AWS plans 1M GPUs
A social post reports AWS plans to deploy over 1 million NVIDIA GPUs across its global cloud by 2027 to scale AI compute. (x.com) The tweet contrasts that centralized GPU roll‑out with decentralized alternatives like Aethir and has driven discussion about where AI capacity will sit. (x.com)
Amazon Web Services plans to deploy more than 1 million NVIDIA graphics processing units across its cloud starting in 2026. (aws.amazon.com) Amazon Web Services and NVIDIA disclosed the expansion at NVIDIA GTC on March 16, 2026, with Amazon saying the rollout will span global cloud regions and include Blackwell and Rubin chip architectures. (aws.amazon.com) Reuters reported on March 19 that deliveries begin in 2026 and run through the end of 2027, citing NVIDIA vice president Ian Buck. Reuters also reported that the agreement includes Spectrum and ConnectX networking gear alongside the processors. (usnews.com) Cloud graphics processing units are rented chips in remote data centers, and they are the hardware that trains large artificial intelligence models and serves their answers after deployment. Amazon is adding them as demand rises for bigger training runs and for inference, the step where a model generates a response for users. (aws.amazon.com; aws.amazon.com) The scale stands out even inside Amazon’s existing buildout. In October 2025, Amazon said Project Rainier, its custom-chip cluster for Anthropic, had come online with nearly 500,000 Trainium2 chips, and that Anthropic was expected to scale to more than 1 million Trainium2 chips by the end of 2025. (aboutamazon.com) That leaves Amazon pursuing both tracks at once: NVIDIA systems for a broad set of customer workloads and Amazon-designed silicon for selected training and inference jobs. Amazon markets Inferentia for lower-cost inference on Amazon Elastic Compute Cloud, while Trainium2 powers the Rainier cluster. (aws.amazon.com; aboutamazon.com) The NVIDIA relationship is not new. In March 2024, the companies said Amazon Web Services would offer Grace Blackwell-based Amazon Elastic Compute Cloud instances and NVIDIA DGX Cloud services on Amazon infrastructure. (nvidianews.nvidia.com) The new 1 million-unit target has also fed a separate argument about where artificial intelligence capacity will sit. Aethir, which pitches a distributed graphics processing unit cloud, says it has 434,790 enterprise-grade GPUs on demand, while also presenting itself as a lower-cost alternative to Amazon Web Services. (aethir.com; aethir.com) Amazon’s spending gives that debate more weight. Amazon told investors in February 2025 that it expected about $100 billion in capital expenditures in 2025, with Chief Executive Andy Jassy saying most of it would support artificial intelligence for Amazon Web Services. (cnbc.com) For now, the clearest signal is that Amazon is expanding artificial intelligence capacity with both its own chips and NVIDIA’s, and doing it at a scale measured in the millions. (aws.amazon.com; aboutamazon.com))
