AWS CEO: A100s still selling
AWS CEO Matt Garman publicly noted that NVIDIA A100 GPU servers — about six years old — are fully sold out at AWS and haven’t been retired, a comment framed to counter worries about rapid chip depreciation. That detail signals AWS customers may still get long service life from older GPU hardware, which affects procurement timing and cost modeling for cloud AI workloads. It’s a useful heads‑up if you’re planning GPU capacity for experiments or production deployments. (x.com) (x.com)
Matt Garman made the remark during a live conversation with Cisco executive Jeetu Patel at the Cisco AI Summit; the exchange was reported by Data Center Dynamics on February 5, 2026. (datacenterdynamics.com) AWS cut list prices for A100-based on‑demand instances by about 33% in a broader June 1, 2025 pricing update, so keeping older A100 servers active sits alongside a clear effort to lower per‑hour compute costs for customers. (aws.amazon.com) The A100 is Nvidia’s data‑center AI chip first introduced in 2020; that chip family was designed to accelerate large numerical models and other compute-heavy workloads, and its continued use matters because some scientific and engineering calculations require higher numeric precision — in other words, the chip must track many decimal places in every calculation for correct results. (datacenterdynamics.com) Garman told the summit that many customers keep running older A100 hardware because overall demand still outstrips supply and because some workloads can’t move to the newest architectures without losing required precision; he contrasted that with other customers who can switch to newer chips by reducing “floating point” accuracy (using fewer decimal places to speed up AI workloads). (datacenterdynamics.com) Nvidia formally placed the A100 family into an end‑of‑life process in early 2024, which affects new orders and factory shipments, yet major cloud operators have continued to operate and rent A100 capacity while they roll out newer generations. (onestopsystems.com) AWS is simultaneously expanding its fleet — the company told industry press it plans to deploy more than one million Nvidia GPUs, including the Blackwell and Rubin architectures, over the following 12 months. (datacenterdynamics.com) For concrete planning: the A100 appears in AWS instance families labeled P4d and P4de, and those on‑demand prices were cut by about 33% in June 2025 while H100/H200 instance families saw larger cuts (up to ~44% for some H100 classes), which changes the math for whether to rent older A100 capacity or wait for newer machines. (aws.amazon.com)
