Since its late-2025 launch, Trainium3 has run at near-full capacity — meaning demand already exceeds what Amazon can supply internally.
Andy Jassy put a number on the opportunity in his April shareholder letter: if Amazon sold its chips externally rather than running them only inside AWS, the annual revenue run rate would reach approximately $50 billion. The chip division already generates $20 billion a year through AWS alone. A direct sales model would shift Amazon from cloud provider to hardware vendor, putting it squarely in competition with Nvidia.
Nvidia controls roughly 80% of the AI accelerator market, with its H100 and H200 cards treated as the industry default. The pricing reflects that position. Trainium3 offers a cost-performance argument for data centers that want powerful AI hardware without locking themselves into the AWS ecosystem. That's a challenge to Nvidia on both the technology and the business model front.
Talks are early and no partners or timelines have been announced. But the direction is hard to miss. When the world's largest cloud provider starts treating its chip business as a potential standalone operation, it suggests the AI hardware market — currently tilted almost entirely toward Nvidia — has more room to shift than present valuations imply.
