NVIDIA Rubin Ultra (V300)
Product Overview
NVIDIA Rubin Ultra (codename V300) is the flagship-enhanced variant of the Rubin architecture, announced at GTC 2026 (2026-03-17) alongside Rubin R200 and positioned as the top product for hyperscale training/inference. The original design uses a 4-die MCM package with 576GB HBM4e per card, expandable via NVL144 racks to 144 GPUs — NVIDIA's flagship answer for sovereign AI and yotta-scale clusters.
⚠️ Architecture revision notice: According to SemiAnalysis, the original 4-die Rubin Ultra design has been cancelled due to packaging manufacturing difficulty; the new version is scaled down to roughly half the size, with performance halved accordingly. The specs below are based on originally published information; final production specs are subject to NVIDIA's official announcements.
📌 2026-09 spec-cut report (SemiAnalysis / STAR Market Daily): Rubin Ultra's HBM configuration may be downgraded from HBM4E 12-Hi (384GB tier) to HBM4 8-Hi (192GB), with Samsung co-developing the 8-high product. The driver: after HBM/DRAM price hikes, memory already accounts for roughly 40% of total system TCO; the downgrade cuts HBM cost by more than 50%, and the saved budget shifts to the Scale-up network — under an NVL576 NPO scheme, the Scale-up network's share of rack spending rises from 4% to 12%. TrendForce also shows NVIDIA evaluating HBM4E 8-Hi / HBM4E 12-Hi / HBM4 8-Hi options in parallel from 2026 Q3. Final configuration subject to official announcements.
Core Specifications (original design, pending final official confirmation)
| Spec | Value |
|---|---|
| GPU Architecture | Rubin architecture (MCM multi-chip module) |
| Package | Original 4-die design (⚠️ under revision) |
| Process | Presumed TSMC 3nm (same as Rubin R200) |
| FP4 Inference Compute | Presumed ~100 PFLOPS (original, dense) |
| FP8 Training Compute | Presumed ~70 PFLOPS (original) |
| Memory Capacity | 576 GB HBM4e (original design) |
| Memory Type | HBM4e |
| Memory Bandwidth | Presumed >44 TB/s (HBM4e) |
| Interconnect | NVLink 7 (presumed, >3.6 TB/s unidirectional) |
| TDP | Presumed 3000W+ (liquid cooling mandatory) |
| Announced | March 17, 2026 (GTC 2026) |
| Mass Production | Presumed 2027 H2 - 2028 (original timeline) |
| Pricing | Presumed $15-21 million per NVL144 rack |
Comparison with Rubin R200
| Item | Rubin R200 | Rubin Ultra (original) |
|---|---|---|
| GPUs/rack | 72 (NVL72) | 144 (NVL144) |
| Memory/GPU | 288 GB HBM4 | 576 GB HBM4e |
| HBM/rack | 20,736 GB | 82,944 GB |
| Mass production timing | 2026 H2 | 2027 H2 - 2028 |
| Positioning | Volume workhorse | Flagship-enhanced |
NVL144 rack: Rubin Ultra forms racks of 144 GPUs with 82,944 GB of total HBM — 4x the memory capacity of NVL72 (Rubin R200).
Platform Configuration (original NVL144)
- 144 Rubin Ultra GPUs
- 72 Vera CPUs (presumed)
- Total memory: 82.9 TB HBM4e
- Total compute: Presumed ~14 EFLOPS FP4 (original)
- TDP: Presumed ~360kW (full liquid cooling mandatory)
- Mass production: Presumed from 2027 H2
Mass Production and Delivery (original timeline)
- GTC 2026: Announced alongside Rubin R200
- Mass production: Presumed 2027 H2 - 2028 (later than Rubin R200's 2026 H2)
- ⚠️ Revision risk: SemiAnalysis reports the original 4-die design has been cancelled; the new version's specs and schedule may change
Use Cases
Rubin Ultra targets hyperscale data centers and sovereign AI clusters, suited for:
- Ten-trillion-parameter LLM training
- yotta-scale AI factories
- Ultra-long-context inference (576GB memory per card)
- National-scale AI infrastructure
References
- NVIDIA GTC 2026 keynote (Rubin Ultra announcement)
- SemiAnalysis: Rubin Ultra's original 4-die design cancelled; new version scaled down by about half
- SemiAnalysis / STAR Market Daily (2026-09): Rubin Ultra memory may be cut from the 384GB tier to 192GB HBM4, with budget shifted to Scale-up optical interconnect
- Industry supply chain analysis (Foxconn): V300 576GB HBM4e / 144-GPU rack, $21,000/rack
Related Products
- NVIDIA Rubin R200 - Same-generation volume GPU
- NVIDIA Rubin CPX - Inference-optimized variant
- AMD MI455X - Same-generation competitor
- Full comparison table