L2 · ChipsAccelerator⬦ ChokepointUnited States

Nvidia Data-Center GPU (Blackwell/Rubin)

The GPU product line that is the default AI training and inference accelerator

Nvidia's data-center GPU line is the product embodiment of the company's dominance: Hopper (H100/H200) gave way to Blackwell (B200, GB200) in 2025 and points to Rubin in 2026. These parts pair large HBM3E/HBM4 stacks with NVLink to scale into rack-level systems like the GB200 NVL72. Every part is fabricated by TSMC on N4/N3-class nodes using CoWoS packaging, making the product line's output physically gated by TSMC packaging capacity.

Generations

Hopper (H100/H200) → Blackwell (B200/GB200) → Rubin

Memory

HBM3E now, HBM4 on Rubin

Scale-up unit

GB200 NVL72 (72 GPUs, NVLink domain)

How it fits the stack

Nvidia Data-Center GPU (Blackwell/Rubin) with what it depends on (above) and what it feeds (below). The figure renders as a crawlable diagram and upgrades to an interactive 3D graph as it scrolls into view.

Nvidia Data-Center GPU (Blackwell/Rubin)Depends on ↑Feeds ↓Related

Nvidia Data-Center GPU (Blackwell/Rubin) in the AI stack. Nvidia Data-Center GPU (Blackwell/Rubin) with its immediate upstream dependencies (top) and downstream dependents (bottom) in the AI value chain. Hover a node in 3D, or read the full relationships below.

Graph data (text) — 18 entities, 18 relationships

Nvidia Data-Center GPU (Blackwell/Rubin) —depends on→ Advanced Packaging Substrates (ABF)
Nvidia Data-Center GPU (Blackwell/Rubin) —used by→ Coolant Distribution Units & cold plates
Nvidia Data-Center GPU (Blackwell/Rubin) —supplies→ HBM4
Nvidia Data-Center GPU (Blackwell/Rubin) —uses→ High-Bandwidth Memory (HBM)
Nvidia Data-Center GPU (Blackwell/Rubin) —uses→ Ibiden
Nvidia Data-Center GPU (Blackwell/Rubin) —supplies→ Infineon Technologies
Nvidia Data-Center GPU (Blackwell/Rubin) —supplies→ KYEC (King Yuan Electronics)
Anthropic —used by→ Nvidia Data-Center GPU (Blackwell/Rubin)
Anthropic —depends on→ Nvidia Data-Center GPU (Blackwell/Rubin)
Black Forest Labs —depends on→ Nvidia Data-Center GPU (Blackwell/Rubin)
Cohere —depends on→ Nvidia Data-Center GPU (Blackwell/Rubin)
CoreWeave (financing) —used by→ Nvidia Data-Center GPU (Blackwell/Rubin)
DeepSeek —depends on→ Nvidia Data-Center GPU (Blackwell/Rubin)
DeepSpeed / Megatron —depends on→ Nvidia Data-Center GPU (Blackwell/Rubin)
Nvidia Data-Center GPU (Blackwell/Rubin) —competes with→ Huawei Ascend (910B/910C)
Nvidia Data-Center GPU (Blackwell/Rubin) —competes with→ Hyperscaler Custom-ASIC Shift
Nvidia Data-Center GPU (Blackwell/Rubin) —designs→ Nvidia
Nvidia Data-Center GPU (Blackwell/Rubin) —invests in→ AI capex super-cycle (hyperscaler spend)

Depends on ↑ · 22

depends onAdvanced Packaging Substrates (ABF)⬦ABF/substrate shortfalls have throttled GPU output independent of wafer/HBM used byCoolant Distribution Units & cold platesLiquid cooling mandatory for GB200/GB300 suppliesHBM4⬦HBM4 capacity is the tightest single input gating 2026 GPU volumes usesHigh-Bandwidth Memory (HBM)⬦Nvidia accelerators depend on HBM stacks usesIbiden⬦large FC-BGA IC substrate is a capacity-limited input to every GPU package suppliesInfineon TechnologiesPower-delivery silicon and VRMs for GPU racks suppliesKYEC (King Yuan Electronics)⬦Final/wafer test for AI accelerators suppliesMurata ManufacturingThousands of MLCCs per AI GPU board for power delivery suppliesNan Ya PCB⬦ABF FC-BGA substrates for accelerators depends onNCCL (NVIDIA Collective Comms)⬦large training runs depend on NCCL collectives across GPUs depends onNCCL / RCCL collectives⬦Distributed training relies on NCCL collectives depends onPCIe / CXLHost attach via PCIe suppliesRetimers & connectivity silicon (Astera, Credo)Connectivity silicon inside GB200 racks suppliesShinko Electric Industries⬦FC-BGA/ABF substrates carry Nvidia AI GPUs manufacturesTSMC (Taiwan Semiconductor Manufacturing Company)⬦Nvidia GPUs fabbed by TSMC on leading-edge nodes manufacturesTSMC Advanced Packaging Fabs (AP7/Chiayi)⬦CoWoS-packaged GPUs constrained by this capacity manufacturesTSMC Arizona (Fab 21)⬦US-fabbed Blackwell wafers partially de-risk Taiwan concentration; packaging still mostly in Taiwan usesTSMC CoWoS (Chip-on-Wafer-on-Substrate)⬦Blackwell/Hopper use TSMC CoWoS advanced packaging usesUCIe (chiplet interconnect)Die-to-die interconnect for multi-chiplet accelerators depends onUnimicron / ABF substrate makers⬦Every accelerator needs an ABF substrate suppliesUnimicron / ABF substrate makers⬦ABF substrates for GPU packages suppliesVanguard (VIS)PMIC/analog around GPU boards

Feeds ↓ · 40

used byAnthropicAnthropic uses Nvidia GPUs alongside TPU/Trainium depends onAnthropicAlso uses Nvidia GPUs alongside Trainium/TPU depends onBlack Forest LabsGPUs to train FLUX image models depends onCohereTrains on Nvidia GPUs; Nvidia is an investor used byCoreWeave (financing)CoreWeave's fleet is Nvidia-GPU-based depends onDeepSeekTrained on export-restricted Nvidia H800/H20 chips depends onDeepSpeed / MegatronDistributed training on GPU clusters depends onElevenLabs (voice AI)real-time voice synthesis inference demand depends onEU AI Factories & GigafactoriesClusters overwhelmingly built on Nvidia GPUs depends onEuropean sovereign AI (Mistral/Aleph Alpha)Sovereign compute still needs Nvidia GPUs depends onFigure / Humanoid RoboticsRobot foundation models trained on Nvidia GPUs suppliesFoxconn (Hon Hai)GPUs shipped to ODMs to build GB200/GB300 NVL72 racks suppliesFoxconn / Hon Hai (AI servers)⬦GPUs/boards flow to Foxconn for NVL72 rack integration sourcesGPU gray market / diversion⬦gray-market diversion of restricted H100/H200/B200 to China depends onGPU-backed SPV / neocloud financingGPUs serve as loan collateral depends onHumanoid / physical AI (Figure, Tesla Optimus)Vision-language-action models run on Nvidia depends onJUPITER Exascale SupercomputerBuilt on ~24,000 Nvidia GH200 superchips depends onMeta AI (FAIR)Meta operates one of the largest Nvidia H100/Blackwell fleets depends onMidjourney / image-video genimage/video gen distinct GPU demand vector depends onMistral AIMistral trains on Nvidia GPUs via partnership usesMLPerf (MLCommons)Benchmarks Nvidia silicon for buyers depends onNeocloud Debt Financing / GPU-backed Credit⬦GPUs are the collateral; depreciation is the risk depends onNscaleGPU fleet for its sovereign cloud depends onNVIDIA TensorRT-LLM / DynamoTensorRT-LLM/Dynamo target NVIDIA GPUs used byOpenAIOpenAI trains/serves models on Nvidia GPUs depends onOpenAIGPT-5/o-series trained and served on Nvidia GPUs suppliesQuanta Computer⬦GB200/GB300 boards to Quanta for rack assembly depends onSakana AITrains/serves on Nvidia despite efficiency thesis depends onSovereign AI programsNational clusters buy NVIDIA GPUs depends onStepFun (Jieyue Xingchen)export-constrained accelerators usesSupermicro / ODM system integratorsIntegrates GPUs into deployable racks depends onSwiss AI Initiative (ETH/EPFL, Apertus)Apertus trained on ~10k GH200 on Alps depends onTesla AI / xAI compute (Colossus)Colossus is one of the largest single GPU superclusters depends onThinking Machines LabUses Nvidia GPU compute for training usesTSMC N3 (3nm)⬦Blackwell-class GPUs fabricated on N3 usesvLLMHigh-throughput serving on Nvidia GPUs depends onvLLMPrimary backend is Nvidia GPUs suppliesWistronHGX/GB200 baseboard assembly suppliesWiwynn⬦NVL72 rack integration by Wiwynn depends onxAI (capital raises)Colossus cluster built on Nvidia H100/H200 GPUs

Context — capital, rivals, policy · · 5

competes withHuawei Ascend (910B/910C)⬦competes withHyperscaler Custom-ASIC ShiftIn-house ASICs reduce NVIDIA dependence invests inAI capex super-cycle (hyperscaler spend)⬦Hyperscaler capex underwrites GPU demand restrictsChina REE / Gallium-Germanium Export Controls⬦upstream leverage over Western AI hardware designsNvidia⬦Nvidia designs the Blackwell/Rubin data-center GPU product line

← Back to the AI Systems Map