Vera Rubin NVL72 vs GB300 vs GB200 NVL72: A Full Comparison Guide to NVIDIA's AI Racks

Three NVIDIA rack SKUs now sit on the same buyer shortlist: GB200 NVL72, GB300 NVL72, and Vera Rubin NVL72. Each pairs 72 GPUs with 36 CPUs in a liquid-cooled 72-GPU NVLink domain. The differences that actually change a purchase order are memory generation, NVLink generation, and how honest the public power and availability numbers are.

This full guide is the sourced comparison: Vera Rubin NVL72 vs GB300 vs GB200 NVL72, using NVIDIA's own product tables as of October 14, 2026. It is the follow-on to Inside Deep Tech's breakout Vera Rubin NVL72 full guide and pairs with the earlier GB200 NVL72 full guide.

Here's the shape of the piece: what each rack is, a side-by-side spec table, memory and fabric deltas, performance claims with their footnotes, facility and cooling constraints, who should buy which generation, and the honest limits. For non-NVIDIA options, keep the AMD Instinct MI350 guide and the best AI inference chips of 2026 comparison open.

Last updated: October 14, 2026.

Key Takeaways

• **Vera Rubin NVL72** lists **20.7 TB** of HBM4 at **1,400 TB/s** and **216 TB/s** of NVLink 6; production shipments began August 2026 per Inside Deep Tech's Vera Rubin guide.
• **GB300 NVL72** keeps fifth-generation NVLink at **130 TB/s** but lifts GPU memory to **20 TB** (up to **576 TB/s**) on Blackwell Ultra.
• **GB200 NVL72** is the Blackwell baseline: **13.4 TB** HBM3E at **576 TB/s** and **130 TB/s** NVLink.
• NVIDIA claims Vera Rubin can cut inference to **one-tenth** the cost per million tokens vs GB200 on Kimi-K2-Thinking (**32K**/**8K** ISL/OSL); treat that as a vendor projection.
• Honest limits: Vera Rubin's public table still omits rack kW; GB200 facility planning still starts near **120 kW** continuous; supply and HBM allocation still gate all three.

What each NVL72 rack actually is

All three products share the MGX NVL72 idea: liquid-cooled trays, a copper NVLink spine, and a single 72-GPU domain. The CPU and GPU names change the bill of materials.

GB200 NVL72 connects 36 Grace CPUs and 72 Blackwell GPUs. NVIDIA markets it for real-time trillion-parameter LLM inference and MoE training, with a 72-GPU NVLink domain quoted at 130 TB/s.

GB300 NVL72 swaps in 72 Blackwell Ultra GPUs with the same 36 Grace CPUs. NVIDIA positions it for test-time scaling and AI reasoning, stating 1.5x more dense FP4 Tensor Core FLOPS and 2x higher attention performance versus Blackwell GPUs.

Vera Rubin NVL72 rebuilds the platform around 72 Rubin GPUs and 36 Vera CPUs, NVLink 6, ConnectX-9, and BlueField-4. NVIDIA's page says it is available now on the third-generation MGX NVL72 design, with cable-free trays and more than 80 MGX partners.

NVIDIA Vera Rubin: A Full Guide to the Rubin GPU, Vera CPU, and NVL72 Racks After Blackwell

Deep dive on Rubin GPU, Vera CPU, HBM4, NVLink 6, Groq 3 LPX, and honest limits. Start here before comparing racks.

Inside Deep Tech

Side-by-side: Vera Rubin NVL72 vs GB300 vs GB200

The table below uses NVIDIA's published rack rows. Sparse and dense footnotes differ by product page, so FLOPS columns are labeled the way NVIDIA labels them.

Source: NVIDIA GB200 NVL72, NVIDIA GB300 NVL72, and NVIDIA Vera Rubin NVL72 product pages, retrieved October 14, 2026. GB200 rack power cross-checked against Inside Deep Tech's GB200 NVL72 full guide. Per-GPU GB200/GB300 memory rows are arithmetic from rack totals where NVIDIA does not print a per-GPU GB figure on the rack page.

Memory: HBM3E capacity vs HBM4 bandwidth

Memory is the cleanest three-way split. GB200 holds **13.4 TB** of HBM3E at **576 TB/s**. GB300 lifts capacity to **20 TB** while quoting the same up-to-**576 TB/s** class of aggregate bandwidth. Vera Rubin moves to HBM4 at **20.7 TB** and **1,400 TB/s**, with each Rubin GPU listed at **288 GB** and **19.2 TB/s**.

That means GB300 is mostly a capacity upgrade inside the Blackwell Ultra memory subsystem. Vera Rubin is a generation change: more bytes and **1,400 TB/s** of aggregate HBM bandwidth versus the **576 TB/s** Blackwell NVL72 figures.

Here's why that matters for you. Long-context and high-interactivity inference burn HBM for KV cache. If your batch sizes are memory-bound on GB200, GB300's 20 TB pool is the incremental Blackwell answer. If you need the HBM4 bandwidth step and the Vera host memory pool (up to **54 TB** LPDDR5X), you are in Rubin territory.

For packaging and stack supply that still gate every GPU, see Inside Deep Tech's adjacent infrastructure guides on liquid cooling and 800 VDC power distribution.

Fabric: NVLink 5 domain vs NVLink 6

GB200 and GB300 both quote **130 TB/s** of NVLink domain bandwidth on fifth-generation NVLink. Vera Rubin quotes **216 TB/s** on sixth-generation NVLink, with **3 TB/s** per GPU and **6 TB/s** per Superchip on the same table.

Scale-out also moves. GB300 materials emphasize ConnectX-8 at **800 Gb/s** per GPU into Quantum-X800 or Spectrum-X. Vera Rubin materials emphasize ConnectX-9 and **32.4 TB/s** of bidirectional networking bandwidth at rack scale, plus Spectrum-X / Quantum-X800 scale-out.

Inside Deep Tech's take: do not treat NVLink upgrades as a substitute for campus optics. Scale-up stays inside the rack. Scale-out still needs InfiniBand or Ethernet engineering. For switch-side optics, the co-packaged optics full guide covers why CPO shows up in Spectrum-class roadmaps.

FLOPS and marketing multipliers, with footnotes

Peak math is where buyers get played. GB200 lists NVFP4 at **1,440** | **720** PFLOPS sparse|dense. GB300 lists FP4 at **1,440** | **1,080** PFLOPS, with a footnote that Tensor Core specs are with sparsity unless noted, and the second figure is without sparsity. Vera Rubin lists **3,600** PFLOPS sparse NVFP4 inference and **2,520** PFLOPS dense NVFP4 training.

NVIDIA's Vera Rubin page claims one-tenth the cost per million tokens versus GB200 NVL72 and up to **10x** more tokens per megawatt, both measured on Kimi-K2-Thinking at 32K/8K input/output lengths, labeled subject to change. It also claims one-fourth the GPUs to train a **10T**-parameter MoE on **100T** tokens in one month, marked projected.

GB300's headline versus Hopper is up to a **50x** overall AI factory output increase, with DeepSeek-R1 at 32K/8K ISL/OSL and FP4 Dynamo disaggregation versus H100 FP8, also projected. GB200's classic multipliers versus H100 remain **30x** LLM inference, **4x** training, **25x** energy efficiency under NVIDIA's published latency and cluster assumptions.

That means a fair bake-off uses one model, one ISL/OSL, one precision, and one interactivity target. Do not paste sparse GB200 numbers next to dense Vera Rubin training numbers and invent a multiplier.

CPU side: Grace vs Vera

GB200 and GB300 both use Grace with **2,592** Arm Neoverse V2 cores and **17 TB** of LPDDR5X at **14 TB/s** across the rack. Vera Rubin swaps the host for Vera: **3,168** custom Olympus cores, **6,336** threads, and up to **54 TB** of LPDDR5X.

Agentic workloads push tool calls and sandboxes onto CPU. That is why NVIDIA sells Vera CPU racks as a sibling product line in the Vera Rubin platform. If your cluster is still GPU-bound decode with light host work, Grace on GB300 may be enough. If you are building agent factories with heavy CPU sandboxes, Vera's core and memory envelope is part of the product, not a footnote.

Facility power, cooling, and what is still unpublished

All three racks assume direct-to-chip liquid cooling. Vera Rubin publishes a **45°C** warm-water inlet, which NVIDIA ties to dry-cooler friendly designs. GB300 is described as fully liquid-cooled. GB200 facility practice, covered in Inside Deep Tech's GB200 guide, still plans around roughly **120 kW** NVIDIA DGX class draw and **125**-**135 kW** OEM operating bands.

What NVIDIA does not print on the Vera Rubin or GB300 marketing tables retrieved for this guide is a single continuous rack-kilowatt figure. Do not invent one. Get the OEM SKU sheet, CDU topology, and whip ampacity in writing before you compare TCO.

Power density is also why Inside Deep Tech keeps the 800 VDC guide and liquid cooling guide next to every NVL72 article. Copper cross-section and CDU redundancy fail before FLOPS slides do.

Who should buy which rack

Source: NVIDIA product pages linked above and Inside Deep Tech peer guides. Preference rows are editorial judgments on those published specs, not NVIDIA recommendations.

Inside Deep Tech's take

Inside Deep Tech's take: Vera Rubin NVL72 vs GB300 is not a beauty contest. GB300 is the Blackwell Ultra capacity refresh on the same 130 TB/s NVLink island. Vera Rubin is the first rack where memory, CPU, and scale-up fabric all jump a generation at once. GB200 remains the volume workhorse while allocation and facility packages catch up.

The expensive mistake is paying Rubin prices for a GB300-shaped workload, or buying GB300 into a hall that still cannot reject 120 kW-class heat. Match the bottleneck, then the SKU.

A practical three-rack deployment checklist

Before any PO, force the integrator to answer the same five questions for GB200, GB300, and Vera Rubin. First, continuous IT kilowatts per rack plus CDU pump draw. Second, facility water chemistry and inlet temperature, including whether a **45°C** Vera Rubin inlet is actually available year-round. Third, NVLink cartridge and switch-tray sparing. Fourth, scale-out NIC generation (ConnectX-8 versus ConnectX-9) and whether Quantum-X800 or Spectrum-X is already in the hall. Fifth, acceptance tests on your model, not on NVIDIA's Kimi or DeepSeek charts.

Teams that already operate GB200 can often reuse cold-plate procedures, leak detection, and MGX logistics when moving to GB300. The Vera Rubin jump is larger: new CPU, new HBM generation, new NVLink generation, and new SuperNIC class. Budget bring-up time accordingly, even when NVIDIA describes a seamless third-generation MGX footprint on the Vera Rubin page.

Of course, software continuity helps. CUDA and NCCL familiarity transfer. What does not transfer automatically is thermal soak behavior under sustained all-reduce at the new memory bandwidth, or host-side sandbox density on Vera versus Grace. Run those tests before you retire capacity planning assumptions from the GB200 guide.

Where this comparison sits in the IDT stack

This article is the rack-level comparison. The platform deep dives remain the Vera Rubin NVL72 full guide and the GB200 NVL72 full guide. Power and cooling prerequisites live in the 800 VDC and liquid cooling guides. Switch optics context is in the CPO guide. Competitive silicon context is in the MI350 and inference chips guides.

Use those pages for depth. Use this page when a CFO or facilities lead asks which of the three NVL72 SKUs belongs on the 2026-2027 floor plan.

Honest limits

First, power opacity. Vera Rubin and GB300 public pages reviewed here do not publish a continuous rack-kW number. GB200 planning still leans on ~**120 kW** class figures from DGX/OEM materials summarized in the GB200 guide.

Second, footnote traps. Sparse versus dense, with-sparsity versus without-sparsity, and projected versus measured claims differ across the three pages. Vera Rubin's **10x** tokens/MW and **1/10** cost claims are workload-bound.

Third, supply. Rubin HBM4 and Blackwell Ultra HBM3E both compete for packaging and DRAM capacity. Allocation, not brochure FLOPS, often decides ship dates.

Fourth, software maturity. CUDA and NCCL continuity help, but early-lifecycle racks still need your own acceptance tests. Vendor charts are not SLAs.

Fifth, non-NVIDIA paths. If lock-in risk dominates, price AMD Instinct MI350 and the wider inference chip comparison before you standardize on NVL72 forever.

FAQ

What to do Monday

Normalize every quote onto one sheet: GPU memory bytes and bandwidth, NVLink domain TB/s, sparse-versus-dense FLOPS, OEM continuous kW, and CDU topology. Then re-run your own model at a fixed interactivity target on whatever generation you can actually touch.

For the platform deep dives behind this comparison, start with Inside Deep Tech's Vera Rubin NVL72 full guide and GB200 NVL72 full guide. For power and cooling prerequisites, use the 800 VDC and liquid cooling guides. For more frontier coverage, return to the Inside Deep Tech homepage.