Entity — NVIDIA
Status: active Owner: Finance / Charlie AGT-002 Entity ID: ENTITY-AI-NVIDIA Visibility: PUBLIC Last verified: 2026-07-10
Snapshot
NVIDIA is the merchant GPU incumbent and reference point for all accelerator comparisons. Current shipping flagship is the B300 (Blackwell Ultra), deployed at rack scale as GB300 NVL72 (72 GPUs, 20 TB HBM3e, 130 TB/s NVLink). The next generation, Rubin (Vera Rubin NVL144, HBM4, NVLink 6), was declared "in full production" at CES 2026 with volume shipments targeted for 2H2026; Rubin Ultra (2027) and Feynman are on the published roadmap.
Chip Table
| Chip | Status | Key spec highlights | As-of | Source | Grade |
|---|---|---|---|---|---|
| B300 / GB300 NVL72 (Blackwell Ultra) | shipping | 288 GB HBM3e, 8 TB/s per GPU; 15 PF dense FP4 (NVFP4) per GPU; rack 1,080 PF dense FP4; NVLink 5 130 TB/s per rack | 2026-07-10 | https://www.nvidia.com/en-us/data-center/gb300-nvl72/ | 🟢 |
| B200 / GB200 NVL72 (Blackwell) | shipping (prior flagship, still ramping in fleet) | dual reticle-sized dies, NV-HBI 10 TB/s die-to-die; per-GPU numbers not re-verified this pass | 2026-07-10 | https://developer.nvidia.com/blog/inside-nvidia-blackwell-ultra-the-chip-powering-the-ai-factory-era/ | 🟢 (architecture facts only) |
| Rubin GPU / Vera Rubin NVL144 | announced (ramp 2H2026) | 336B transistors; up to 288 GB HBM4 at 22 TB/s; vendor claims 50 PF NVFP4 inference, 35 PF NVFP4 training per GPU; NVL144 rack 3.6 EF NVFP4; NVLink 6 | 2026-07-10 | https://www.servethehome.com/nvidia-launches-next-generation-rubin-ai-compute-platform-at-ces-2026/ | 🟡 (keynote via credible secondary) |
| Rubin CPX | announced | inference-specialized GPU for massive-context prefill; part of disaggregated Rubin platform | 2026-07-10 | https://nvidianews.nvidia.com/news/nvidia-unveils-rubin-cpx-a-new-class-of-gpu-designed-for-massive-context-inference | 🟢 (existence/positioning; specs not verified this pass) |
| Rubin Ultra / Feynman | announced (2027 / later) | roadmap placeholders only; no verified specs | 2026-07-10 | https://www.tomshardware.com/pc-components/gpus/nvidia-announces-rubin-gpus-in-2026-rubin-ultra-in-2027-feynam-after | 🟡 |
Technical Trajectory
- Cadence: roughly annual data-center platform steps — Blackwell (2024) → Blackwell Ultra (2025) → Rubin (2H2026) → Rubin Ultra (2027) → Feynman (roadmap). Sources: CES 2026 coverage (2026-01), Tom's Hardware roadmap piece (🟡).
- Memory: HBM3e 288 GB / 8 TB/s per GPU (Blackwell Ultra, 🟢) → HBM4 up to 288 GB / 22 TB/s per GPU claimed for Rubin (🟡, keynote) — bandwidth, not capacity, is the generation jump.
- Interconnect: NVLink 5 (1.8 TB/s per GPU; 130 TB/s per NVL72 rack, 🟢) → NVLink 6 doubling per-link throughput on Rubin (🟡). Rack-scale NVL72/NVL144 is now the product unit, not the chip.
- Node: Blackwell Ultra on TSMC 4NP (🟢, NVIDIA dev blog); Rubin reported on TSMC 3nm-class (🟡). Per-rack power keeps climbing (secondary reports of ~1,400 W per B300 GPU and 600 kW-class future racks are 🟡 — not vendor spec-page numbers).
- Precision direction: NVFP4 is the marketing headline unit from Blackwell Ultra onward; dense-vs-sparse basis is not always stated for Rubin claims — treat unqualified Rubin PFLOPs as A1/A2 until a spec page exists.
Watch Items
- Official Rubin spec page / whitepaper (upgrade Rubin row 🟡 → 🟢; confirm dense-vs-sparse basis of 50/35 PF NVFP4 claims).
- Rubin volume-shipment confirmation in 2H2026 (FY earnings calls, partner system availability).
- MLPerf Training v6.0 round (~mid/late 2026): first Rubin submissions?
- GB300 NVL72 per-GPU TDP from an official source (currently secondary-only).
Claims
- CLAIM-AI-SILICON-001 — NVIDIA is the merchant-GPU denominator in the custom-ASIC-share question; spec/benchmark rows here feed the baseline.
- CLAIM-AI-BOTTLENECK-001 — HBM4 bandwidth step and rack power escalation are direct inputs to where the bottleneck (and pricing power) sits.
Sources
- 🟢 NVIDIA GB300 NVL72 spec page — https://www.nvidia.com/en-us/data-center/gb300-nvl72/ (accessed 2026-07-10)
- 🟢 NVIDIA dev blog, "Inside NVIDIA Blackwell Ultra" — https://developer.nvidia.com/blog/inside-nvidia-blackwell-ultra-the-chip-powering-the-ai-factory-era/ (accessed 2026-07-10)
- 🟢 NVIDIA newsroom, Rubin CPX announcement — https://nvidianews.nvidia.com/news/nvidia-unveils-rubin-cpx-a-new-class-of-gpu-designed-for-massive-context-inference (accessed 2026-07-10)
- 🟡 ServeTheHome, Rubin platform launch at CES 2026 — https://www.servethehome.com/nvidia-launches-next-generation-rubin-ai-compute-platform-at-ces-2026/ (accessed 2026-07-10)
- 🟡 Tom's Hardware, Vera Rubin platform in depth — https://www.tomshardware.com/pc-components/gpus/nvidias-vera-rubin-platform-in-depth-inside-nvidias-most-complex-ai-and-hpc-platform-to-date (accessed 2026-07-10)
- 🟡 Tom's Hardware, roadmap (Rubin/Rubin Ultra/Feynman) — https://www.tomshardware.com/pc-components/gpus/nvidia-announces-rubin-gpus-in-2026-rubin-ultra-in-2027-feynam-after (accessed 2026-07-10)
Changelog
- 2026-07-10: Created; first verification pass.