The NVIDIA A100 is a data-center GPU based on the Ampere architecture, launched in 2020. The A100 introduced several capabilities that became standard for AI training: third- generation Tensor Cores with BF16 and TF32 support, multi-instance GPU (MIG) partitioning, and third-generation NVLink for high- bandwidth GPU-to-GPU communication.
A100 was the dominant training GPU from 2020 through 2022 and the backbone of the initial generation of large language model training (GPT-3, early Claude and Gemini models). Availability was notoriously constrained through this period, and multi-year supply commitments drove major cloud capital expenditure decisions.
As of 2026, A100s have been superseded at the frontier by H100s and H200s, but remain widely deployed for inference workloads and for training runs that do not require the latest hardware capabilities. The secondary market for A100s is substantial, and the useful operational lifetime of the hardware extends through at least 2028-2029 for workloads where the cost-per-token economics favor older silicon.