SEATTLE — June 18, 2025
Amazon Web Services (AWS) is seeing clear results from its ambitious custom silicon strategy, as growing adoption of its Trainium and Graviton chip families starts to erode Nvidia’s long-standing dominance in AI infrastructure.
In recent months, AWS has reported strong momentum for its internally developed Trainium chips, designed specifically for machine learning training. The demand has grown so rapidly that supply is struggling to keep up. Meanwhile, anticipation is building for Trainium3, the next-generation chip expected to double performance over its predecessor while reducing energy consumption by 50%.
“We’re building chips that are optimized for the workloads of today and tomorrow,” said Rami Sinno, Director of Engineering at AWS Annapurna Labs, the Israeli team behind the company’s silicon efforts. “Trainium3 is coming up this year, and it’s going to bring huge performance-per-watt improvements.”
AWS’s in-house chips now span the entire compute stack. From Graviton4—a new general-purpose Arm-based CPU capable of 600 Gbps of network bandwidth—to Trainium, which powers large-scale AI training workloads for AWS clients like Anthropic, the company has created an ecosystem that is vertically integrated and performance-tuned.
Compared to Nvidia GPUs, AWS claims its chips offer 30–40% better cost-efficiency on common AI tasks. This is especially important for cloud clients operating at scale, where even marginal improvements translate into significant savings.
This performance edge isn’t just technical—it’s strategic. As Nvidia’s hardware becomes increasingly expensive and in limited supply, companies like AWS, Google, and Microsoft are turning to custom accelerators to control costs, improve scalability, and reduce reliance on third-party vendors.
An internal AWS report also suggests that many clients are now training large models entirely on Trainium chips. According to sources familiar with the matter, AWS is already experimenting with Stargate, a new initiative to build one of the world’s largest AI data centers powered by its own silicon.
Despite Nvidia’s dominance with its CUDA ecosystem and recent Blackwell GPU launch, analysts warn that AWS’s performance gains could pressure Nvidia’s margins. “The biggest risk to Nvidia isn’t performance—it’s pricing leverage,” said one industry expert. “AWS’s stack gives it full control, from chip to rack to software.”
This shift reflects a broader trend across the tech industry. As AI models become more complex, the infrastructure behind them needs to be more specialized. AWS, by investing early in Annapurna Labs and developing its own chips, is reaping the benefits—and setting a new competitive standard in the process.
Whether this signals the beginning of a long-term erosion of Nvidia’s cloud dominance remains to be seen. But one thing is clear: AWS’s bet on silicon is no longer a side project. It’s a core strategy that’s beginning to reshape the AI landscape.
