The NVIDIA H100 is a data-center GPU based on the Hopper architecture, launched in 2023. The H100 introduced fourth- generation Tensor Cores with FP8 support (halving memory requirements for many workloads), the Transformer Engine for accelerated transformer model inference, and 80GB of HBM3 memory with 3 TB/s bandwidth.
H100 availability was the binding constraint on AI training capacity through 2023-2024. The shortage shaped the competitive dynamics of the AI industry: labs with secured H100 allocations could train frontier models; labs without them could not. Anthropic, OpenAI, Meta, and Google (for internal training) competed aggressively for H100 capacity from NVIDIA, Microsoft Azure, AWS, and Oracle.
As of 2026, H100s remain widely deployed but are being supplemented by H200s (for workloads requiring more memory) and Blackwell- architecture B200s (for the latest frontier training runs). The H100 is likely to remain in production for another three to five years, moving progressively from training workloads toward inference as newer silicon displaces it at the capability frontier.