Nvidia to AWS: 1M+ GPUs
- Nvidia said Amazon Web Services will buy 1 million Nvidia graphics processors by the end of 2027, with deliveries starting in 2026 as Amazon expands cloud infrastructure for artificial intelligence workloads. - Nvidia executive Ian Buck said the AWS deal also includes Spectrum networking chips and Groq inference chips, with AWS using seven Nvidia chip types because inference is “wickedly hard.” - The agreement extends an AWS-Nvidia tie-up that already brought Blackwell systems to Amazon’s cloud and shows hyperscalers are buying full AI stacks, not standalone GPUs. (aboutamazon.com)
Artificial intelligence models are trained on giant clusters, but serving answers to users is a different job. Nvidia says that is why Amazon Web Services is buying more than 1 million Nvidia GPUs through 2027. (finance.yahoo.com) Nvidia vice president Ian Buck told Reuters the deliveries start in 2026 and run through 2027. Nvidia and Amazon did not disclose the value of the contract. (finance.yahoo.com) (financialexpress.com) A GPU is the main math engine, but large cloud systems also need memory links and network switches to move data between chips. Buck said AWS is also buying Nvidia Spectrum networking chips and Groq inference chips as part of the same package. (finance.yahoo.com) Inference is the step where a trained model produces a reply, image, or action for a live user request. Buck told Reuters that inference is “wickedly hard” and said AWS plans to use Groq plus six other Nvidia chip types to handle those workloads. (finance.yahoo.com) That helps explain why this is not just a bulk chip order. Nvidia is selling Amazon a broader system of compute, networking, and inference silicon that fits inside AWS data centers and cloud services. (finance.yahoo.com) (capacityglobal.com) AWS and Nvidia had already deepened their partnership in March 2024, when they said the Nvidia Blackwell platform would come to Amazon Elastic Compute Cloud and Nvidia DGX Cloud on AWS. That announcement also tied AWS security and networking software into Blackwell-based systems. (aboutamazon.com) (nvidianews.nvidia.com) AWS has since rolled out Blackwell-based instances, including P6-B200 and P6-B300, and says its P6e UltraServers use Nvidia GB200 and GB300 systems for training and inference. Amazon markets those systems for trillion-parameter-scale models and large distributed GPU clusters. (aws.amazon.com 1) (aws.amazon.com 2) (aws.amazon.com 3) Nvidia has framed the AWS timeline against a wider sales push. Reuters reported that Buck linked the 2027 end date to Jensen Huang’s projection of a $1 trillion sales opportunity for Nvidia’s Rubin and Blackwell chip families through 2027. (financialexpress.com) (economictimes.indiatimes.com) The deal leaves AWS using both in-house silicon and Nvidia hardware at the same time. Amazon has separately said its AI Factories combine Nvidia systems, Trainium chips, AWS services, and Amazon networking into dedicated customer infrastructure. (aboutamazon.com) The headline number is 1 million GPUs, but the more revealing detail is the seven-chip inference stack around them. Nvidia is telling customers that cloud AI now runs on coordinated systems, not single chips. (finance.yahoo.com)
