Amazon and NVIDIA announced on August 26 that they are deepening their collaboration. The former will deploy an additional 2 million NVIDIA GPUs in its data centers over the next two years. These chips, which include Blackwell Ultra, Rubin, and Rubin Ultra, are mainly used for large model training and inference, and are planned to be gradually integrated into AWS data centers in 2027 and 2028.
Self-developed and purchased chips go hand in hand, but NVIDIA still leads
The news came as NVIDIA revealed its quarterly earnings call. Just five months ago, Amazon had committed to deploying over 1 million GPUs, but the unexpectedly fast growth in demand has forced the order to double; the amount was not disclosed, but the potential value is estimated to be in the hundreds of billions of dollars based on the price of high-end GPUs. Notably, Amazon is accelerating the development of its self-made Trainium and Graviton chips to reduce dependency, with the annualized revenue run rate of custom chips already exceeding $25 billion, yet NVIDIA still maintains its leading position through this major order.
The cooperation goes beyond just buying cards. NVIDIA's networking, open models, CPUs, data software, and robotics platforms will be more widely integrated into AWS. Vera CPUs are partially integrated with Rubin and partially delivered independently, and have been progressing since the third quarter. Both sides are extending towards warehouse automation, with Amazon adopting "physical AI" technology stacks such as Omniverse, Cosmos, and Isaac to drive a fleet of robots.
NVIDIA's same-day quarterly report confirms the fervor for computing power: second-quarter revenue reached $96.2 billion, exceeding expectations, with $89 billion from the data center segment, up 117% year-over-year. It is expected to reach $108 billion in the third quarter, with Rubin already in mass production this quarter. The company also promised to invest $279 billion to secure supply and capacity. Huang Renxun said that AI has generated profitable "tokens," but the market is still focused on the core question—whether the tens of billions of infrastructure investments can be converted into profit as expected.
