Amazon is expanding its AI newchip footprint, developing its own Trainium chips and deploying a massive AI accelerator cluster (Project Rainier) with nearly half a million Trainium2 units, which could loosen Nvidia's grip on the AI data-center market through stronger in-house capabilities and partnerships like Anthropic.

This shift places pressure on Nvidia as major cloud customers reduce reliance on third-party GPUs by building or collaborating on in-house AI silicon and optimized systems.

The common thread across customers is a move toward internal or collaborative AI chip solutions, raising the risk that demand for Nvidia GPUs in data centers could erode over time.

Alphabet's Gemini 3 training on its own TPUs demonstrates a viable path for hyperscalers to scale AI workloads without dependence on Nvidia, signaling a broader industry trend.

The Gemini 3 training workflow using Alphabet's TPUs suggests GPUs are not indispensable for large-scale AI training, enabling custom hardware to compete with Nvidia.

The AI market is expanding rapidly, with projections from about $235 billion to roughly $631 billion by 2028, a trajectory that has historically benefited Nvidia but now faces increasing competition for data-center workloads.

Microsoft is reducing its reliance on Nvidia GPUs by deepening collaboration with OpenAI and pursuing joint development of custom AI chips, including OpenAI’s Broadcom chip efforts, which could challenge Nvidia’s dominance.