{"data":{"slug":"nvidia-h100","etype":"chip","name":"NVIDIA H100","cat":"GPU","subtitle":"The GPU of the ChatGPT era — the unit of account for the 2023–25 AI build-out.","summary":"Hopper's H100 became the defining commodity of the AI boom: clusters were sized in H100s, startups raised funding denominated in them, and export policy was written around them. Its Transformer Engine pairs fourth-generation tensor cores with FP8 precision, roughly tripling A100 training throughput per chip.\n\nxAI's Colossus assembled roughly 100,000 of them in 122 days; Meta, Microsoft, OpenAI, Google and every major cloud deployed them in the hundreds of thousands. Blackwell succeeded it in 2025, but H100/H200 fleets still carry a large share of world AI compute.","attrs":[{"g":"Overview","l":"Designer","v":"NVIDIA"},{"g":"Overview","l":"Architecture","v":"Hopper (GH100)"},{"g":"Overview","l":"Announced","v":"March 2022 (GTC); shipping volume from early 2023"},{"g":"Overview","l":"Status","v":"Shipping — succeeded by Blackwell at the frontier"},{"b":"Vendor spec","g":"Manufacturing","l":"Foundry / process","v":"TSMC 4N (custom 5 nm-class)"},{"b":"Vendor spec","g":"Manufacturing","l":"Transistors","v":"80 billion"},{"b":"Vendor spec","g":"Manufacturing","l":"Die size","u":"mm²","v":"814"},{"b":"Vendor spec","g":"Manufacturing","l":"Packaging","v":"CoWoS-S with 5 or 6 HBM stacks"},{"b":"Vendor spec","g":"Memory","l":"Memory","v":"80 GB HBM3 (SXM); 94 GB HBM3 (NVL); 80 GB HBM2e (PCIe)"},{"b":"Vendor spec","g":"Memory","l":"Memory bandwidth","u":"TB/s","v":"3.35 (SXM) / 3.9 (NVL)"},{"b":"Vendor spec","g":"Compute","l":"TF32 tensor (dense)","n":"NVIDIA datasheets often headline throughput with structured sparsity; figures here are dense unless noted.","u":"TFLOPS","v":"495"},{"b":"Vendor spec","g":"Compute","l":"BF16/FP16 tensor (dense)","u":"TFLOPS","v":"990"},{"b":"Vendor spec","g":"Compute","l":"FP8 tensor (dense)","n":"FP8 via the Transformer Engine was Hopper's signature feature","u":"TFLOPS","v":"1,979"},{"b":"Vendor spec","g":"Compute","l":"INT8 tensor (dense)","u":"TOPS","v":"1,979"},{"b":"Vendor spec","g":"Compute","l":"FP64 tensor","u":"TFLOPS","v":"67"},{"b":"MLPerf","g":"AI Performance","l":"MLPerf","v":"Set the reference results for training and inference benchmarks across 2023–24 submissions"},{"b":"Vendor spec","g":"Power","l":"TDP","u":"W","v":"up to 700 (SXM) / 350 (PCIe)"},{"g":"Physical","l":"Form factors","v":"SXM5, PCIe, NVL (paired PCIe)"},{"b":"Vendor spec","g":"Physical","l":"Interconnect","v":"NVLink 4 — 900 GB/s; PCIe 5.0"},{"g":"Price","l":"Launch price","v":"Not publicly disclosed"},{"b":"Reported","g":"Price","l":"Market price","v":"~$25,000–40,000 per GPU at the 2023–24 peak"},{"b":"Reported","g":"Price","l":"Cloud rental","n":"varies with commitment and provider","v":"Roughly $2–7 per GPU-hour across major and specialist clouds during 2025 — falling steadily"}],"kf":{"node":"TSMC 4N","year":2022,"class":"GPU","tdp_w":700,"mem_bw":3.35,"memory":"80 GB HBM3","designer":"NVIDIA"},"aliases":["H100","Hopper","GH100","H100 SXM","H100 NVL"],"ticker":null,"exchange":null,"hq":null,"founded":null,"website":null,"status":"Shipping","region":null,"asof":"2026-01-15","sources":[{"url":"https://www.nvidia.com/en-us/data-center/h100/","label":"NVIDIA H100 page"},{"url":"https://resources.nvidia.com/en-us-tensor-core","label":"Hopper architecture whitepaper"}]},"meta":{"found":true,"count":1,"served_from":"site cache (no live upstream calls)","as_of":"2026-10-14T01:55:51+00:00","disclaimer":"Research and information only — not investment advice. Structured records carry basis labels (vendor spec / reported / estimate); market quotes are delayed/indicative.","attribution":"Records compiled from vendor documentation, official filings and reputable public reporting; market quotes via financial data providers. See /data-sources."}}