{"meta":{"title":"Inference Stack","as_of":"2026-08","notes":"Family-level taxonomy, not exhaustive SKUs. Each product is tagged with multiple facets so the same dataset can be re-grouped by segment, silicon type, or architecture generation. market_share_pct and revenue_usd are placeholders for a later phase. `url` fields point to NVIDIA/AWS product pages as of Aug 2026 — spot-check before relying on them; product pages get reorganized. Data Center hardware products additionally carry specs/availability_status/availability_note/ideal_workload/sources fields (added Aug 2026 via web research; roadmap-stage products are explicitly flagged as such, not presented as shipped specs). As of Sep 2026, coverage expanded beyond NVIDIA to include AMD, Google, and Intel data-center AI accelerator families — each new entry carries a `vendor` field (missing/omitted `vendor` still means NVIDIA, for backward compatibility with the original 33 entries). Amazon (Trainium/Inferentia) added Sep 2026.","official_segments":["Data Center","Gaming","Professional Visualization","Automotive & Robotics","OEM & Other"],"silicon_types":["GPU","CPU","DPU","LPU","TPU","Networking","System","Software/Platform"]},"count":51,"products":[{"id":"dc-gpu-hopper","name":"Hopper (H100 / H200)","segment":"Data Center","silicon_type":"GPU","family":"Data Center GPU","generation":"Hopper","status":"legacy, still shipping","description":"Prior-generation AI/HPC GPU, still widely deployed and sold alongside Blackwell.","market_share_pct":null,"revenue_usd":null,"url":"https://www.nvidia.com/en-us/data-center/h100/","specs":[{"label":"Memory (H100 SXM)","value":"80GB HBM3"},{"label":"Memory (H200 SXM)","value":"141GB HBM3e"},{"label":"Memory bandwidth (H100)","value":"3.35 TB/s"},{"label":"Memory bandwidth (H200)","value":"4.8 TB/s"},{"label":"FP8 compute (sparse)","value":"3,958 TFLOPS"},{"label":"FP16 compute (sparse)","value":"1,979 TFLOPS"},{"label":"TDP","value":"up to 700W (configurable)"},{"label":"NVLink bandwidth","value":"900 GB/s"},{"label":"Process node","value":"TSMC 4N"}],"availability_status":"shipping","availability_note":"Mainstream production. H100 is now prior-generation but still widely deployed; H200 (higher-memory refresh) has shipped since 2024.","ideal_workload":"H100: large-scale LLM training/inference where 80GB is sufficient. H200: memory-bound inference of very large models (large KV-cache) and HPC needing more/faster memory.","sources":[{"label":"NVIDIA H100","url":"https://www.nvidia.com/en-us/data-center/h100/"},{"label":"NVIDIA H200","url":"https://www.nvidia.com/en-us/data-center/h200/"},{"label":"NVIDIA Hopper architecture","url":"https://www.nvidia.com/en-us/data-center/technologies/hopper-architecture/"}],"vendor":"nvidia"},{"id":"dc-gpu-blackwell","name":"Blackwell (B200 / B300 / GB300 Ultra)","segment":"Data Center","silicon_type":"GPU","family":"Data Center GPU","generation":"Blackwell","status":"shipping, in high demand through 2027","description":"Current flagship AI/HPC GPU generation; basis for GB200/GB300 NVL72 systems.","market_share_pct":null,"revenue_usd":null,"url":"https://www.nvidia.com/en-us/data-center/technologies/blackwell-architecture/","specs":[{"label":"Memory (B200)","value":"180GB HBM3e"},{"label":"Memory (B300 / Blackwell Ultra)","value":"288GB HBM3e"},{"label":"Memory bandwidth (B200)","value":"7.7 TB/s"},{"label":"Memory bandwidth (B300)","value":"8 TB/s"},{"label":"FP4 compute (B200)","value":"18 PFLOPS"},{"label":"FP4 compute (B300, dense/sparse)","value":"15 / 20 PFLOPS"},{"label":"TDP (B200)","value":"up to 1,000W"},{"label":"TDP (B300)","value":"up to 1,400W"},{"label":"NVLink 5 bandwidth","value":"1.8 TB/s per GPU (14.4 TB/s per 8-GPU HGX node)"},{"label":"Process node","value":"TSMC 4NP"}],"availability_status":"shipping","availability_note":"B200 in production since 2025; B300 / GB300 Ultra ramping through 2025-2026 as the current flagship generation.","ideal_workload":"Frontier-scale LLM training and inference (trillion-parameter class), reasoning/agentic inference needing large KV-cache and FP4 throughput. Premium, highest-cost tier.","sources":[{"label":"NVIDIA HGX","url":"https://www.nvidia.com/en-us/data-center/hgx/"},{"label":"NVIDIA Blackwell B200 datasheet (via Primeline)","url":"https://www.primeline-solutions.com/media/categories/server/nach-gpu/nvidia-hgx-h200/nvidia-blackwell-b200-datasheet.pdf"},{"label":"TechPowerUp: GB300 Blackwell Ultra","url":"https://www.techpowerup.com/330154/nvidia-gb300-blackwell-ultra-will-feature-288-gb-hbm3e-memory-1400-w-tdp"}],"vendor":"nvidia"},{"id":"dc-gpu-rubin","name":"Vera Rubin (Rubin GPU)","segment":"Data Center","silicon_type":"GPU","family":"Data Center GPU","generation":"Rubin","status":"launching H2 2026","description":"Next-gen GPU paired with the Vera CPU; part of the annual architecture cadence.","market_share_pct":null,"revenue_usd":null,"url":"https://www.nvidia.com/en-us/data-center/technologies/rubin/","specs":[{"label":"Memory (per GPU)","value":"288GB HBM4"},{"label":"Memory bandwidth (per GPU)","value":"~13 TB/s"},{"label":"Rack configuration","value":"NVL144"},{"label":"Compute (rack-level FP4 dense)","value":"~3.6 EFLOPS"},{"label":"Compute (rack-level FP8 training)","value":"~1.2 EFLOPS"},{"label":"NVLink 6 rack bandwidth","value":"260 TB/s"},{"label":"Inter-rack interconnect (CX9)","value":"28.8 TB/s"},{"label":"Paired CPU","value":"Vera (88-core)"}],"availability_status":"roadmap","availability_note":"Announced/roadmap disclosure (NVIDIA GTC keynotes), not a published datasheet. NVIDIA states a 2H 2026 target; no independent confirmation of shipping was found as of Aug 2026 — treat all figures as preliminary.","ideal_workload":"Positioned by NVIDIA as the successor to Blackwell for next-gen frontier training and high-throughput inference at rack scale.","sources":[{"label":"Tom's Hardware: NVIDIA Rubin/Rubin Ultra/Feynman roadmap","url":"https://www.tomshardware.com/pc-components/gpus/nvidia-announces-rubin-gpus-in-2026-rubin-ultra-in-2027-feynam-after"},{"label":"VideoCardz: Vera Rubin NVL72 detailed","url":"https://videocardz.com/newz/nvidia-vera-rubin-nvl72-detailed-72-gpus-36-cpus-260-tb-s-scale-up-bandwidth"}],"vendor":"nvidia"},{"id":"dc-gpu-rubin-ultra","name":"Rubin Ultra","segment":"Data Center","silicon_type":"GPU","family":"Data Center GPU","generation":"Rubin Ultra","status":"roadmap, 2027","description":"Four compute chiplets, 1TB HBM4E memory; powers Kyber NVL144 racks.","market_share_pct":null,"revenue_usd":null,"url":"https://developer.nvidia.com/blog/nvidia-vera-rubin-pod-seven-chips-five-rack-scale-systems-one-ai-supercomputer/","specs":[{"label":"Rack configuration","value":"NVL576 (576 GPU dies, 4 per package)"},{"label":"Memory (per GPU)","value":"HBM4e, ~8 TB/s"},{"label":"Total rack memory","value":"~365 TB"},{"label":"Compute (rack FP4 inference)","value":"~15 EFLOPS"},{"label":"Compute (rack FP8 training)","value":"~5 EFLOPS"},{"label":"NVLink 7 throughput","value":"~1.5 PB/s"},{"label":"Inter-rack interconnect (CX9)","value":"115.2 TB/s"}],"availability_status":"roadmap","availability_note":"Far-out roadmap disclosure only. NVIDIA target 2H 2027; the associated \"Kyber\" rack system has been reported by press (SemiAnalysis-sourced) to have slipped to 2028. Essentially unverified beyond NVIDIA's own roadmap slides.","ideal_workload":"Positioned for the largest-scale rack/cluster training and inference deployments beyond Rubin — the next wave of frontier model scale.","sources":[{"label":"Tom's Hardware: Rubin/Rubin Ultra/Feynman roadmap","url":"https://www.tomshardware.com/pc-components/gpus/nvidia-announces-rubin-gpus-in-2026-rubin-ultra-in-2027-feynam-after"},{"label":"Tom's Hardware: Kyber rack delay to 2028","url":"https://www.tomshardware.com/pc-components/gpus/nvidias-kyber-rack-for-rubin-ultra-slips-to-2028"}],"vendor":"nvidia"},{"id":"dc-gpu-feynman","name":"Feynman","segment":"Data Center","silicon_type":"GPU","family":"Data Center GPU","generation":"Feynman","status":"roadmap, 2028","description":"Future architecture introducing die stacking, custom HBM, and co-packaged optics in NVLink switches.","market_share_pct":null,"revenue_usd":null,"url":"https://www.datacenterdynamics.com/en/news/nvidia-updates-data-center-product-roadmap-following-lpu-launch-at-gtc-2026/","specs":[{"label":"Target timeframe","value":"~2028"},{"label":"Paired CPU","value":"\"Rosa\" (successor to Vera)"},{"label":"Memory","value":"\"Custom HBM\" beyond HBM4e — spec undisclosed"},{"label":"Process node (candidate, unconfirmed)","value":"TSMC A16 (1.6nm-class)"}],"availability_status":"roadmap","availability_note":"Earliest-stage roadmap mention only. No memory capacity, bandwidth, TDP, NVLink, or TFLOPS figures are public. Some reporting suggests the platform design itself may not yet be finalized.","ideal_workload":"Not yet determined — currently a placeholder for \"the generation after Rubin Ultra\" on NVIDIA's public roadmap.","sources":[{"label":"Guru3D: Feynman architecture targets 2028","url":"https://www.guru3d.com/story/nvidia-feynman-architecture-targets-2028-with-custom-hbm-and-rosa-cpu-platform/"},{"label":"TweakTown: Feynman coming 2028","url":"https://www.tweaktown.com/news/110521/nvidia-updates-roadmap-with-new-details-on-its-next-gen-gpu-feynman-coming-in-2028/index.html"}],"vendor":"nvidia"},{"id":"dc-gpu-ada","name":"L4 / L40S (Ada Lovelace)","segment":"Data Center","silicon_type":"GPU","family":"Data Center GPU","generation":"Ada Lovelace","status":"shipping","description":"Inference- and graphics-optimized data center GPUs, one tier below Hopper/Blackwell. This is the workhorse tier for cost-sensitive inference, virtual workstations, and media/graphics workloads — added to this dataset after mapping AWS instance families surfaced how much of the deployed fleet runs on this tier rather than the flagship one. See deployments.aws.","market_share_pct":null,"revenue_usd":null,"url":"https://www.nvidia.com/en-us/data-center/l4/","specs":[{"label":"Memory (L4)","value":"24GB GDDR6"},{"label":"Memory (L40S)","value":"48GB GDDR6 with ECC"},{"label":"Memory bandwidth (L4)","value":"300 GB/s"},{"label":"Memory bandwidth (L40S)","value":"864 GB/s"},{"label":"FP8 compute (L4)","value":"485 TFLOPS"},{"label":"FP8 compute (L40S, dense/sparse)","value":"733 / 1,466 TFLOPS"},{"label":"TDP (L4)","value":"72W"},{"label":"TDP (L40S)","value":"350W"},{"label":"Process node","value":"TSMC 4N (Ada Lovelace)"}],"availability_status":"shipping","availability_note":"Current-generation efficiency/versatility tier. PCIe-only cards — no NVLink.","ideal_workload":"L4: low-power, high-density inference (video/vision/small-mid LLM), edge/cost-sensitive deployments. L40S: fine-tuning, graphics/rendering, and mid-size inference between L4 and full HGX-class hardware.","sources":[{"label":"NVIDIA L4","url":"https://www.nvidia.com/en-us/data-center/l4/"},{"label":"NVIDIA L40S","url":"https://www.nvidia.com/en-us/data-center/l40s/"}],"vendor":"nvidia"},{"id":"dc-gpu-ampere-legacy","name":"A100 / A10G (Ampere)","segment":"Data Center","silicon_type":"GPU","family":"Data Center GPU","generation":"Ampere","status":"legacy, still widely deployed","description":"Two-generations-prior data center/inference GPU family. Long past the shipping frontier but still common in existing cloud fleets and on-prem clusters — e.g. AWS P4 (A100) and G5 (A10G) instances.","market_share_pct":null,"revenue_usd":null,"url":"https://www.nvidia.com/en-us/data-center/a100/","specs":[{"label":"Memory (A100)","value":"80GB HBM2e"},{"label":"Memory bandwidth (A100 SXM)","value":"2,039 GB/s"},{"label":"FP16 compute (A100, dense/sparse)","value":"312 / 624 TFLOPS"},{"label":"TDP (A100 SXM)","value":"400W"},{"label":"NVLink (A100)","value":"600 GB/s"},{"label":"Memory (A10G)","value":"24GB GDDR6"},{"label":"Memory bandwidth (A10G)","value":"~600 GB/s"},{"label":"TDP (A10G)","value":"150W"}],"availability_status":"legacy","availability_note":"Being phased out of new deployments in favor of Hopper/Blackwell, but still widely available/deployed as the lowest-cost cloud tier.","ideal_workload":"A100: cost-sensitive training/fine-tuning of small-to-mid models, legacy inference, price-sensitive multi-GPU clusters. A10G: lightweight inference, graphics/VDI, small-model workloads at low cost.","sources":[{"label":"NVIDIA A100","url":"https://www.nvidia.com/en-us/data-center/a100/"},{"label":"NVIDIA A10","url":"https://www.nvidia.com/en-us/data-center/products/a10-gpu/"},{"label":"AWS A10G datasheet (PDF)","url":"https://d1.awsstatic.com/product-marketing/ec2/NVIDIA_AWS_A10G_DataSheet_FINAL_02_17_2022.pdf"}],"vendor":"nvidia"},{"id":"dc-cpu-grace","name":"Grace CPU (incl. Grace Hopper / Grace Blackwell superchips)","segment":"Data Center","silicon_type":"CPU","family":"Data Center CPU","generation":"Grace","status":"shipping","description":"Arm-based data center CPU; paired with GPUs in GH200/GB200/GB300 superchip modules.","market_share_pct":null,"revenue_usd":null,"url":"https://www.nvidia.com/en-us/data-center/grace-cpu/","specs":[{"label":"Cores (Grace C1, single)","value":"72x Arm Neoverse V2"},{"label":"Cores (Grace Superchip, dual)","value":"144x Arm Neoverse V2"},{"label":"Memory (single)","value":"LPDDR5X, up to 480GB, up to 512 GB/s"},{"label":"Memory (Superchip)","value":"LPDDR5X, up to 960GB, up to 1,024 GB/s"},{"label":"NVLink-C2C","value":"up to 900 GB/s (Grace-Grace or Grace-Hopper)"},{"label":"TDP","value":"140–250W configurable"}],"availability_status":"shipping","availability_note":"Grace, GH200 (Grace Hopper), GB200 and GB300 (Grace Blackwell) superchip combinations are all shipping now — the standard CPU host across current NVL72-class racks.","ideal_workload":"Coherent-memory CPU host feeding GPU-heavy training/inference nodes via high-bandwidth NVLink-C2C, rather than a general-purpose server CPU.","sources":[{"label":"NVIDIA Grace CPU Superchip","url":"https://www.nvidia.com/en-us/data-center/grace-cpu-superchip/"},{"label":"NVIDIA Grace Hopper Superchip","url":"https://www.nvidia.com/en-us/data-center/grace-hopper-superchip/"},{"label":"NVIDIA GB200 NVL72","url":"https://www.nvidia.com/en-us/data-center/gb200-nvl72/"}],"vendor":"nvidia"},{"id":"dc-cpu-vera","name":"Vera CPU","segment":"Data Center","silicon_type":"CPU","family":"Data Center CPU","generation":"Vera","status":"launching H2 2026","description":"Next-gen CPU paired with Rubin GPU in the Vera Rubin platform.","market_share_pct":null,"revenue_usd":null,"url":"https://www.nvidia.com/en-us/data-center/technologies/rubin/","specs":[{"label":"Cores","value":"88x custom \"Olympus\" Arm-compatible cores"},{"label":"Threading","value":"176 logical threads (NVIDIA \"Spatial Multithreading\")"},{"label":"Memory","value":"LPDDR5X via SOCAMM2, up to 1.5TB, ~1.2 TB/s"},{"label":"NVLink-C2C (to Rubin GPU)","value":"up to 1.8 TB/s"},{"label":"TDP (third-party estimate)","value":"~250–450W"}],"availability_status":"roadmap","availability_note":"Officially unveiled May 31 2026. Not yet broadly available — NVIDIA states production shipments begin fall 2026. Early named partners include Anthropic, OpenAI, Oracle Cloud Infrastructure, CoreWeave, and Lambda. Power/TDP figures are third-party estimates, not NVIDIA-published.","ideal_workload":"Purpose-built as the \"CPU for agents\" — host node for agentic AI/reinforcement-learning workloads (code execution, orchestration) and the CPU half of Vera Rubin GPU superchips, replacing Grace in next-gen racks.","sources":[{"label":"NVIDIA: Vera, the CPU for agents","url":"https://nvidianews.nvidia.com/news/nvidia-unveils-vera-the-cpu-for-agents"},{"label":"The Register: Vera CPU / Olympus cores deep dive","url":"https://www.theregister.com/systems/2026/08/01/nvidias-vera-cpu-and-the-olympus-cores-that-power-it-deep-dive/5282056"}],"vendor":"nvidia"},{"id":"dc-dpu-bluefield","name":"BlueField DPU (BlueField-3 / BlueField-4)","segment":"Data Center","silicon_type":"DPU","family":"Data Processing Unit","generation":"BlueField-3 shipping, BlueField-4 2026","status":"shipping / upcoming","description":"Offloads networking, storage, and security processing from the host CPU.","market_share_pct":null,"revenue_usd":null,"url":"https://www.nvidia.com/en-us/networking/products/data-processing-unit/","specs":[{"label":"Cores (BlueField-3)","value":"up to 16x Armv8.2+ A78 \"Hercules\""},{"label":"Memory (BlueField-3)","value":"16GB DDR5, ECC"},{"label":"Networking (BlueField-3)","value":"up to 400 Gb/s"},{"label":"Cores (BlueField-4)","value":"64x Arm (Grace-derived)"},{"label":"Memory (BlueField-4, 3rd-party reported)","value":"128GB LPDDR5"},{"label":"Networking (BlueField-4)","value":"800 Gb/s (integrated ConnectX-9)"}],"availability_status":"shipping","availability_note":"BlueField-3 is shipping and broadly deployed now. BlueField-4 is in early availability in 2026 as part of Vera Rubin platforms — not yet broadly GA; its memory/PCIe/power figures come from third-party reporting, not a confirmed NVIDIA datasheet.","ideal_workload":"Offloads networking, storage (NVMe-oF), and security processing from host CPUs/GPUs. BlueField-4 specifically targets KV-cache acceleration and data movement for disaggregated LLM inference.","sources":[{"label":"NVIDIA BlueField datasheet","url":"https://resources.nvidia.com/en-us-accelerated-networking-resource-library/datasheet-nvidia-bluefield"},{"label":"NVIDIA: BlueField-4 for AI factories","url":"https://blogs.nvidia.com/blog/bluefield-4-ai-factory/"}],"vendor":"nvidia"},{"id":"dc-lpu-groq","name":"Groq 3 LPX","segment":"Data Center","silicon_type":"LPU","family":"Inference Accelerator (non-GPU)","generation":"LPX gen 1 (further generations expected)","status":"launched GTC 2026","description":"SRAM-based, non-GPU inference chip from the ~$20B Groq acqui-hire/licensing deal. NVIDIA's own product branding is 'LPX' (distinct from Groq's original 'LPU' terminology, kept here as the silicon-type label since it's the industry-generic term); targets agentic AI inference in liquid-cooled racks. Signals NVIDIA diversifying beyond GPU silicon for inference workloads.","market_share_pct":null,"revenue_usd":null,"url":"https://www.nvidia.com/en-us/data-center/lpx/","specs":[{"label":"SRAM (per LPU)","value":"500MB, 150 TB/s bandwidth"},{"label":"Scale-up bandwidth (per LPU)","value":"2.5 TB/s"},{"label":"Rack (256 LPUs)","value":"128GB aggregate SRAM, 40 PB/s aggregate bandwidth, 640 TB/s scale-up, 12TB DDR5"},{"label":"Process node (3rd-party, unconfirmed)","value":"Samsung 4nm"}],"availability_status":"roadmap","availability_note":"\"Groq 3 LPX\" was announced at GTC in March 2026, following NVIDIA's ~$20B non-exclusive technology licensing deal with Groq (Dec 2025) and the hire of Groq's founder/leadership. Targeted to launch H2 2026 alongside the Vera Rubin platform — not yet GA. Groq Inc. continues to operate independently (GroqCloud, legacy GroqCard hardware) under new leadership.","ideal_workload":"Ultra-low-latency, high-token-throughput inference — deterministic SRAM-resident execution avoids HBM round-trips, suited to real-time/agentic inference where latency matters more than raw FLOPS-per-dollar.","sources":[{"label":"NVIDIA LPX","url":"https://www.nvidia.com/en-us/data-center/lpx/"},{"label":"DataCenterDynamics: NVIDIA announces Groq 3 LPU","url":"https://www.datacenterdynamics.com/en/news/nvidia-announces-groq-3-lpu-ai-inference-chip-plans-256-lpu-rack/"},{"label":"Groq: Groq/NVIDIA licensing agreement","url":"https://groq.com/newsroom/groq-and-nvidia-enter-non-exclusive-inference-technology-licensing-agreement-to-accelerate-ai-inference-at-global-scale"}],"vendor":"nvidia"},{"id":"dc-net-nvlink","name":"NVLink / NVLink Switch","segment":"Data Center","silicon_type":"Networking","family":"Chip-to-chip / rack-scale interconnect","generation":"NVLink 5 shipping, NVLink 6 with Rubin, NVLink 7 with Rubin Ultra","status":"shipping / roadmap","description":"High-bandwidth interconnect linking GPUs within a server and across a rack.","market_share_pct":null,"revenue_usd":null,"url":"https://www.nvidia.com/en-us/data-center/nvlink/","specs":[{"label":"Bandwidth per GPU (NVLink 4 / Hopper)","value":"900 GB/s"},{"label":"Bandwidth per GPU (NVLink 5 / Blackwell)","value":"1,800 GB/s"},{"label":"Bandwidth per GPU (NVLink 6 / Rubin)","value":"3,600 GB/s"},{"label":"Rack-domain aggregate (NVLink 5, NVL72)","value":"130 TB/s"},{"label":"Rack-domain aggregate (NVLink 6, NVL72)","value":"260 TB/s"},{"label":"Max domain scale","value":"up to 72 GPUs in one all-to-all NVLink domain"}],"availability_status":"shipping","availability_note":"NVLink 5 (Blackwell, GB200/GB300 NVL72) is generally available and shipping in volume. NVLink 6 (Rubin) was declared \"in full production\" at CES 2026, ramping through 2026 in Vera Rubin NVL144 systems.","ideal_workload":"The highest-bandwidth, lowest-latency GPU-to-GPU fabric within a single rack/domain — lets GPUs act as one large memory-coherent accelerator for tensor/expert-parallel training and inference.","sources":[{"label":"NVIDIA NVLink","url":"https://www.nvidia.com/en-us/data-center/nvlink/"},{"label":"VideoCardz: Vera Rubin NVL72 detailed","url":"https://videocardz.com/newz/nvidia-vera-rubin-nvl72-detailed-72-gpus-36-cpus-260-tb-s-scale-up-bandwidth"}],"vendor":"nvidia"},{"id":"dc-net-connectx","name":"ConnectX SuperNIC","segment":"Data Center","silicon_type":"Networking","family":"Network Interface Card","generation":"ConnectX-8 shipping, ConnectX-9 with Rubin","status":"shipping / roadmap","description":"High-performance NIC for GPU-to-GPU and GPU-to-storage networking (Ethernet/InfiniBand).","market_share_pct":null,"revenue_usd":null,"url":"https://www.nvidia.com/en-us/networking/ethernet-adapters/","specs":[{"label":"Bandwidth (ConnectX-8)","value":"up to 800 Gb/s, PCIe Gen6 x16"},{"label":"Bandwidth (ConnectX-9)","value":"up to 1.6 Tb/s per GPU aggregate, single-port 800G Ethernet mode"}],"availability_status":"shipping","availability_note":"ConnectX-8 is shipping (supply-constrained) in GB300 NVL72. ConnectX-9 is announced/design-in for Vera Rubin NVL144 — firmware GA'd Feb 2026 but broad hardware availability and full spec (PCIe generation/lane count) are unconfirmed.","ideal_workload":"The network interface into the scale-out fabric — connects each GPU server to Spectrum-X Ethernet or Quantum InfiniBand beyond the NVLink domain.","sources":[{"label":"ServeTheHome: ConnectX-8 SuperNIC detailed","url":"https://www.servethehome.com/nvidia-connectx-8-supernic-pcie-gen6-800g-nic-detailed/"},{"label":"NVIDIA ConnectX SuperNIC","url":"https://www.nvidia.com/en-us/networking/products/ethernet/supernic"}],"vendor":"nvidia"},{"id":"dc-net-spectrum","name":"Spectrum-X Ethernet Switch","segment":"Data Center","silicon_type":"Networking","family":"Switch","generation":"Spectrum-6 with Rubin, Spectrum-7 roadmap","status":"shipping / roadmap","description":"Ethernet switching fabric purpose-built for AI data center traffic patterns.","market_share_pct":null,"revenue_usd":null,"url":"https://www.nvidia.com/en-us/networking/spectrumx/","specs":[{"label":"Spectrum-4 / SN5000","value":"up to 800 Gb/s/port, 51.2 Tb/s aggregate per switch"},{"label":"Spectrum-6 / SN6810","value":"128x 800G ports, 102.4 Tb/s aggregate"},{"label":"Spectrum-6 / SN6800","value":"512x 800G ports, 409.6 Tb/s aggregate"}],"availability_status":"shipping","availability_note":"Spectrum-4/SN5000 is GA and shipping in volume. Spectrum-6/SN6000 launched at CES 2026 for gigascale AI factories; datasheets exist but a broad-market ship date beyond early-adopter deployment wasn't confirmed.","ideal_workload":"Ethernet-based scale-out AI fabric connecting GPU servers across racks/pods — an alternative to InfiniBand for customers standardizing on Ethernet operations.","sources":[{"label":"NVIDIA Spectrum-X","url":"https://www.nvidia.com/en-us/networking/ethernet-switching/"},{"label":"SDxCentral: NVIDIA debuts Spectrum-6","url":"https://www.sdxcentral.com/news/nvidia-debuts-spectrum-6-switches-to-power-next-gen-ai-networking/"}],"vendor":"nvidia"},{"id":"dc-net-quantum","name":"Quantum InfiniBand Switch","segment":"Data Center","silicon_type":"Networking","family":"Switch","generation":"Quantum-X800","status":"shipping","description":"InfiniBand switching fabric, the alternative/complement to Spectrum-X Ethernet for AI clusters.","market_share_pct":null,"revenue_usd":null,"url":"https://www.nvidia.com/en-us/networking/products/infiniband/","specs":[{"label":"Quantum-2 (NDR)","value":"400 Gb/s/port, 64 ports, 51.2 Tb/s aggregate"},{"label":"Quantum-X800 (XDR)","value":"800 Gb/s/port, up to 144 ports, ~115.2 Tb/s aggregate"}],"availability_status":"shipping","availability_note":"Quantum-X800 has been shipping since 2024/2025 and is the current generation. A next-gen \"Quantum-X1600\" name appears in analyst coverage only — not an NVIDIA-confirmed product.","ideal_workload":"InfiniBand scale-out fabric for the largest distributed training jobs — hardware in-network computing (SHARP), adaptive routing; historically the standard for the largest-scale AI training networks.","sources":[{"label":"NVIDIA Quantum-X800","url":"https://www.nvidia.com/en-us/networking/products/infiniband/quantum-x800/"},{"label":"NVIDIA Quantum-2","url":"https://www.nvidia.com/en-gb/networking/quantum2/"}],"vendor":"nvidia"},{"id":"dc-sys-dgx","name":"DGX Systems","segment":"Data Center","silicon_type":"System","family":"AI Server (single/multi-node)","generation":"DGX B200 / DGX GB200","status":"shipping","description":"NVIDIA's own-branded AI servers — the reference design for enterprise AI infrastructure.","market_share_pct":null,"revenue_usd":null,"url":"https://www.nvidia.com/en-us/data-center/dgx-platform/","specs":[{"label":"DGX B200","value":"8x B200 GPUs, 1.44TB HBM3e, 72 PFLOPS FP8 sparse, ~14.3kW"},{"label":"DGX B300","value":"8x B300 GPUs, 2.1TB HBM3e, 72 PFLOPS FP8 sparse, ~14kW"},{"label":"Cooling","value":"Air-cooled, MGX-standard 10U chassis"}],"availability_status":"shipping","availability_note":"Both DGX B200 and the newer, higher-memory DGX B300 are shipping now and sold in parallel.","ideal_workload":"Single-node enterprise/departmental AI server — turnkey rack-unit deployment, not a multi-rack cluster building block.","sources":[{"label":"NVIDIA DGX B200","url":"https://www.nvidia.com/en-us/data-center/dgx-b200/"},{"label":"NVIDIA DGX B300","url":"https://www.nvidia.com/en-us/data-center/dgx-b300/"}],"vendor":"nvidia"},{"id":"dc-sys-hgx","name":"HGX Baseboard","segment":"Data Center","silicon_type":"System","family":"AI Server building block","generation":"HGX B200","status":"shipping","description":"Reference baseboard design that OEMs (Dell, HPE, Supermicro, etc.) build servers around.","market_share_pct":null,"revenue_usd":null,"url":"https://www.nvidia.com/en-us/data-center/hgx/","specs":[{"label":"HGX B200","value":"8x B200 GPUs, 1.44TB HBM3e, ~72 PFLOPS FP8 sparse"},{"label":"HGX B300","value":"8x B300 GPUs, 2.3TB HBM3e, ~72 PFLOPS FP8"},{"label":"Integration","value":"OEM baseboard (Dell, HPE, Supermicro, Lenovo, ASUS, etc.)"}],"availability_status":"shipping","availability_note":"Both B200 and B300 baseboard tiers are shipping now via OEM partners; cooling varies by OEM (air or direct-liquid).","ideal_workload":"The standard 8-GPU building block OEM vendors integrate into hyperscale/cloud server nodes — not a rack-scale NVLink domain like NVL72.","sources":[{"label":"NVIDIA HGX","url":"https://www.nvidia.com/en-us/data-center/hgx/"}],"vendor":"nvidia"},{"id":"dc-sys-nvl72","name":"GB200 / GB300 NVL72 Rack","segment":"Data Center","silicon_type":"System","family":"Rack-scale system","generation":"Blackwell generation","status":"shipping","description":"Liquid-cooled, rack-scale system linking 72 GPUs via NVLink as one large accelerator.","market_share_pct":null,"revenue_usd":null,"url":"https://www.nvidia.com/en-us/data-center/gb200-nvl72/","specs":[{"label":"GB200 NVL72","value":"72x B200 + 36x Grace, 13.4TB HBM3e, 1,440 PFLOPS FP4 sparse, ~120kW/rack"},{"label":"GB300 NVL72","value":"72x B300 + 36x Grace, ~20–21TB HBM3e, 1,440 PFLOPS FP4 sparse, ~120–140kW/rack"},{"label":"NVLink bandwidth","value":"130 TB/s across the 72-GPU domain"},{"label":"Cooling","value":"Fully liquid-cooled, direct-to-chip with in-rack CDU"}],"availability_status":"shipping","availability_note":"GB200 NVL72 is shipping via OEM partners (HPE, Dell, Supermicro, Lenovo). GB300 NVL72 is \"available now\" per NVIDIA and in active hyperscaler deployment (e.g. reported in Azure ND v6 clusters).","ideal_workload":"Rack-scale building block for hyperscale training/inference — the standard unit multiple racks are networked together to form a SuperPOD.","sources":[{"label":"NVIDIA GB200 NVL72","url":"https://www.nvidia.com/en-us/data-center/gb200-nvl72/"},{"label":"NVIDIA GB300 NVL72","url":"https://www.nvidia.com/en-us/data-center/gb300-nvl72/"}],"vendor":"nvidia"},{"id":"dc-sys-kyber","name":"Kyber NVL144 Rack","segment":"Data Center","silicon_type":"System","family":"Rack-scale system","generation":"Rubin Ultra generation","status":"roadmap, 2027","description":"Next-gen rack-scale system, ~4x the performance of a Blackwell NVL72 rack.","market_share_pct":null,"revenue_usd":null,"url":"https://developer.nvidia.com/blog/nvidia-vera-rubin-pod-seven-chips-five-rack-scale-systems-one-ai-supercomputer/","specs":[{"label":"GPU packages per rack","value":"144 (576 across 4 optically-linked racks, \"Kyber NVL576\")"},{"label":"Memory (per package, demo figure)","value":"up to 1TB HBM4e"},{"label":"Compute (rack FP4 inference, roadmap projection)","value":"~15 EFLOPS (NVL576 config)"},{"label":"Power (directional estimate)","value":"~600kW-class per rack"},{"label":"Cooling","value":"Liquid-cooled, vertical-tray design with a high-layer-count PCB midplane"}],"availability_status":"roadmap","availability_note":"Not shipping. Originally targeted for 2H 2027 as part of the Rubin Ultra platform; press (SemiAnalysis-sourced reporting via CNBC, Tom's Hardware, and others) reports a ~12+ month slip to 2028, attributed to manufacturing difficulty with a 78-layer PCB midplane. NVIDIA has not confirmed or denied the delay. No official NVIDIA datasheet exists for this product — every figure above should be treated as a roadmap/analyst estimate, not a confirmed spec.","ideal_workload":"Positioned as the next rack-scale building block beyond NVL72/GB300 for frontier-scale training and inference — roughly double the GPU density per rack at far higher power density.","sources":[{"label":"Tom's Hardware: Kyber rack delay to 2028","url":"https://www.tomshardware.com/pc-components/gpus/nvidias-kyber-rack-for-rubin-ultra-slips-to-2028"},{"label":"Tom's Hardware: Rubin Ultra 600,000W Kyber racks","url":"https://www.tomshardware.com/pc-components/gpus/nvidia-shows-off-rubin-ultra-with-600-000-watt-kyber-racks-and-infrastructure-coming-in-2027"}],"vendor":"nvidia"},{"id":"sw-cuda","name":"CUDA / cuDNN / TensorRT","segment":"Data Center","silicon_type":"Software/Platform","family":"Core compute software stack","generation":null,"status":"shipping","description":"The programming model and libraries that are arguably NVIDIA's deepest moat — harder to displace than any single chip.","market_share_pct":null,"revenue_usd":null,"url":"https://developer.nvidia.com/cuda-toolkit","vendor":"nvidia"},{"id":"sw-nim-nemo","name":"NIM Microservices / NeMo","segment":"Data Center","silicon_type":"Software/Platform","family":"AI model deployment & customization","generation":null,"status":"shipping","description":"NIM packages pretrained models as deployable microservices; NeMo is the framework for building/customizing them.","market_share_pct":null,"revenue_usd":null,"url":"https://developer.nvidia.com/nim","vendor":"nvidia"},{"id":"sw-ai-enterprise","name":"NVIDIA AI Enterprise","segment":"Data Center","silicon_type":"Software/Platform","family":"Enterprise AI software suite","generation":null,"status":"shipping","description":"Licensed, supported software layer for enterprises running AI in production.","market_share_pct":null,"revenue_usd":null,"url":"https://www.nvidia.com/en-us/data-center/products/ai-enterprise/","vendor":"nvidia"},{"id":"sw-dgx-cloud","name":"DGX Cloud","segment":"Data Center","silicon_type":"Software/Platform","family":"Cloud service","generation":null,"status":"shipping","description":"NVIDIA-managed AI infrastructure-as-a-service, sold directly and through cloud partners (AWS, Google Cloud, etc.).","market_share_pct":null,"revenue_usd":null,"url":"https://www.nvidia.com/en-us/data-center/dgx-cloud/","vendor":"nvidia"},{"id":"sw-omniverse","name":"Omniverse","segment":"Data Center","silicon_type":"Software/Platform","family":"Simulation / digital twin platform","generation":null,"status":"shipping","description":"Platform for 3D simulation and digital twins; also underpins robotics training (Isaac) and creative/pro-viz workflows.","market_share_pct":null,"revenue_usd":null,"url":"https://www.nvidia.com/en-us/omniverse/","vendor":"nvidia"},{"id":"gaming-geforce-rtx","name":"GeForce RTX 50-series","segment":"Gaming","silicon_type":"GPU","family":"Consumer GPU","generation":"Blackwell","status":"shipping, from $299","description":"Consumer gaming GPU lineup built on the Blackwell architecture, spanning entry-level to flagship.","market_share_pct":null,"revenue_usd":null,"url":"https://www.nvidia.com/en-us/geforce/graphics-cards/","vendor":"nvidia"},{"id":"gaming-dlss","name":"DLSS","segment":"Gaming","silicon_type":"Software/Platform","family":"Rendering software","generation":null,"status":"shipping","description":"AI upscaling/frame-generation software tied to GeForce RTX hardware.","market_share_pct":null,"revenue_usd":null,"url":"https://www.nvidia.com/en-us/geforce/technologies/dlss/","vendor":"nvidia"},{"id":"gaming-geforce-now","name":"GeForce NOW","segment":"Gaming","silicon_type":"Software/Platform","family":"Cloud gaming service","generation":null,"status":"shipping","description":"Cloud game-streaming service, NVIDIA's consumer-facing cloud compute product.","market_share_pct":null,"revenue_usd":null,"url":"https://www.nvidia.com/en-us/geforce-now/","vendor":"nvidia"},{"id":"proviz-rtx-pro","name":"RTX PRO 6000 Blackwell (Workstation / Server / Max-Q)","segment":"Professional Visualization","silicon_type":"GPU","family":"Workstation / server GPU","generation":"Blackwell","status":"shipping","description":"Professional-grade GPU line for design, engineering, data science, and agentic AI workloads on desktops and servers.","market_share_pct":null,"revenue_usd":null,"url":"https://www.nvidia.com/en-us/data-center/rtx-pro-6000-blackwell-server-edition/","vendor":"nvidia"},{"id":"auto-drive-thor","name":"DRIVE Thor","segment":"Automotive & Robotics","silicon_type":"GPU","family":"Automotive SoC","generation":"Thor","status":"shipping","description":"System-on-chip for autonomous driving and in-cabin AI, central to the DRIVE platform.","market_share_pct":null,"revenue_usd":null,"url":"https://www.nvidia.com/en-us/self-driving-cars/","vendor":"nvidia"},{"id":"auto-drive-hyperion","name":"DRIVE Hyperion","segment":"Automotive & Robotics","silicon_type":"System","family":"Reference architecture","generation":null,"status":"shipping","description":"Full sensor + compute reference architecture that automakers build autonomous vehicle platforms on.","market_share_pct":null,"revenue_usd":null,"url":"https://www.nvidia.com/en-us/self-driving-cars/","vendor":"nvidia"},{"id":"robotics-jetson","name":"Jetson (Thor / Orin)","segment":"Automotive & Robotics","silicon_type":"GPU","family":"Edge/robotics compute module","generation":"Thor (latest), Orin (prior gen)","status":"shipping","description":"Small-form-factor compute modules for robots, drones, and other edge/embedded AI devices.","market_share_pct":null,"revenue_usd":null,"url":"https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-thor/","vendor":"nvidia"},{"id":"robotics-isaac","name":"Isaac (incl. GR00T foundation models)","segment":"Automotive & Robotics","silicon_type":"Software/Platform","family":"Robotics simulation & foundation models","generation":null,"status":"shipping","description":"Robotics development platform: simulation (built on Omniverse) plus GR00T foundation models for humanoid/general-purpose robots.","market_share_pct":null,"revenue_usd":null,"url":"https://developer.nvidia.com/isaac","vendor":"nvidia"},{"id":"oem-other","name":"OEM & Other (legacy GPU SKUs, chips sold through OEM channel)","segment":"OEM & Other","silicon_type":"GPU","family":"Catch-all","generation":null,"status":"shipping","description":"Smallest reported segment; legacy and miscellaneous silicon sold through OEM partners.","market_share_pct":null,"revenue_usd":null,"url":"https://www.nvidia.com/en-us/data-center/products/","vendor":"nvidia"},{"id":"amd-gpu-mi300","name":"Instinct MI300X / MI325X","segment":"Data Center","silicon_type":"GPU","family":"Data Center GPU","generation":"CDNA 3","status":"shipping","description":"AMD's widely-deployed data-center GPU, built on CDNA 3 with the largest per-chip HBM capacity of its generation.","market_share_pct":null,"revenue_usd":null,"url":"https://www.amd.com/en/products/accelerators/instinct/mi300x.html","vendor":"amd","cloud_providers":["azure"],"specs":[{"label":"Memory (MI300X)","value":"192GB HBM3"},{"label":"Memory (MI325X)","value":"256GB HBM3e"},{"label":"Memory bandwidth (MI300X)","value":"5.3 TB/s"},{"label":"FP8 compute (sparse)","value":"5.2 PFLOPS"},{"label":"FP32 compute","value":"163.4 TFLOPS"},{"label":"Process node","value":"TSMC 5nm/6nm (chiplet)"}],"availability_status":"shipping","availability_note":"MI300X shipping since December 2023; MI325X (memory-upgraded refresh) shipping since Q4 2024. Available on Microsoft Azure (ND MI300X v5).","ideal_workload":"Large-model inference and training where per-GPU memory capacity is the bottleneck — 192GB (vs. 80GB on H100) fits bigger models per GPU with less sharding.","sources":[{"label":"AMD Instinct MI300X","url":"https://www.amd.com/en/products/accelerators/instinct/mi300x.html"},{"label":"AMD Instinct MI325X","url":"https://www.amd.com/en/products/accelerators/instinct/mi325x.html"}]},{"id":"amd-gpu-mi350","name":"Instinct MI350X / MI355X","segment":"Data Center","silicon_type":"GPU","family":"Data Center GPU","generation":"CDNA 4","status":"shipping","description":"AMD's current-generation data-center GPU, roughly doubling dense AI throughput over CDNA 3 with native low-precision (FP4/FP6) formats for inference.","market_share_pct":null,"revenue_usd":null,"url":"https://www.amd.com/en/products/accelerators/instinct/mi350/mi355x.html","vendor":"amd","specs":[{"label":"Memory","value":"288GB HBM3e"},{"label":"Memory bandwidth","value":"8 TB/s"},{"label":"FP8/INT8 compute (sparse, MI355X)","value":"10.1 PFLOPS"},{"label":"FP4/MXFP4 compute (sparse, MI355X)","value":"10.1 PFLOPS"},{"label":"Process node","value":"TSMC 3nm"}],"availability_status":"shipping","availability_note":"Launched June 2025. MI355X is the higher-clocked, liquid-cooled variant of the air-cooled MI350X. Broad hyperscaler availability beyond direct/OEM channels was not yet confirmed as of this writing — spot-check before relying on it.","ideal_workload":"Frontier-scale training and high-throughput inference needing maximum memory bandwidth and native low-precision (FP4/FP6) formats.","sources":[{"label":"AMD Instinct MI355X","url":"https://www.amd.com/en/products/accelerators/instinct/mi350/mi355x.html"},{"label":"AMD Instinct MI350X","url":"https://www.amd.com/en/products/accelerators/instinct/mi350/mi350x.html"}]},{"id":"amd-sw-rocm","name":"ROCm Platform","segment":"Data Center","silicon_type":"Software/Platform","family":"Core compute software stack","generation":null,"status":"shipping","description":"AMD's open-source GPU compute stack — drivers, libraries, and the HIP programming model that let inference frameworks target Instinct GPUs.","market_share_pct":null,"revenue_usd":null,"url":"https://rocm.docs.amd.com","vendor":"amd"},{"id":"google-tpu-v5e","name":"Cloud TPU v5e","segment":"Data Center","silicon_type":"TPU","family":"Cloud TPU","generation":"v5e","status":"shipping","description":"Google's cost-efficient TPU generation for inference and mid-scale training and fine-tuning, available only through Google Cloud.","market_share_pct":null,"revenue_usd":null,"url":"https://cloud.google.com/tpu/docs/v5e","vendor":"google","cloud_providers":["gcp"],"specs":[{"label":"Peak compute (BF16)","value":"197 TFLOPS/chip"},{"label":"HBM per chip","value":"16GB"},{"label":"Pod scale","value":"up to 256 chips"}],"availability_status":"shipping","availability_note":"Generally available on Google Cloud.","ideal_workload":"Cost-efficient LLM inference and fine-tuning at moderate scale.","sources":[{"label":"Google Cloud TPU v5e","url":"https://cloud.google.com/tpu/docs/v5e"}]},{"id":"google-tpu-v6e","name":"Cloud TPU v6e (Trillium)","segment":"Data Center","silicon_type":"TPU","family":"Cloud TPU","generation":"v6e / Trillium","status":"shipping","description":"Google's sixth-generation TPU, roughly doubling peak compute and expanding memory bandwidth over v5e for training and inference.","market_share_pct":null,"revenue_usd":null,"url":"https://cloud.google.com/tpu/docs/v6e","vendor":"google","cloud_providers":["gcp"],"specs":[{"label":"Peak compute (BF16)","value":"~918 TFLOPS/chip"},{"label":"Inter-chip interconnect bandwidth","value":"13 TB/s"},{"label":"Native precision","value":"BF16 (no native FP8)"}],"availability_status":"shipping","availability_note":"Generally available on Google Cloud since 2025.","ideal_workload":"Large-scale training and inference needing more headroom than v5e without moving to Ironwood-class hardware.","sources":[{"label":"Google Cloud TPU v6e","url":"https://cloud.google.com/tpu/docs/v6e"}]},{"id":"google-tpu-v7","name":"Cloud TPU v7 (Ironwood)","segment":"Data Center","silicon_type":"TPU","family":"Cloud TPU","generation":"v7 / Ironwood","status":"shipping","description":"Google's newest TPU generation, built for the inference era — the first TPU with native FP8 hardware support and the largest TPU pods to date.","market_share_pct":null,"revenue_usd":null,"url":"https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/ironwood-tpu-age-of-inference/","vendor":"google","cloud_providers":["gcp"],"specs":[{"label":"Peak compute (FP8)","value":"4,614 TFLOPS/chip"},{"label":"Memory","value":"192GB HBM3e"},{"label":"Memory bandwidth","value":"7.37 TB/s"},{"label":"Inter-chip interconnect bandwidth","value":"9.6 Tb/s"},{"label":"Pod scale","value":"9,216 chips / 42.5 FP8 ExaFLOPS"}],"availability_status":"shipping","availability_note":"Announced at Google Cloud Next 2025; generally available since late 2025.","ideal_workload":"Very large-scale inference serving where pod-level throughput and native FP8 matter most.","sources":[{"label":"Ironwood: the first Google TPU for the age of inference","url":"https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/ironwood-tpu-age-of-inference/"},{"label":"Google Cloud TPU7x (Ironwood) docs","url":"https://docs.cloud.google.com/tpu/docs/tpu7x"}]},{"id":"google-sw-jax-xla","name":"JAX / XLA","segment":"Data Center","silicon_type":"Software/Platform","family":"Core compute software stack","generation":null,"status":"shipping","description":"Google's compiler and framework stack for TPUs — XLA compiles JAX and TensorFlow programs into optimized TPU kernels.","market_share_pct":null,"revenue_usd":null,"url":"https://github.com/openxla/xla","vendor":"google"},{"id":"intel-gaudi3","name":"Gaudi 3","segment":"Data Center","silicon_type":"GPU","family":"Data Center GPU","generation":"Gaudi 3","status":"shipping","description":"Intel's current-generation AI accelerator (from its Habana Labs acquisition) — an open, Ethernet-networked alternative to NVLink-based clusters.","market_share_pct":null,"revenue_usd":null,"url":"https://www.intel.com/content/www/us/en/products/details/discrete-gpus/gaudi.html","vendor":"intel","specs":[{"label":"Memory","value":"128GB HBM2e"},{"label":"On-chip SRAM","value":"96MB"},{"label":"Networking","value":"24x 200GbE integrated"}],"availability_status":"shipping","availability_note":"Available via cloud partners since late 2024, with expanded availability through 2025. Adoption has been slower than Intel initially targeted. Not offered as a named instance family on AWS, Google Cloud, or Azure — available via Intel Tiber Developer Cloud and select partners.","ideal_workload":"Cost-sensitive training and inference for teams wanting an open, non-NVLink networking alternative.","sources":[{"label":"Intel Gaudi AI accelerators","url":"https://www.intel.com/content/www/us/en/products/details/discrete-gpus/gaudi.html"}]},{"id":"intel-crescent-island","name":"Crescent Island","segment":"Data Center","silicon_type":"GPU","family":"Data Center GPU","generation":"Xe3P","status":"roadmap, sampling H2 2026","description":"Intel's upcoming inference-focused data-center GPU, built on the Xe3P architecture with HBM4 memory — Intel's pivot away from the Gaudi line toward a GPU-based roadmap.","market_share_pct":null,"revenue_usd":null,"url":"https://www.tomshardware.com/pc-components/gpus/hot-chips-2026-intel-dives-deep-on-crescent-island-ai-accelerator-larger-caches-and-deeper-xmx-engines-target-maximum-ai-flops-per-watt","vendor":"intel","specs":[{"label":"Architecture","value":"Xe3P"},{"label":"Memory","value":"HBM4 (capacity not yet disclosed)"},{"label":"Cooling","value":"Liquid-cooled"}],"availability_status":"roadmap","availability_note":"Not yet shipping — entering customer sampling in H2 2026 per Intel's Hot Chips 2026 disclosures. A successor, 'Jaguar Shores,' is expected in 2027. Treat all specs here as pre-launch and subject to change.","ideal_workload":"Not yet available for production workloads — inference-focused once shipping.","sources":[{"label":"Tom's Hardware: Hot Chips 2026 Crescent Island deep dive","url":"https://www.tomshardware.com/pc-components/gpus/hot-chips-2026-intel-dives-deep-on-crescent-island-ai-accelerator-larger-caches-and-deeper-xmx-engines-target-maximum-ai-flops-per-watt"}]},{"id":"intel-sw-oneapi","name":"oneAPI / Gaudi Software","segment":"Data Center","silicon_type":"Software/Platform","family":"Core compute software stack","generation":null,"status":"shipping","description":"Intel's unified programming model spanning CPUs, GPUs, and Gaudi accelerators, including the Gaudi software stack used for framework integration.","market_share_pct":null,"revenue_usd":null,"url":"https://www.intel.com/content/www/us/en/developer/tools/oneapi/overview.html","vendor":"intel"},{"id":"amazon-trainium2","name":"Trainium2","segment":"Data Center","silicon_type":"GPU","family":"Data Center GPU","generation":"Trainium2","status":"shipping","description":"AWS's second-generation custom training/inference accelerator, built to reduce reliance on NVIDIA for large-scale model training on EC2.","market_share_pct":null,"revenue_usd":null,"url":"https://aws.amazon.com/ai/machine-learning/trainium/","vendor":"amazon","cloud_providers":["aws"],"specs":[{"label":"Memory","value":"96GB HBM3e"},{"label":"Memory bandwidth (per chip)","value":"~2.9 TB/s"},{"label":"Compute (BF16, dense)","value":"667 TFLOPS/chip"},{"label":"TDP","value":"~500W"},{"label":"UltraServer scale","value":"up to 64 chips / 83.2 PFLOPS FP8"}],"availability_status":"shipping","availability_note":"Generally available since December 2024 via EC2 Trn2 instances and UltraServers. AWS-exclusive — not sold as standalone hardware.","ideal_workload":"Large-scale LLM training and inference on AWS at lower cost than NVIDIA-based instances, for workloads compatible with the Neuron SDK.","sources":[{"label":"AWS Trainium","url":"https://aws.amazon.com/ai/machine-learning/trainium/"},{"label":"Amazon EC2 Trn2 Instances","url":"https://aws.amazon.com/ec2/instance-types/trn2/"}]},{"id":"amazon-trainium3","name":"Trainium3","segment":"Data Center","silicon_type":"GPU","family":"Data Center GPU","generation":"Trainium3","status":"shipping","description":"AWS's newest custom accelerator, with roughly 1.7x the memory bandwidth of Trainium2 and native FP8 support at large UltraServer scale.","market_share_pct":null,"revenue_usd":null,"url":"https://aws.amazon.com/about-aws/whats-new/2025/12/amazon-ec2-trn3-ultraservers/","vendor":"amazon","cloud_providers":["aws"],"specs":[{"label":"Memory","value":"144GB HBM3e"},{"label":"Memory bandwidth","value":"4.9 TB/s"},{"label":"Compute (FP8)","value":"2.52 PFLOPS/chip"},{"label":"UltraServer scale","value":"up to 144 chips / 362 MXFP8 PFLOPS"}],"availability_status":"shipping","availability_note":"Announced at re:Invent 2025; Trn3 UltraServers generally available since December 2025. AWS-exclusive.","ideal_workload":"Frontier-scale training and high-throughput inference on AWS needing maximum per-chip memory and bandwidth.","sources":[{"label":"Announcing Amazon EC2 Trn3 UltraServers","url":"https://aws.amazon.com/about-aws/whats-new/2025/12/amazon-ec2-trn3-ultraservers/"}]},{"id":"amazon-inferentia2","name":"Inferentia2","segment":"Data Center","silicon_type":"GPU","family":"Data Center GPU","generation":"Inferentia2","status":"shipping","description":"AWS's cost-efficient, inference-only accelerator for production LLM and computer-vision serving.","market_share_pct":null,"revenue_usd":null,"url":"https://aws.amazon.com/machine-learning/inferentia/","vendor":"amazon","cloud_providers":["aws"],"specs":[{"label":"Memory","value":"32GB HBM"},{"label":"Compute (FP16)","value":"190 TFLOPS"}],"availability_status":"shipping","availability_note":"Generally available since 2023 via EC2 Inf2 instances. AWS-exclusive.","ideal_workload":"Cost-sensitive inference serving where Trainium's training flexibility isn't needed.","sources":[{"label":"AWS Inferentia","url":"https://aws.amazon.com/machine-learning/inferentia/"}]},{"id":"amazon-sw-neuron","name":"AWS Neuron SDK","segment":"Data Center","silicon_type":"Software/Platform","family":"Core compute software stack","generation":null,"status":"shipping","description":"AWS's SDK and compiler stack for Trainium and Inferentia, integrating with PyTorch and JAX.","market_share_pct":null,"revenue_usd":null,"url":"https://aws.amazon.com/machine-learning/neuron/","vendor":"amazon"},{"id":"amd-dpu-pensando","name":"Pensando Salina DPU / Vulcano SmartNIC","segment":"Data Center","silicon_type":"DPU","family":"Data Center DPU","generation":"Salina (3rd-gen)","status":"shipping","description":"AMD's data-processing-unit and SmartNIC line (from the Pensando acquisition) — offloads networking, storage, and security from the host CPU and extends GPU memory via DPU-managed NVMe.","market_share_pct":null,"revenue_usd":null,"url":"https://www.amd.com/en/blogs/2026/ai-networking-built-for-scale.html","vendor":"amd","specs":[{"label":"Salina 400 DPU networking","value":"Dual 400GE, PCIe Gen 5"},{"label":"Salina 400 DPU compute","value":"232 P4 MPU engines, 16x Arm Neoverse-N1 cores"},{"label":"Salina 400 DPU memory","value":"up to 128GB DDR5"},{"label":"Vulcano 800 SmartNIC","value":"up to 800 Gbps Ethernet per card"}],"availability_status":"shipping","availability_note":"Salina is AMD/Pensando's third-generation DPU, announced 2026 as part of the Helios rack platform; claimed up to 1.4x the performance of NVIDIA BlueField-3.","ideal_workload":"Offloading networking/security/storage from host CPUs in large AI clusters; scale-out and scale-across fabric for multi-rack training.","sources":[{"label":"AMD: AI Networking Built for Scale","url":"https://www.amd.com/en/blogs/2026/ai-networking-built-for-scale.html"}]},{"id":"amd-sys-helios","name":"Helios AI Rack","segment":"Data Center","silicon_type":"System","family":"Rack-scale AI system","generation":null,"status":"launching H2 2026","description":"AMD's fully integrated rack-scale AI system — Instinct GPUs, EPYC CPUs, and Pensando networking co-designed together, AMD's first real answer to NVIDIA's DGX/NVL72 vertical integration.","market_share_pct":null,"revenue_usd":null,"url":"https://www.amd.com/en/blogs/2026/amd-launches-helios-the-highest-performing-rackscale-ai-infrastructure-solution.html","vendor":"amd","specs":[{"label":"GPUs per rack","value":"72x Instinct MI450-series (MI455X)"},{"label":"Memory","value":"31 TB unified HBM4"},{"label":"Compute (FP8 / FP4)","value":"1.4 / 2.9 ExaFLOPS"},{"label":"Networking","value":"All-Ethernet, open-standards (Pensando) vs. NVIDIA's proprietary interconnect"},{"label":"Host CPU","value":"6th Gen AMD EPYC"}],"availability_status":"roadmap","availability_note":"Announced 2026; AMD expects to begin shipping to customers in H2 2026. List price reported around -5.5M per rack.","ideal_workload":"Frontier-scale training and inference for customers wanting an open-standards (Ethernet) alternative to NVIDIA's NVLink-based rack systems.","sources":[{"label":"AMD Launches Helios","url":"https://www.amd.com/en/blogs/2026/amd-launches-helios-the-highest-performing-rackscale-ai-infrastructure-solution.html"}]},{"id":"google-sys-tpu-pod","name":"TPU Superpod (Ironwood)","segment":"Data Center","silicon_type":"System","family":"Cloud TPU","generation":"v7 / Ironwood","status":"shipping","description":"Google's largest TPU deployment unit — thousands of Ironwood chips wired together with proprietary inter-chip interconnect, offered only as a managed Google Cloud service, never sold as standalone hardware.","market_share_pct":null,"revenue_usd":null,"url":"https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/ironwood-tpu-age-of-inference/","vendor":"google","cloud_providers":["gcp"],"specs":[{"label":"Pod scale","value":"9,216 chips"},{"label":"Pod compute (FP8)","value":"42.5 ExaFLOPS"},{"label":"Access model","value":"Google Cloud only — not sold as hardware"}],"availability_status":"shipping","availability_note":"Generally available on Google Cloud since late 2025, as part of the Ironwood (v7) generation.","ideal_workload":"The largest-scale inference and training jobs that need pod-level throughput, for teams already committed to Google Cloud.","sources":[{"label":"Ironwood: the first Google TPU for the age of inference","url":"https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/ironwood-tpu-age-of-inference/"}]},{"id":"amazon-sys-ultraserver","name":"Trn2 / Trn3 UltraServer","segment":"Data Center","silicon_type":"System","family":"Rack-scale AI system","generation":"Trainium2 / Trainium3","status":"shipping","description":"AWS's rack-scale Trainium deployment — many chips wired together with NeuronLink and EFA networking, available only as EC2 instances, never sold as standalone hardware.","market_share_pct":null,"revenue_usd":null,"url":"https://aws.amazon.com/ec2/instance-types/trn2/","vendor":"amazon","cloud_providers":["aws"],"specs":[{"label":"Trn2 UltraServer scale","value":"64 Trainium2 chips"},{"label":"Trn2 UltraServer compute (FP8)","value":"83.2 PFLOPS"},{"label":"Trn2 UltraServer memory","value":"6 TB HBM, 185 TB/s aggregate bandwidth"},{"label":"Trn2 UltraServer networking","value":"12.8 Tbps EFAv3"},{"label":"Trn3 UltraServer scale","value":"up to 144 Trainium3 chips / 362 MXFP8 PFLOPS"},{"label":"Access model","value":"AWS only — not sold as hardware"}],"availability_status":"shipping","availability_note":"Trn2 UltraServers generally available since December 2024; Trn3 UltraServers since December 2025.","ideal_workload":"Frontier-scale training and inference on AWS needing more chips than a single instance provides.","sources":[{"label":"Amazon EC2 Trn2 Instances","url":"https://aws.amazon.com/ec2/instance-types/trn2/"},{"label":"Announcing Amazon EC2 Trn3 UltraServers","url":"https://aws.amazon.com/about-aws/whats-new/2025/12/amazon-ec2-trn3-ultraservers/"}]}]}