NVIDIA H100

NVIDIA H100 is a data-centre graphics processing unit introduced by NVIDIA in March 2022 as the first chip built on its Hopper architecture, named after computing pioneer Grace Hopper. Designed to train and run large artificial intelligence models, the H100 became the standard unit of AI compute during the generative AI boom of the 2020s — referenced in corporate results, export-control rules and national technology strategies alike — and powered the first wave of AI supercomputers in Malaysia.^[1][2][5]

Background

The H100 descends from a decades-long shift in which graphics processors, originally built for rendering images, became the engines of machine learning. NVIDIA's CUDA platform, introduced in 2006, allowed researchers to use GPUs for general computation, and the 2012 AlexNet breakthrough in image recognition demonstrated deep learning at scale. Successive data-centre generations — the V100 in 2017 and the A100 in 2020 — industrialised the approach, and demand for the H100 quickly outstripped supply after its launch.^[1]

The chip combines 80 gigabytes of high-bandwidth HBM3 memory with roughly 3.3 terabytes per second of memory bandwidth, and includes a "Transformer Engine" that manages reduced-precision arithmetic for the transformer architectures behind modern language models. NVIDIA cited performance of up to around 30 times the previous generation for large-language-model workloads. H100s are sold in eight-chip DGX servers and larger SuperPOD clusters, and rented by the hour from cloud providers — chips, servers and clusters together forming much of the physical infrastructure of the AI industry.^[1][2]

The AI boom and export controls

Through 2023 and 2024, technology companies built training clusters containing tens of thousands of H100s, and the chip's scarcity shaped strategy across the industry. The H100 also became a focus of United States export controls aimed at limiting China's access to advanced computing: October 2022 rules restricted A100 and H100-class exports, prompting NVIDIA to create reduced-bandwidth models for China, and October 2023 rules closed those loopholes. In January 2025 the outgoing US administration issued the "AI Diffusion" rule, which would have imposed country-level caps on advanced chip exports — affecting Malaysia and other data-centre hubs — before it was rescinded in May 2025 and replaced with case-by-case guidance and continuing concern about diversion to third countries.^[3][4]

>See Also

• NVIDIA
• NVIDIA Blackwell
• GPU Cluster
• AI Chip Export Controls
• YTL AI Cloud