Hyperscalers are currently facing a “compute crunch,” meaning their existing infrastructure can’t keep up with the huge demand from AI and related services. Giants like Amazon, Google, and Meta are turning to their own custom chips to close this gap. Among them, Amazon’s Trainium chips for AI training and Graviton server CPUs have become quite strong, leading CEO Andy Jassy to express great confidence in them. Jassy told shareholders that Amazon’s custom chip business is on track to reach $50 billion in annual recurring revenue (ARR), a figure surprisingly higher than what AMD and Intel have achieved.
The $50 billion ARR is an estimate of what Amazon’s custom chip business would earn if it were a separate company, providing computing power like NVIDIA. The growth of Trainium and Graviton mainly comes from AWS’s cloud services. Amazon has not yet offered its custom chips to external customers. Jassy claims that the financial benefits from AWS and their custom chips are better than NVIDIA’s. While he emphasized his commitment to NVIDIA, he also noted that customers want an option that leads in “price-performance,” and he believes Trainium offers this much better.
Amazon’s CEO also hinted at Intel’s CPU market share, stating that since the launch of the ARM-based Graviton chips, they now dominate AWS’s infrastructure. Jassy says a similar situation is happening with Trainium for training and inference. His statement carries an interesting idea that the industry might not fully grasp: Custom infrastructure isn’t meant to replace popular computing options like NVIDIA’s. Instead, it aims to fill such a massive computing gap that hyperscalers like Amazon cannot rely solely on GPU makers.
Interestingly, Amazon also suggested it might sell its server racks to other companies. This could mean it directly competes with NVIDIA and others. With “hundreds of billions” in capital expenditures planned for investment in the business, it will be fascinating to see how the Trainium + Graviton combination develops in the future.
