Amazon is trying to build an AI chip alternative, but its next big infrastructure move shows just how firmly Nvidia still holds the keys to the computing boom.
Five months after committing to deploy more than 1 million Nvidia GPUs across AWS, Amazon and Nvidia said they would add another 2 million chips to Amazon data centers in 2027 and 2028. The expanded plan takes AWS’s intended Nvidia fleet beyond 3 million GPUs, after demand “exceeded those expectations.”1
The new capacity will include Nvidia’s Blackwell Ultra, Rubin and Rubin Ultra GPUs, aimed at the immense workloads required to train and run AI models. The companies attribute the surge to customers spanning startups, enterprises, AI labs and governments — a broad-based appetite that has made more computing power the immediate priority.
The scale of the order is also a reminder of the tension inside Amazon’s AI strategy. AWS has been developing Trainium chips as an alternative for deep-learning work and has promoted its Graviton processors, seeking both lower dependence on Nvidia and a bigger role in the semiconductor market. Its custom-chip business has reached a $25 billion annualized revenue run rate, according to the companies’ broader account of the expansion.1
Yet this is not merely a purchase of more accelerators. Amazon plans to bring Nvidia’s Vera CPUs to AWS, while Nvidia’s networking, software, open models and robotics technology are set to spread across the cloud provider’s infrastructure. Another report summed up the shift bluntly: Amazon is adding 2 million GPUs on top of the 1 million it had already planned, while the partnership also brings Vera CPUs to AWS.2
Nvidia frames the spending as a response to AI becoming commercially productive. “If we had more compute, we could generate more profitable tokens,” CEO Jensen Huang said, arguing that additional capacity will feed returns for AI services.1 Investors, however, will be left to test whether the industry’s enormous infrastructure bill delivers profits as quickly as its GPU orders are growing.