{"count":3,"providers":[{"id":"aws","label":"AWS","instance_group_label":"EC2 instance family","as_of":"2026-08","note":"First cloud in the 'where this hardware actually shows up' series. EC2 instance family → the silicon inside it (NVIDIA and AWS's own Trainium/Inferentia), plus the software integration points and what else on AWS is neither.","instances":[{"family":"G6","accelerator":"L4","product_id":"dc-gpu-ada","gpu_count":"up to 8","status":"shipping","note":"Cost-efficient inference and graphics workloads.","url":"https://aws.amazon.com/ec2/instance-types/g6/"},{"family":"G6e","accelerator":"L40S","product_id":"dc-gpu-ada","gpu_count":"up to 8","status":"shipping","note":"Higher-memory Ada tier for inference and generative AI.","url":"https://aws.amazon.com/ec2/instance-types/g6e/"},{"family":"G7","accelerator":"RTX PRO 4500 Blackwell Server Edition","product_id":"proviz-rtx-pro","gpu_count":"up to 8","status":"shipping (GA Jun 2026)","note":"A distinct Pro Viz SKU from the RTX PRO 6000 tracked as the base product.","url":"https://aws.amazon.com/ec2/instance-types/g7/"},{"family":"G7e","accelerator":"RTX PRO 6000 Blackwell Server Edition","product_id":"proviz-rtx-pro","gpu_count":"up to 8","status":"shipping (GA Jan 2026)","note":"Pro-viz/agentic-AI GPU brought to cloud instances.","url":"https://aws.amazon.com/ec2/instance-types/g7e/"},{"family":"G5","accelerator":"A10G","product_id":"dc-gpu-ampere-legacy","gpu_count":"up to 8","status":"legacy, still offered","note":null,"url":"https://aws.amazon.com/ec2/instance-types/g5/"},{"family":"P5","accelerator":"H100","product_id":"dc-gpu-hopper","gpu_count":"up to 8","status":"legacy, still offered","note":null,"url":"https://aws.amazon.com/ec2/instance-types/p5/"},{"family":"P5e / P5en","accelerator":"H200","product_id":"dc-gpu-hopper","gpu_count":"up to 8","status":"shipping","note":"P5en pairs H200 with Sapphire Rapids CPUs for ~4x CPU-GPU bandwidth vs P5/P5e.","url":"https://aws.amazon.com/ec2/instance-types/p5/"},{"family":"P6-B200","accelerator":"Blackwell B200","product_id":"dc-gpu-blackwell","gpu_count":"8","status":"shipping","note":"Up to 2x P5en performance; EFAv4 networking, not NVLink, ties instances together.","url":"https://aws.amazon.com/ec2/instance-types/p6/"},{"family":"P6-B300","accelerator":"Blackwell Ultra B300","product_id":"dc-gpu-blackwell","gpu_count":"8","status":"shipping (GA Nov 2025)","note":null,"url":"https://aws.amazon.com/ec2/instance-types/p6/"},{"family":"P6e-GB200 UltraServer","accelerator":"GB200 NVL72 (Grace + Blackwell, NVLink-domain rack)","product_id":"dc-sys-nvl72","gpu_count":"rack-scale (NVL72 domain)","status":"shipping (GA Jul 2025)","note":"Also implicates dc-gpu-blackwell and dc-cpu-grace.","url":"https://aws.amazon.com/ec2/instance-types/p6/"},{"family":"P6e-GB300 UltraServer","accelerator":"GB300 NVL72 (Grace + Blackwell Ultra, NVLink-domain rack)","product_id":"dc-sys-nvl72","gpu_count":"rack-scale (NVL72 domain)","status":"shipping (GA Dec 2025)","note":"Also implicates dc-gpu-blackwell and dc-cpu-grace.","url":"https://aws.amazon.com/ec2/instance-types/p6/"},{"family":"Trn2","accelerator":"Trainium2","product_id":"amazon-trainium2","gpu_count":"16 (up to 64 in UltraServers)","status":"shipping","note":"AWS's own custom training/inference accelerator — not NVIDIA silicon.","url":"https://aws.amazon.com/ec2/instance-types/trn2/"},{"family":"Inf2","accelerator":"Inferentia2","product_id":"amazon-inferentia2","gpu_count":"up to 12","status":"shipping","note":"AWS's own cost-efficient, inference-only accelerator — not NVIDIA silicon.","url":"https://aws.amazon.com/machine-learning/inferentia/"}],"platform_integrations":[{"name":"DGX Cloud on AWS","product_id":"sw-dgx-cloud","description":"NVIDIA's managed AI infrastructure service, sold and run on AWS capacity.","url":"https://www.nvidia.com/en-us/data-center/dgx-cloud/"},{"name":"NIM on Bedrock / SageMaker / EKS / Batch","product_id":"sw-nim-nemo","description":"NIM microservices (including NVIDIA Nemotron open models and Cosmos world models) deployable through multiple AWS services.","url":"https://developer.nvidia.com/nim"},{"name":"Spectrum-X in AWS AI Factories","product_id":"dc-net-spectrum","description":"AWS's new sovereign-AI infrastructure offering combines AWS services with NVIDIA Blackwell GPUs and Spectrum-X networking.","url":"https://www.nvidia.com/en-us/networking/spectrumx/"},{"name":"NVLink Fusion + AWS Trainium4","product_id":"dc-net-nvlink","description":"AWS's own next-gen custom silicon (Trainium4) is being designed to interconnect via NVIDIA NVLink and the MGX rack architecture — NVIDIA's interconnect reaching beyond NVIDIA's own chips.","url":"https://www.nvidia.com/en-us/data-center/nvlink/"},{"name":"cuVS in Amazon OpenSearch","product_id":"sw-cuda","description":"NVIDIA's GPU-accelerated vector search library now powers OpenSearch Serverless vector search.","url":"https://developer.nvidia.com/cuvs"}],"not_nvidia":[{"name":"AWS Nitro System","description":"The host/hypervisor offload layer on every modern EC2 GPU instance is AWS's own Nitro System, not NVIDIA BlueField — a common point of confusion.","url":"https://aws.amazon.com/ec2/nitro/"}]},{"id":"gcp","label":"Google Cloud","instance_group_label":"Compute Engine machine family","as_of":"2026-08","note":"Second cloud in the 'where this hardware actually shows up' series — same shape as aws.json. Machine family → the silicon inside it (NVIDIA and Google's own TPU), plus the software integration points and what else on Google Cloud is neither. Verify current GA/allocation status before relying on it — the A3/A4 family is expanding quickly and access is often reservation-gated.","instances":[{"family":"G2","accelerator":"L4","product_id":"dc-gpu-ada","gpu_count":"up to 8","status":"shipping","note":"Cost-efficient inference and graphics workloads — Google Cloud's counterpart to AWS G6.","url":"https://cloud.google.com/compute/docs/gpus#l4-gpus"},{"family":"G4","accelerator":"RTX PRO 6000 Blackwell Server Edition","product_id":"proviz-rtx-pro","gpu_count":"up to 8","status":"shipping","note":"Pro-viz/agentic-AI GPU brought to cloud instances, mirroring AWS G7e.","url":"https://cloud.google.com/compute/docs/gpus#rtx-pro-6000-gpus"},{"family":"A2","accelerator":"A100","product_id":"dc-gpu-ampere-legacy","gpu_count":"up to 16 (A2 Ultra)","status":"legacy, still offered","note":"A2 Standard (40GB) and A2 Ultra (80GB) variants.","url":"https://cloud.google.com/blog/products/compute/announcing-google-cloud-a2-vm-family-based-on-nvidia-a100-gpu"},{"family":"A3 High","accelerator":"H100 SXM","product_id":"dc-gpu-hopper","gpu_count":8,"status":"shipping","note":"Baseline A3 tier.","url":"https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines"},{"family":"A3 Mega","accelerator":"H100 SXM","product_id":"dc-gpu-hopper","gpu_count":8,"status":"shipping","note":"Same H100 SXM as A3 High, with roughly double the GPU-to-GPU networking bandwidth for large training clusters.","url":"https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines"},{"family":"A3 Ultra","accelerator":"H200 SXM","product_id":"dc-gpu-hopper","gpu_count":8,"status":"shipping","note":"Pairs H200 with GCP's Titanium networking offload and RDMA over Converged Ethernet (RoCE).","url":"https://docs.cloud.google.com/compute/docs/gpus/create-gpu-vm-a3u-a4"},{"family":"A4","accelerator":"B200","product_id":"dc-gpu-blackwell","gpu_count":8,"status":"shipping","note":"Often reservation-gated given demand; check quota/allocation before planning around it.","url":"https://cloud.google.com/blog/products/compute/introducing-a4-vms-powered-by-nvidia-blackwell-gpus"},{"family":"A4X","accelerator":"GB200 NVL72 (Grace + Blackwell, NVLink-domain rack)","product_id":"dc-sys-nvl72","gpu_count":"rack-scale (NVL72 domain)","status":"limited availability","note":"Also implicates dc-gpu-blackwell and dc-cpu-grace. Built with NVIDIA for large-scale training; capacity reservation required.","url":"https://docs.cloud.google.com/compute/docs/gpus/create-gpu-vm-a3u-a4"},{"family":"A4X Max","accelerator":"GB300 NVL72 (Grace + Blackwell Ultra, NVLink-domain rack)","product_id":"dc-sys-nvl72","gpu_count":"rack-scale (NVL72 domain)","status":"limited availability","note":"Newest rack-scale tier; standard Compute Engine SLA does not apply as of Aug 2026 — verify current terms before committing.","url":"https://docs.cloud.google.com/compute/docs/gpus/create-gpu-vm-a3u-a4"},{"family":"Cloud TPU v5e","accelerator":"TPU v5e","product_id":"google-tpu-v5e","gpu_count":"up to 256 (pod scale)","status":"shipping","note":"Google's own AI accelerator — not NVIDIA silicon.","url":"https://cloud.google.com/tpu/docs/v5e"},{"family":"Cloud TPU v6e (Trillium)","accelerator":"TPU v6e","product_id":"google-tpu-v6e","gpu_count":"pod scale","status":"shipping","note":"Google's own AI accelerator — not NVIDIA silicon.","url":"https://cloud.google.com/tpu/docs/v6e"},{"family":"Cloud TPU v7 (Ironwood)","accelerator":"TPU v7","product_id":"google-tpu-v7","gpu_count":"up to 9,216 (pod scale)","status":"shipping","note":"Google's own AI accelerator — not NVIDIA silicon.","url":"https://docs.cloud.google.com/tpu/docs/tpu7x"}],"platform_integrations":[{"name":"NIM on Vertex AI / GKE","product_id":"sw-nim-nemo","description":"NIM microservices deployable through Vertex AI Model Garden and Google Kubernetes Engine.","url":"https://developer.nvidia.com/nim"},{"name":"DGX Cloud on Google Cloud","product_id":"sw-dgx-cloud","description":"NVIDIA's managed AI infrastructure service, also sold and run on Google Cloud capacity.","url":"https://www.nvidia.com/en-us/data-center/dgx-cloud/"},{"name":"Quantum InfiniBand in AI Hypercomputer clusters","product_id":"dc-net-quantum","description":"Google's AI Hypercomputer / Cluster Director large-scale training clusters use NVIDIA Quantum InfiniBand fabric alongside GCP's own Jupiter/Titanium networking stack.","url":"https://docs.cloud.google.com/ai-hypercomputer/docs/gpu"}],"not_nvidia":[{"name":"Google Titanium","description":"The host/hypervisor offload and networking layer on modern Compute Engine GPU instances is Google's own Titanium architecture, not NVIDIA BlueField — the same distinction as AWS Nitro on EC2.","url":"https://cloud.google.com/titanium"},{"name":"Google Axion","description":"Google's custom Arm-based CPU (C4A and successors) is Google's own answer to general-purpose compute — a different lane from NVIDIA's Grace CPU, which pairs specifically with NVIDIA GPUs.","url":"https://cloud.google.com/products/axion"}]},{"id":"azure","label":"Microsoft Azure","instance_group_label":"Azure VM series","as_of":"2026-08","note":"Third cloud in the 'where this hardware actually shows up' series — same shape as aws.json / gcp.json. VM series → NVIDIA silicon inside it, plus the software integration points and a short list of what on Azure is deliberately NOT NVIDIA. Verify current GA/region/quota status before relying on it.","instances":[{"family":"NVadsA10 v5","accelerator":"A10","product_id":"dc-gpu-ampere-legacy","gpu_count":"fractional or up to 4","status":"shipping","note":"Graphics/light-inference tier; fractional GPU sizes let you rent partial A10 capacity.","url":"https://learn.microsoft.com/en-us/azure/virtual-machines/sizes/gpu-accelerated/nvadsa10v5-series"},{"family":"NC A100 v4","accelerator":"A100 (PCIe, 80GB)","product_id":"dc-gpu-ampere-legacy","gpu_count":"up to 4","status":"legacy, still offered","note":null,"url":"https://learn.microsoft.com/en-us/azure/virtual-machines/sizes/gpu-accelerated/nc-family"},{"family":"ND A100 v4 / NDm A100 v4","accelerator":"A100 (40GB / 80GB SXM)","product_id":"dc-gpu-ampere-legacy","gpu_count":"up to 8","status":"legacy, still offered","note":null,"url":"https://learn.microsoft.com/en-us/azure/virtual-machines/sizes/gpu-accelerated/nd-family"},{"family":"NCads H100 v5 / NCCads H100 v5","accelerator":"H100 NVL","product_id":"dc-gpu-hopper","gpu_count":"1 or 2","status":"shipping","note":"Smaller, fractional-node H100 tier for teams that don't need a full 8-GPU host.","url":"https://learn.microsoft.com/en-us/azure/virtual-machines/sizes/gpu-accelerated/nc-family"},{"family":"ND H100 v5","accelerator":"H100 SXM","product_id":"dc-gpu-hopper","gpu_count":8,"status":"shipping","note":null,"url":"https://learn.microsoft.com/en-us/azure/virtual-machines/sizes/gpu-accelerated/ndh100v5-series"},{"family":"ND H200 v5","accelerator":"H200 SXM","product_id":"dc-gpu-hopper","gpu_count":8,"status":"shipping","note":"Memory-bound inference tier — Azure's counterpart to AWS P5e/P5en.","url":"https://blogs.nvidia.com/blog/microsoft-azure-hopper-gpu-instances/"},{"family":"ND GB200 v6","accelerator":"GB200 NVL72 (Grace + Blackwell, NVLink-domain rack)","product_id":"dc-sys-nvl72","gpu_count":"rack-scale (NVL72 domain)","status":"shipping","note":"Also implicates dc-gpu-blackwell and dc-cpu-grace. GA per Microsoft's Aug 2026 announcement.","url":"https://techcommunity.microsoft.com/blog/azurehighperformancecomputingblog/accelerating-the-intelligence-age-with-azure-ai-infrastructure-and-the-ga-of-nd-/4394575"},{"family":"ND GB300 v6","accelerator":"GB300 NVL72 (Grace + Blackwell Ultra, NVLink-domain rack)","product_id":"dc-sys-nvl72","gpu_count":"rack-scale (NVL72 domain)","status":"ramping — large dedicated clusters (e.g. for OpenAI) confirmed, broad general-purpose GA not yet verified","note":"Also implicates dc-gpu-blackwell and dc-cpu-grace.","url":"https://azure.microsoft.com/en-us/blog/microsoft-azure-delivers-the-first-large-scale-cluster-with-nvidia-gb300-nvl72-for-openai-workloads/"}],"platform_integrations":[{"name":"NIM on Azure AI Foundry / AKS","product_id":"sw-nim-nemo","description":"NIM microservices deployable through Azure AI Foundry and Azure Kubernetes Service.","url":"https://developer.nvidia.com/nim"},{"name":"DGX Cloud on Microsoft Azure","product_id":"sw-dgx-cloud","description":"NVIDIA's managed AI infrastructure service, also sold and run on Azure capacity.","url":"https://www.nvidia.com/en-us/data-center/dgx-cloud/"},{"name":"Quantum InfiniBand in Azure AI infrastructure","product_id":"dc-net-quantum","description":"Azure's large ND GB200/GB300 clusters use NVIDIA Quantum-class InfiniBand fabric to tie racks together.","url":"https://azure.microsoft.com/en-us/blog/microsoft-and-nvidia-accelerate-ai-development-and-performance/"}],"not_nvidia":[{"name":"Azure Boost","description":"The host/hypervisor offload layer on modern Azure VMs is Microsoft's own Azure Boost, not NVIDIA BlueField — the same distinction as AWS Nitro and Google Titanium.","url":"https://learn.microsoft.com/en-us/azure/azure-boost/overview-azure-boost"},{"name":"Microsoft Maia","description":"Microsoft's own AI accelerator silicon (Maia 200, announced Jan 2026, inference-focused) competes directly with NVIDIA GPUs on Azure.","url":"https://blogs.microsoft.com/blog/2026/01/26/maia-200-the-ai-accelerator-built-for-inference/"},{"name":"Microsoft Cobalt","description":"Microsoft's custom Arm-based CPU is its own answer to general-purpose compute — a different lane from NVIDIA's Grace CPU, which pairs specifically with NVIDIA GPUs.","url":"https://azure.microsoft.com/en-us/blog/microsoft-and-nvidia-accelerate-ai-development-and-performance/"}]}]}