L2 · ChipsAccelerator⬦ ChokepointUnited StatesNVDA

Nvidia

Dominant AI GPU designer and de facto industry chokepoint

Nvidia designs the GPUs that train and serve the majority of the world's frontier AI models, paired with its CUDA software moat and NVLink/NVSwitch interconnect. Its Blackwell family (B200, GB200 NVL72 rack-scale systems, and the mid-cycle Blackwell Ultra / GB300) ramped hard through 2025, with the next-gen Rubin platform (Vera Rubin, Rubin CPX) slated for 2026. Nvidia is fabless and depends entirely on TSMC for leading-edge manufacturing and CoWoS advanced packaging plus SK Hynix/Micron/Samsung HBM. It is the single largest source of leverage in the AI stack and a genuine supply chokepoint.

Flagship 2025 platform

Blackwell (B200 / GB200 NVL72), Blackwell Ultra GB300

Next-gen 2026

Rubin / Vera Rubin, Rubin CPX

Software moat

CUDA + NVLink/NVSwitch, ~80-90% AI training share

Manufacturing

Fabless; TSMC N4/N3 + CoWoS packaging

How it fits the stack

Nvidia 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.

NvidiaDepends on ↑Feeds ↓Related

Nvidia in the AI stack. Nvidia 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) — 17 entities, 18 relationships

Nvidia —uses→ Arm Holdings
Nvidia —uses→ Arm Neoverse / CSS server IP
Nvidia —used by→ Cadence Design Systems
Nvidia —depends on→ Electricity Grid & Utilities
Nvidia —depends on→ High-Bandwidth Memory (HBM)
Nvidia —uses→ High-Bandwidth Memory (HBM)
Nvidia —uses→ InfiniBand (NVIDIA Quantum)
Amazon Web Services (AWS) —uses→ Nvidia
Arm Holdings —uses→ Nvidia
CoreWeave (financing) —depends on→ Nvidia
CoreWeave Data Centers —uses→ Nvidia
Crusoe Energy —depends on→ Nvidia
Fireworks AI —depends on→ Nvidia
G42 —depends on→ Nvidia
Nvidia —partners with→ AI Infrastructure Partnership
Nvidia —competes with→ AMD
Nvidia —partners with→ Andreessen Horowitz (a16z)
Nvidia —competes with→ AWS Trainium / Inferentia

Depends on ↑ · 18

usesArm Holdings⬦Grace CPU (in GB200) uses Arm Neoverse cores usesArm Neoverse / CSS server IP⬦Grace CPU built on Arm Neoverse cores used byCadence Design Systems⬦Nvidia uses Cadence for analog/custom and emulation depends onElectricity Grid & Utilities⬦Deployment of Nvidia GPU clusters is gated by available datacenter power, not just chip supply depends onHigh-Bandwidth Memory (HBM)⬦Blackwell/Rubin need HBM3E/HBM4 stacks from SK Hynix/Micron/Samsung usesHigh-Bandwidth Memory (HBM)⬦HBM3E/HBM4 is co-packaged with Nvidia GPUs (Hopper/Blackwell/Rubin); the tightest supply chokepoint alongside CoWoS usesInfiniBand (NVIDIA Quantum)⬦Scale-out fabric for training clusters used byInfiniBand (NVIDIA Quantum)⬦Scale-out fabric connecting NVIDIA GPU servers suppliesInnoLight (Eoptolink peer)⬦800G/1.6T optical modules for NVIDIA AI clusters suppliesMicron TechnologyHBM3E qualified into Nvidia platforms; touts power-efficiency edge used byNAND FlashEnterprise SSDs store AI training datasets and checkpoints in GPU data centers usesNVLink / NVSwitch⬦Unifies 72 Blackwell GPUs in NVL72 domain used bySiemens EDA (Mentor)Tessent DFT and Calibre used in accelerator design/sign-off suppliesSK Hynix⬦SK Hynix is lead HBM supplier for Nvidia GPUs used bySynopsys⬦Nvidia designs GPUs with Synopsys synthesis/verification/IP depends onTSMC (Taiwan Semiconductor Manufacturing Company)⬦All cloud GPUs ultimately fabbed by TSMC on N4/N3 with CoWoS packaging suppliesTSMC (Taiwan Semiconductor Manufacturing Company)⬦Fabricates Nvidia Blackwell (N4/4NP) and Rubin (N3) GPUs depends onTSMC CoWoS (Chip-on-Wafer-on-Substrate)⬦Requires TSMC CoWoS advanced packaging; capacity is the key constraint

Feeds ↓ · 23

usesAmazon Web Services (AWS)Buys massive volumes of Nvidia GPUs (H100/H200/GB200) for EC2 GPU instances usesArm Holdings⬦Grace CPU licenses Arm Neoverse cores depends onCoreWeave (financing)Nvidia-only fleet with early GB200 access; Nvidia is equity investor usesCoreWeave Data CentersBuilt around NVIDIA GB200/GB300 systems depends onCrusoe EnergyDeploys Nvidia GB200 GPUs in Stargate campuses depends onFireworks AIServes open-model inference on Nvidia GPUs depends onG425GW UAE cluster to run Nvidia chips pending Commerce approval usesGoogle Cloud (GCP)Offers Nvidia GPU instances alongside TPUs depends onHUMAINfew-hundred-K GB300 GPUs pending US export licenses depends onLambda (Lambda Labs)Pure Nvidia GPU cloud with early-access allocation; Nvidia investor depends onMeta (AI Infrastructure)Among Nvidia's largest GPU customers for Llama training and ranking depends onMicrosoftamong the largest Nvidia GPU buyers for Azure AI capacity depends onMicrosoft AzureAmong Nvidia's largest buyers of GB200/Blackwell for OpenAI + Copilot depends onMistral Compute / Essonne datacenterBuilt on Nvidia hardware depends onNebius GroupBuilds large Nvidia GPU clusters; Nvidia is an investor depends onOracle Cloud Infrastructure (OCI)OCI superclusters built on Nvidia GB200/GB300; largely resells Nvidia silicon depends onScalewayOperates Nvidia H100/GH200 supercomputers (Jeroboam/Nabuchodonosor)depends onSovereign AI programsSovereign buyers are a major new Nvidia demand class usesStargate ProjectNVIDIA GB200/GB300 systems at Abilene depends onStargate ProjectNvidia systems power the multi-GW campuses depends onTogether AIRuns open-model inference/training on Nvidia GPU clusters depends onVultrOffers Nvidia GPU instances globally depends onxAI (capital raises)Colossus scaling toward 1M+ Nvidia GPUs

Context — capital, rivals, policy · · 42

partners withAI Infrastructure PartnershipNvidia is a consortium technology partner competes withAMDInstinct is the leading merchant-GPU challenger to Nvidia partners withAndreessen Horowitz (a16z)'Oxygen' program secures Nvidia GPU capacity for portfolio startups competes withAWS Trainium / InferentiaIn-house silicon reduces AWS Nvidia dependence competes withBroadcom⬦Broadcom-enabled custom ASICs are the main structural threat to merchant GPUs competes withCambricon competes withCerebras SystemsWafer-scale training + fast-inference cloud vs GPUs partners withCiscoIntegrates NVIDIA Ethernet AI fabrics partners withCoreWeave (financing)Nvidia invests in and prioritizes GPU allocation to CoreWeave competes withd-MatrixIn-memory-compute inference vs GPU serving partners withDatacenter debt & structured financeVendor financing / circular investment concern competes withEtchedTransformer-only ASIC targets GPU inference economics competes withGoogle TPULargest non-Nvidia training fleet competes withGroqLPU inference-only alternative to GPU serving partners withInfineon TechnologiesCo-develops 800V-class rack power architecture competes withIntelGaudi positioned as lower-cost Nvidia alternative partners withIntel Foundry2025 Nvidia $5B investment and x86/RTX collaboration partners withLambda (Lambda Labs)Nvidia strategic investor and supply partner competes withMeta MTIAReduces Meta's marginal Nvidia GPU spend competes withMicrosoft MaiaIn-house accelerator to cut Nvidia dependence partners withNebius GroupNvidia equity backer and supply partner competes withPreferred Networks (PFN)MN-Core targets training/inference vs Nvidia competes withSambaNova SystemsEnterprise/sovereign AI systems vs Nvidia partners withSchneider ElectricCo-developed liquid-cooling / power reference designs for AI racks competes withSoftBank Compute / Izanagi & Arm AIOwn accelerator ambitions vs Nvidia competes withTenstorrentOpen, non-HBM, CUDA-free accelerators + IP competes withTesla DojoWas an in-house alternative to Nvidia training GPUs partners withVertivGB200/GB300 power-and-cooling reference designs partners withvLLMOptimizes open serving stacks for its GPUs invests inCoreWeave (financing)equity stake plus capacity backstop in the GPU cloud invests inIntel~$5B stake (2025) + custom x86 CPU with NVLink integration invests inOpenAIup to $100B committed in 2025 tied to gigawatts of Nvidia systems ('circular' financing)invests inxAI (capital raises)participated in xAI raises funding Colossus GPU clusters restrictsChina (industrial policy & retaliation)⬦2025 pressure on firms to avoid H20 + antitrust/security scrutiny restrictsUS Export Controls (BIS)⬦US export controls cap/ban advanced GPU sales to China (H20 etc.)designsCUDA / software moat⬦CUDA/cuDNN/NCCL is the software moat that makes the silicon sticky designsNvidia Data-Center GPU (Blackwell/Rubin)⬦Nvidia designs the Blackwell/Rubin data-center GPU product line designsNvidia NeMo / NIM / DynamoNvidia's training-to-serving software designsNVIDIA TensorRT-LLM / DynamoNvidia's optimized inference library designsNvidia Vera Rubin platformRubin GPU + Vera CPU, 2026 successor to Grace Blackwell designsNVLink / NVSwitch⬦NVIDIA's proprietary scale-up interconnect designsSpectrum-X / Ethernet-for-AISpectrum-X Ethernet is Nvidia's scale-out fabric play

In the news · 7

NVIDIA · 2026-03-16

NVIDIA NemoClaw: Secure, Always-On AI Agents With One Command

Announced at GTC 2026, NemoClaw is NVIDIA's open-source stack that adds privacy and security controls to OpenClaw. It bundles Nemotron models and the new OpenShell runtime into a single-command install for autonomous AI agents.

NVIDIA · 2026-03-16

NVIDIA Vera Rubin Platform: Seven Chips, Five Racks, One AI Supercomputer

The headline hardware announcement at GTC 2026: the Vera Rubin platform combines seven new chips and five rack-scale systems into one coherent AI supercomputer. Jensen claims 40 million times more compute in 10 years since DGX-1.

NVIDIA · 2026-03-16

Dynamo 1.0: NVIDIA's Operating System for AI Factories Enters Production

Dynamo 1.0 is NVIDIA's inference optimization software that delivers up to 7x performance boost on Blackwell GPUs. It rearchitects inference by splitting work between GPUs (prefill/attention) and Groq LPUs (decode/generation).

NVIDIA · 2026-03-16

NVIDIA Launches Nemotron Coalition: Open Frontier Models With Mistral, Cursor, Perplexity

NVIDIA announced the Nemotron Coalition — a first-of-its-kind collaboration with Mistral AI, Cursor, LangChain, Perplexity, and others to jointly develop Nemotron 4, the next-gen open frontier model, on NVIDIA DGX Cloud.

NVIDIA · 2026-03-16

NVIDIA-Powered Robotaxis to Launch With Uber Across 28 Cities by 2028

Jensen Huang declared 'the ChatGPT moment for self-driving cars has arrived.' NVIDIA-powered robotaxis will launch with Uber across 28 cities on four continents by 2028, starting with LA and SF Bay Area in H1 2027.

NVIDIA · 2026-03-16

NVIDIA DLSS 5: Fusing 3D Graphics With Generative AI — 'The GPT Moment for Graphics'

NVIDIA calls DLSS 5 their most significant graphics breakthrough since real-time ray tracing in 2018. It fuses controllable 3D graphics with generative AI to produce photoreal lighting and materials in real-time 16ms frames.

NVIDIA · 2026-03-16

GTC 2026 Keynote: Jensen's Big Numbers — $1T Demand, 40M× Compute, Agentic AI Era

Jensen Huang's 2-hour GTC 2026 keynote laid out paradigm shifts: $1 trillion in AI infrastructure demand through 2027, computing demand up 1 million times in 2 years, 'every SaaS company will become an AaaS company,' and 'tokens per watt' as the new CEO metric.

← Back to the AI Systems Map