Amazon to add 2 million Nvidia GPUs for cloud AI workloads while continuing to build its own chips

Amazon Web Services is continuing to scale its use of Nvidia graphics processing units, according to a report that says the cloud unit plans to add 2 million Nvidia GPUs. The move highlights a practical reality of today’s AI infrastructure buildout: even large cloud providers investing in custom chips often keep buying major third-party accelerators to meet near-term demand and to support existing software and model pipelines.

The same report frames the GPU expansion as happening alongside Amazon’s efforts to build its own AI chips. In practice, that can mean AWS uses custom silicon for some workloads while still relying on Nvidia for others, such as training runs that depend on mature performance optimizations or for customers that want consistent results across hardware.

For Nvidia, cloud customers are a central driver of data center demand. Nvidia’s high-end AI GPUs are widely used for training large models and for running them at scale, and cloud purchases tend to translate quickly into near-term procurement activity. A plan to add a large batch of GPUs is also notable because it suggests AWS is not treating custom hardware as a complete replacement for Nvidia in the immediate term.

For Amazon, continuing to buy Nvidia GPUs while developing its own accelerators can be read as a hedge. Custom chips can offer cost and performance advantages once the company and customers have validated them across common use cases. But that validation takes time, and cloud providers often maintain capacity on multiple platforms to avoid service gaps or to support a broader range of customer requirements.

The report does not, in the available material here, specify how AWS will allocate the additional Nvidia GPUs between training and inference, what customers or services the capacity will support, or whether the GPUs are destined for specific regions or data center clusters. It also does not disclose any internal target dates for how quickly Amazon expects its custom AI chips to reduce GPU consumption.

More broadly, the cloud AI race has increasingly become a two-track strategy. Providers build custom chips to lower effective costs and to differentiate their hardware stack, but they also keep buying Nvidia and other ecosystem hardware because the software ecosystem, developer tooling, and performance tuning are already established.

A key caveat is that the available information is limited to the headline-level report and does not include direct quotes from Amazon AWS, Nvidia, or procurement documentation. Without more detail, it is not possible to confirm the exact procurement timing, pricing, or the degree to which Nvidia capacity will overlap with or eventually replace custom-chip deployments.

Watching next, investors and industry watchers will likely focus on whether Amazon’s custom AI chips expand beyond targeted workloads, and whether Nvidia data center demand continues to be supported by large cloud additions. Any follow-up disclosure from Amazon on AWS’s chip roadmap, capacity planning, or customer adoption would sharpen the picture of how quickly Nvidia GPU demand could soften or stay resilient.