{"data":[{"slug":"amd-epyc-turin","etype":"chip","name":"AMD EPYC 9005 (Turin)","cat":"CPU","subtitle":"The 192-core x86 host CPU under a large share of the world's GPU nodes.","status":"Shipping","region":null,"ticker":null,"kf":{"node":"TSMC N4/N3","year":2024,"class":"CPU","designer":"AMD"},"asof":"2026-01-15"},{"slug":"amd-mi250x","etype":"chip","name":"AMD Instinct MI250X","cat":"GPU","subtitle":"The chiplet GPU that powered Frontier, the first exascale supercomputer.","status":"Prior generation","region":null,"ticker":null,"kf":{"node":"TSMC N6","year":2021,"class":"GPU","tdp_w":560,"mem_bw":3.2,"memory":"128 GB HBM2e","designer":"AMD"},"asof":"2026-01-15"},{"slug":"amd-mi300a","etype":"chip","name":"AMD Instinct MI300A","cat":"APU","subtitle":"CPU and GPU on one package with unified HBM — the chip inside El Capitan.","status":"Shipping","region":null,"ticker":null,"kf":{"node":"TSMC N5/N6","year":2023,"class":"APU","tdp_w":760,"mem_bw":5.3,"memory":"128 GB HBM3 (unified)","designer":"AMD"},"asof":"2026-01-15"},{"slug":"amd-mi300x","etype":"chip","name":"AMD Instinct MI300X","cat":"GPU","subtitle":"AMD's serious challenge to Hopper — more memory than H100, adopted by Microsoft and Meta.","status":"Shipping","region":null,"ticker":null,"kf":{"node":"TSMC N5/N6","year":2023,"class":"GPU","tdp_w":750,"mem_bw":5.3,"memory":"192 GB HBM3","designer":"AMD"},"asof":"2026-01-15"},{"slug":"amd-mi325x","etype":"chip","name":"AMD Instinct MI325X","cat":"GPU","subtitle":"MI300X with 256 GB of HBM3E — the memory-capacity flagship of 2025.","status":"Shipping","region":null,"ticker":null,"kf":{"node":"TSMC N5/N6","year":2024,"class":"GPU","tdp_w":1000,"mem_bw":6.0,"memory":"256 GB HBM3E","designer":"AMD"},"asof":"2026-01-15"},{"slug":"amd-mi355x","etype":"chip","name":"AMD Instinct MI355X","cat":"GPU","subtitle":"CDNA4 on 3 nm: 288 GB HBM3E, FP4/FP6 support, and Blackwell-class ambitions.","status":"Shipping","region":null,"ticker":null,"kf":{"node":"TSMC N3P","year":2025,"class":"GPU","tdp_w":1400,"mem_bw":8.0,"memory":"288 GB HBM3E","designer":"AMD"},"asof":"2026-01-15"},{"slug":"amd-mi400","etype":"chip","name":"AMD Instinct MI400 / MI450","cat":"GPU","subtitle":"AMD's 2026 rack-scale answer — and the chip behind the OpenAI 6-gigawatt deal.","status":"Ramping","region":null,"ticker":null,"kf":{"year":2026,"class":"GPU","memory":"432 GB HBM4","designer":"AMD"},"asof":"2026-08-31"},{"slug":"aws-graviton4","etype":"chip","name":"AWS Graviton4","cat":"CPU","subtitle":"The Arm CPU that quietly runs a huge share of AWS — including AI data plumbing.","status":"Shipping","region":null,"ticker":null,"kf":{"year":2024,"class":"CPU","designer":"AWS Annapurna"},"asof":"2026-01-15"},{"slug":"aws-inferentia2","etype":"chip","name":"AWS Inferentia2","cat":"AI Accelerator","subtitle":"Amazon's inference chip — cheap tokens for production serving.","status":"Shipping","region":null,"ticker":null,"kf":{"year":2023,"class":"AI Accelerator","memory":"32 GB HBM","designer":"AWS Annapurna"},"asof":"2026-01-15"},{"slug":"aws-trainium1","etype":"chip","name":"AWS Trainium","cat":"AI Accelerator","subtitle":"Amazon's first training chip — the opening move of its custom-silicon strategy.","status":"Prior generation","region":null,"ticker":null,"kf":{"year":2022,"class":"AI Accelerator","memory":"32 GB HBM","designer":"AWS Annapurna"},"asof":"2026-01-15"},{"slug":"aws-trainium2","etype":"chip","name":"AWS Trainium2","cat":"AI Accelerator","subtitle":"The chip under Project Rainier — Anthropic's half-million-chip training fleet.","status":"Shipping","region":null,"ticker":null,"kf":{"year":2024,"class":"AI Accelerator","mem_bw":2.9,"memory":"96 GB HBM3","designer":"AWS Annapurna"},"asof":"2026-01-15"},{"slug":"aws-trainium3","etype":"chip","name":"AWS Trainium3","cat":"AI Accelerator","subtitle":"The 3-nanometer follow-on, previewed for 2026 UltraServers.","status":"Announced","region":null,"ticker":null,"kf":{"node":"3 nm-class","year":2026,"class":"AI Accelerator","designer":"AWS Annapurna"},"asof":"2026-01-15"},{"slug":"biren-br100","etype":"chip","name":"Biren BR100","cat":"AI Accelerator","subtitle":"The Chinese GPU startup flagship that US sanctions cut off at the knees.","status":"Historical","region":null,"ticker":null,"kf":{"node":"TSMC N7","year":2022,"class":"AI Accelerator","designer":"Biren"},"asof":"2026-01-15"},{"slug":"cambricon-siyuan-590","etype":"chip","name":"Cambricon Siyuan 590","cat":"AI Accelerator","subtitle":"The listed Chinese AI-chip champion riding domestic substitution.","status":"Shipping","region":null,"ticker":null,"kf":{"year":2023,"class":"AI Accelerator","designer":"Cambricon"},"asof":"2026-01-15"},{"slug":"cerebras-wse-3","etype":"chip","name":"Cerebras WSE-3","cat":"Wafer-Scale","subtitle":"A single chip the size of a dinner plate: 4 trillion transistors, 900,000 cores.","status":"Shipping","region":null,"ticker":null,"kf":{"node":"TSMC N5","year":2024,"class":"Wafer-Scale","tdp_w":23000,"mem_bw":21000,"memory":"44 GB on-chip SRAM","designer":"Cerebras"},"asof":"2026-01-15"},{"slug":"etched-sohu","etype":"chip","name":"Etched Sohu","cat":"AI Accelerator","subtitle":"The transformer-only ASIC bet: burn the architecture into silicon.","status":"In development","region":null,"ticker":null,"kf":{"year":2026,"class":"AI Accelerator","memory":"144 GB HBM3E","designer":"Etched"},"asof":"2026-01-15"},{"slug":"furiosa-rngd","etype":"chip","name":"FuriosaAI RNGD","cat":"AI Accelerator","subtitle":"Korea's inference challenger — the startup that turned down Meta's $800M.","status":"Ramping","region":null,"ticker":null,"kf":{"node":"TSMC N5","year":2024,"class":"AI Accelerator","tdp_w":150,"memory":"48 GB HBM3","designer":"FuriosaAI"},"asof":"2026-01-15"},{"slug":"google-axion","etype":"chip","name":"Google Axion","cat":"CPU","subtitle":"Google's Arm server CPU for the general-purpose half of AI workloads.","status":"Shipping","region":null,"ticker":null,"kf":{"year":2024,"class":"CPU","designer":"Google"},"asof":"2026-01-15"},{"slug":"google-tpu-v1-v3","etype":"chip","name":"Google TPU v1–v3","cat":"AI Accelerator","subtitle":"The originals: the 2015 inference ASIC that started the custom-silicon era.","status":"Historical","region":null,"ticker":null,"kf":{"node":"28 nm (v1)","year":2015,"class":"AI Accelerator","designer":"Google"},"asof":"2026-01-15"},{"slug":"google-tpu-v4","etype":"chip","name":"Google TPU v4","cat":"AI Accelerator","subtitle":"The TPU generation that proved optical circuit switching at 4,096-chip scale.","status":"Prior generation","region":null,"ticker":null,"kf":{"node":"TSMC N7","year":2021,"class":"AI Accelerator","tdp_w":192,"mem_bw":1.2,"memory":"32 GB HBM2","designer":"Google"},"asof":"2026-01-15"},{"slug":"google-tpu-v5e","etype":"chip","name":"Google TPU v5e","cat":"AI Accelerator","subtitle":"The efficiency TPU: cost-per-token optimized inference and mid-scale training.","status":"Shipping","region":null,"ticker":null,"kf":{"year":2023,"class":"AI Accelerator","mem_bw":0.82,"memory":"16 GB HBM2","designer":"Google"},"asof":"2026-01-15"},{"slug":"google-tpu-v5p","etype":"chip","name":"Google TPU v5p","cat":"AI Accelerator","subtitle":"The training TPU behind Gemini 1.0 — 8,960-chip pods.","status":"Shipping","region":null,"ticker":null,"kf":{"year":2023,"class":"AI Accelerator","mem_bw":2.77,"memory":"95 GB HBM2e","designer":"Google"},"asof":"2026-01-15"},{"slug":"google-tpu-v6e-trillium","etype":"chip","name":"Google TPU v6e (Trillium)","cat":"AI Accelerator","subtitle":"The sixth generation: ~4.7x v5e compute, used to train Gemini 2.0.","status":"Shipping","region":null,"ticker":null,"kf":{"year":2024,"class":"AI Accelerator","mem_bw":1.6,"memory":"32 GB HBM","designer":"Google"},"asof":"2026-01-15"},{"slug":"google-tpu-v7-ironwood","etype":"chip","name":"Google TPU v7 (Ironwood)","cat":"AI Accelerator","subtitle":"Google's inference-age flagship: 4.6 PFLOPS FP8 and 9,216-chip superpods.","status":"Ramping","region":null,"ticker":null,"kf":{"year":2025,"class":"AI Accelerator","mem_bw":7.37,"memory":"192 GB HBM3E","designer":"Google"},"asof":"2026-01-15"},{"slug":"graphcore-bow","etype":"chip","name":"Graphcore Bow IPU","cat":"AI Accelerator","subtitle":"The wafer-on-wafer pioneer whose company became SoftBank's AI chip arm.","status":"Historical","region":null,"ticker":null,"kf":{"node":"TSMC N7 (WoW)","year":2022,"class":"AI Accelerator","memory":"900 MB SRAM","designer":"Graphcore"},"asof":"2026-01-15"},{"slug":"groq-lpu","etype":"chip","name":"Groq LPU","cat":"AI Accelerator","subtitle":"The deterministic inference engine famous for extreme tokens-per-second.","status":"Shipping","region":null,"ticker":null,"kf":{"node":"GF 14 nm","year":2020,"class":"AI Accelerator","memory":"230 MB SRAM","designer":"Groq"},"asof":"2026-01-15"},{"slug":"huawei-ascend-910c","etype":"chip","name":"Huawei Ascend 910C","cat":"AI Accelerator","subtitle":"China's frontier chip under sanctions: two 910B dies, SMIC 7 nm, no EUV.","status":"Shipping","region":null,"ticker":null,"kf":{"node":"SMIC N+2 (7 nm-class)","year":2025,"class":"AI Accelerator","memory":"128 GB HBM2E","designer":"Huawei (HiSilicon)"},"asof":"2026-01-15"},{"slug":"intel-gaudi2","etype":"chip","name":"Intel Gaudi 2","cat":"AI Accelerator","subtitle":"The Habana chip that put credible MLPerf numbers against the A100.","status":"Prior generation","region":null,"ticker":null,"kf":{"node":"TSMC N7","year":2022,"class":"AI Accelerator","tdp_w":600,"mem_bw":2.45,"memory":"96 GB HBM2E","designer":"Intel (Habana)"},"asof":"2026-01-15"},{"slug":"intel-gaudi3","etype":"chip","name":"Intel Gaudi 3","cat":"AI Accelerator","subtitle":"Intel's Ethernet-native accelerator — strong specs, struggling traction.","status":"Shipping","region":null,"ticker":null,"kf":{"node":"TSMC N5","year":2024,"class":"AI Accelerator","tdp_w":900,"mem_bw":3.7,"memory":"128 GB HBM2E","designer":"Intel (Habana)"},"asof":"2026-01-15"},{"slug":"intel-xeon-6","etype":"chip","name":"Intel Xeon 6","cat":"CPU","subtitle":"Intel's host-CPU line — the x86 incumbent defending the AI head-node.","status":"Shipping","region":null,"ticker":null,"kf":{"node":"Intel 3","year":2024,"class":"CPU","designer":"Intel"},"asof":"2026-01-15"},{"slug":"meta-mtia-v2","etype":"chip","name":"Meta MTIA v2","cat":"AI Accelerator","subtitle":"Meta's recommendation-inference chip — custom silicon for the ads engine.","status":"Shipping","region":null,"ticker":null,"kf":{"node":"TSMC N5","year":2024,"class":"AI Accelerator","tdp_w":90,"designer":"Meta"},"asof":"2026-01-15"},{"slug":"microsoft-cobalt-100","etype":"chip","name":"Microsoft Cobalt 100","cat":"CPU","subtitle":"Azure's 128-core Arm CPU for the cloud around the AI.","status":"Shipping","region":null,"ticker":null,"kf":{"year":2024,"class":"CPU","designer":"Microsoft"},"asof":"2026-01-15"},{"slug":"microsoft-maia-100","etype":"chip","name":"Microsoft Maia 100","cat":"AI Accelerator","subtitle":"Microsoft's first in-house AI accelerator, built with OpenAI workloads in mind.","status":"Shipping","region":null,"ticker":null,"kf":{"node":"TSMC N5","year":2024,"class":"AI Accelerator","tdp_w":700,"mem_bw":1.8,"memory":"64 GB HBM2E","designer":"Microsoft"},"asof":"2026-01-15"},{"slug":"nvidia-a100","etype":"chip","name":"NVIDIA A100","cat":"GPU","subtitle":"The Ampere workhorse behind GPT-3-era AI — still deployed in huge numbers.","status":"Prior generation","region":null,"ticker":null,"kf":{"node":"TSMC N7","year":2020,"class":"GPU","tdp_w":400,"mem_bw":2.0,"memory":"40/80 GB HBM2e","designer":"NVIDIA"},"asof":"2026-01-15"},{"slug":"nvidia-a800","etype":"chip","name":"NVIDIA A800","cat":"GPU","subtitle":"The export-compliant A100 built for China after the October 2022 controls.","status":"Historical","region":null,"ticker":null,"kf":{"node":"TSMC N7","year":2022,"class":"GPU","tdp_w":400,"mem_bw":2.0,"memory":"40/80 GB HBM2e","designer":"NVIDIA"},"asof":"2026-01-15"},{"slug":"nvidia-b100","etype":"chip","name":"NVIDIA B100","cat":"GPU","subtitle":"The air-cooled Blackwell that the market skipped on the way to B200.","status":"Historical","region":null,"ticker":null,"kf":{"node":"TSMC 4NP","year":2024,"class":"GPU","tdp_w":700,"mem_bw":8.0,"memory":"192 GB HBM3E","designer":"NVIDIA"},"asof":"2026-01-15"},{"slug":"nvidia-b200","etype":"chip","name":"NVIDIA B200","cat":"GPU","subtitle":"Blackwell: two reticle-limit dies fused into one GPU — the 2025 workhorse.","status":"Shipping","region":null,"ticker":null,"kf":{"node":"TSMC 4NP","year":2024,"class":"GPU","tdp_w":1000,"mem_bw":8.0,"memory":"192 GB HBM3E","designer":"NVIDIA"},"asof":"2026-01-15"},{"slug":"nvidia-b300","etype":"chip","name":"NVIDIA B300 (Blackwell Ultra)","cat":"GPU","subtitle":"Blackwell Ultra: 288 GB of HBM3E and ~1.5x the FP4 for the reasoning-inference era.","status":"Ramping","region":null,"ticker":null,"kf":{"node":"TSMC 4NP","year":2025,"class":"GPU","tdp_w":1400,"mem_bw":8.0,"memory":"288 GB HBM3E","designer":"NVIDIA"},"asof":"2026-01-15"},{"slug":"nvidia-bluefield-3","etype":"chip","name":"NVIDIA BlueField-3","cat":"DPU","subtitle":"The data-processing unit that offloads networking, storage and security from hosts.","status":"Shipping","region":null,"ticker":null,"kf":{"year":2023,"class":"DPU","designer":"NVIDIA"},"asof":"2026-01-15"},{"slug":"nvidia-gb200","etype":"chip","name":"NVIDIA GB200 Grace Blackwell","cat":"Superchip","subtitle":"One Grace CPU + two B200 GPUs: the building block of the NVL72 rack.","status":"Shipping","region":null,"ticker":null,"kf":{"node":"TSMC 4NP","year":2024,"class":"Superchip","tdp_w":2700,"mem_bw":16.0,"memory":"384 GB HBM3E + 480 GB LPDDR5X","designer":"NVIDIA"},"asof":"2026-01-15"},{"slug":"nvidia-gh200","etype":"chip","name":"NVIDIA GH200 Grace Hopper","cat":"Superchip","subtitle":"A Grace CPU and Hopper GPU fused over NVLink-C2C into one coherent superchip.","status":"Shipping","region":null,"ticker":null,"kf":{"node":"TSMC 4N","year":2023,"class":"Superchip","tdp_w":1000,"mem_bw":4.9,"memory":"96 GB HBM3 or 144 GB HBM3E + 480 GB LPDDR5X","designer":"NVIDIA"},"asof":"2026-01-15"},{"slug":"nvidia-grace","etype":"chip","name":"NVIDIA Grace CPU","cat":"CPU","subtitle":"NVIDIA's 72-core Arm server CPU — the host half of every superchip.","status":"Shipping","region":null,"ticker":null,"kf":{"node":"TSMC 4N","year":2023,"class":"CPU","tdp_w":500,"memory":"480 GB LPDDR5X","designer":"NVIDIA"},"asof":"2026-01-15"},{"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.","status":"Shipping","region":null,"ticker":null,"kf":{"node":"TSMC 4N","year":2022,"class":"GPU","tdp_w":700,"mem_bw":3.35,"memory":"80 GB HBM3","designer":"NVIDIA"},"asof":"2026-01-15"},{"slug":"nvidia-h200","etype":"chip","name":"NVIDIA H200","cat":"GPU","subtitle":"Hopper with 141 GB of HBM3E — the memory upgrade inference was waiting for.","status":"Shipping","region":null,"ticker":null,"kf":{"node":"TSMC 4N","year":2024,"class":"GPU","tdp_w":700,"mem_bw":4.8,"memory":"141 GB HBM3E","designer":"NVIDIA"},"asof":"2026-01-15"},{"slug":"nvidia-h800","etype":"chip","name":"NVIDIA H800","cat":"GPU","subtitle":"The export-compliant H100 for China — and the chip DeepSeek trained V3 on.","status":"Historical","region":null,"ticker":null,"kf":{"node":"TSMC 4N","year":2023,"class":"GPU","tdp_w":700,"mem_bw":3.35,"memory":"80 GB HBM3","designer":"NVIDIA"},"asof":"2026-01-15"},{"slug":"nvidia-l40s","etype":"chip","name":"NVIDIA L40S","cat":"GPU","subtitle":"The Ada-generation utility GPU: inference, fine-tuning and graphics without HBM.","status":"Shipping","region":null,"ticker":null,"kf":{"node":"TSMC 4N","year":2023,"class":"GPU","tdp_w":350,"mem_bw":0.864,"memory":"48 GB GDDR6","designer":"NVIDIA"},"asof":"2026-01-15"},{"slug":"nvidia-rubin","etype":"chip","name":"NVIDIA Rubin","cat":"GPU","subtitle":"The 2026 architecture: HBM4, the Vera CPU, and NVL144 racks.","status":"Ramping","region":null,"ticker":null,"kf":{"node":"TSMC N3 (reported)","year":2026,"class":"GPU","memory":"288 GB HBM4","designer":"NVIDIA"},"asof":"2026-08-31"},{"slug":"nvidia-rubin-cpx","etype":"chip","name":"NVIDIA Rubin CPX","cat":"AI Accelerator","subtitle":"A context-processing accelerator: GDDR7 instead of HBM, built for prefill.","status":"Announced","region":null,"ticker":null,"kf":{"year":2026,"class":"AI Accelerator","memory":"128 GB GDDR7","designer":"NVIDIA"},"asof":"2026-01-15"},{"slug":"nvidia-v100","etype":"chip","name":"NVIDIA V100","cat":"GPU","subtitle":"The 2017 Volta GPU that made tensor cores — and modern AI training — mainstream.","status":"Historical","region":null,"ticker":null,"kf":{"node":"TSMC 12FFN","year":2017,"class":"GPU","tdp_w":300,"mem_bw":0.9,"memory":"16/32 GB HBM2","designer":"NVIDIA"},"asof":"2026-01-15"},{"slug":"nvidia-vera","etype":"chip","name":"NVIDIA Vera CPU","cat":"CPU","subtitle":"Grace's successor: 88 custom Arm cores for the Rubin platform.","status":"Announced","region":null,"ticker":null,"kf":{"year":2026,"class":"CPU","designer":"NVIDIA"},"asof":"2026-01-15"},{"slug":"qualcomm-ai200","etype":"chip","name":"Qualcomm Cloud AI / AI200","cat":"AI Accelerator","subtitle":"Qualcomm's re-entry into the data center: rack-scale inference from 2026.","status":"Announced","region":null,"ticker":null,"kf":{"year":2026,"class":"AI Accelerator","designer":"Qualcomm"},"asof":"2026-01-15"},{"slug":"rebellions-rebel","etype":"chip","name":"Rebellions REBEL","cat":"AI Accelerator","subtitle":"Korea's other AI-chip bet, merged with SK's Sapeon and backed by Samsung fabs.","status":"In development","region":null,"ticker":null,"kf":{"node":"Samsung","year":2025,"class":"AI Accelerator","designer":"Rebellions"},"asof":"2026-01-15"},{"slug":"sambanova-sn40l","etype":"chip","name":"SambaNova SN40L","cat":"AI Accelerator","subtitle":"A three-tier memory design: SRAM, HBM and terabytes of DDR on one accelerator.","status":"Shipping","region":null,"ticker":null,"kf":{"node":"TSMC N5","year":2023,"class":"AI Accelerator","memory":"64 GB HBM + 1.5 TB DDR","designer":"SambaNova"},"asof":"2026-01-15"},{"slug":"tenstorrent-blackhole","etype":"chip","name":"Tenstorrent Blackhole","cat":"AI Accelerator","subtitle":"Jim Keller's open-stack RISC-V accelerator, sold as cards anyone can buy.","status":"Shipping","region":null,"ticker":null,"kf":{"year":2024,"class":"AI Accelerator","memory":"32 GB GDDR6","designer":"Tenstorrent"},"asof":"2026-01-15"},{"slug":"dmatrix-corsair","etype":"chip","name":"d-Matrix Corsair","cat":"AI Accelerator","subtitle":"Digital in-memory compute aimed at fast, cheap small-batch inference.","status":"Ramping","region":null,"ticker":null,"kf":{"year":2024,"class":"AI Accelerator","designer":"d-Matrix"},"asof":"2026-01-15"}],"meta":{"count":55,"served_from":"site cache (no live upstream calls)","as_of":"2026-10-14T05:59:46+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."}}