NVL72 HBM Capacity Comparison: GB200 vs GB300 vs Rubin (2026)

GB200 NVL72, GB300 NVL72, and Vera Rubin NVL72 are the three racks any NVL72 HBM capacity comparison has to cover, and each shipped with more HBM than the rack before it. That's the headline most coverage leads with, and it's true: 13.4TB, then 20TB, then 20.7TB per rack. What almost nobody says is that the number actually gating how large an MoE expert-parallel group can scale, the NVLink domain size, hasn't moved at all across those three racks. It's been fixed at 72 GPUs since Blackwell shipped, through two full NVLink generations.

That distinction matters if you're sizing a rack reservation around a mixture-of-experts model rather than a dense one. This post separates what actually changed (HBM capacity, HBM type, NVLink bandwidth) from what didn't (domain size), and names the one system that will actually move it: Rubin Ultra's Kyber rack, which isn't shipping until 2028.

TL;DR: NVL72 HBM Capacity Comparison, GB200 vs GB300 vs Rubin

• GB200 NVL72: 72 B200 GPUs, 13.4TB total HBM3e (186GB/GPU in-rack), one 72-GPU NVLink 5 domain at 130TB/s, per NVIDIA's GB200 NVL72 spec page.
• GB300 NVL72: 72 B300 GPUs, 20TB total HBM3e (288GB/GPU), same 72-GPU NVLink 5 domain at 130TB/s, per NVIDIA's GB300 NVL72 spec page.
• Vera Rubin NVL72: 72 Rubin GPUs, 20.7TB total HBM4 (288GB/GPU at 22TB/s), still a 72-GPU domain, now on NVLink 6 at 260TB/s.
• Domain size: unchanged at 72 GPUs across three racks and two NVLink generations. The jump happened one generation earlier, Hopper's 8-GPU NVLink 4 domain to Blackwell's 72-GPU NVLink 5 domain.
• The real jump: Rubin Ultra's Kyber rack doubles the domain to 144 GPUs, but it slipped from 2027 to 2028.
• Spheron takes reservations on GB200 NVL72 and GB300 NVL72 today, with Vera Rubin open for pre-order; B200 and B300 rent on-demand right now for workloads that fit a single node.

What NVLink Domain Size Actually Controls (and Why It Isn't the Same as GPU Count)

NVLink domain size is the number of GPUs that can reach every other GPU in one hop, at full NVLink bandwidth, through a shared NVSwitch fabric. It is not the same thing as how many GPUs sit in your cluster. A training job can span 32,000 GPUs, but those GPUs are organized into many separate NVLink domains, each one a rack, connected to each other by InfiniBand or Ethernet instead of NVLink. Traffic that stays inside one domain moves at terabytes per second in a single switch hop. Traffic that crosses domain boundaries drops to whatever the inter-rack network provides, and usually takes more than one hop to get there. Our NVLink explainer covers the PCIe-versus-NVLink gap in more detail if you need that foundation first.

NVIDIA's own description of GB200 NVL72 calls it "a single 72-chip domain", and is explicit that a larger domain is possible with additional switch tiers, up to 576 GPUs at over 1 PB/s of aggregate bandwidth. That 576-GPU figure describes a different, larger topology than the 72-GPU rack sold as the standard NVL72 product; it is not what ships when you reserve a GB200, GB300, or Vera Rubin NVL72 rack today. The rack-scale product's domain is 72 GPUs, full stop.

NVL72 HBM Capacity Comparison: GB200 vs GB300 vs Vera Rubin, Side by Side

The GB200 row comes straight from NVIDIA's GB200 NVL72 spec page, the GB300 row from NVIDIA's GB300 NVL72 spec page, and the NVLink-generation bandwidth figures from GPUSmith's NVLink 5 vs NVLink 6 comparison.

Two things stand out reading this table left to right instead of generation to generation. First, GB300 and Vera Rubin land on an identical 288GB-per-GPU figure despite being different memory generations (HBM3e versus HBM4), which is why the real differentiator between those two racks is bandwidth, not capacity: Vera Rubin's HBM4 runs at 22TB/s per GPU against GB300's 8TB/s HBM3e. Second, and the point of this whole comparison, the rightmost column never changes. Every rack in this table is a 72-GPU domain. GB200 NVL72's full rack spec table and the deeper GB300 vs GB200 pricing and availability comparison cover the compute-side differences this table leaves out (dense FP4 throughput, FP64, networking, and rack power and liquid cooling requirements), if you're deciding between those two specifically.

One figure worth flagging on its own: a standalone B200 card is rated at 192GB of HBM3e, not the 186GB shown above. NVIDIA publishes 186GB specifically for the per-GPU figure inside the NVL72 rack configuration, and conflating the two is a common error in rack-spec writeups.

Why Domain Size Matters More for MoE Models Than Dense Ones

A dense model shards its one set of weights across GPUs with tensor or pipeline parallelism, and the communication pattern is predictable: mostly neighbor-to-neighbor, with an occasional all-reduce. A mixture-of-experts model has a different problem. Every token has to be routed to whichever experts it was assigned to, which are scattered across GPUs, and the result is a genuinely all-to-all communication pattern at every MoE layer, not an occasional one.

That's where domain size stops being an abstraction. When an expert-parallel group fits inside one NVLink domain, every GPU-to-GPU hop in that all-to-all shuffle stays on the switch fabric. When it doesn't, the group gets forced onto InfiniBand instead, multi-hop, at a fraction of the bandwidth, and that's reportedly why GB200 and GB300's one-hop, all-to-all 72-GPU fabric is what lets these racks dominate wide-expert-parallel serving of large MoE models in the first place, according to SemiAnalysis's InferenceX analysis.

NVIDIA's own numbers back up how much that's worth: it credits the 72-GPU NVLink domain with a 30x inference speedup on a 1.8-trillion-parameter GPT-MoE model at FP4, and 4x faster training on a 32,000-GPU GB200 NVL72 cluster, versus equivalent H100 deployments. A real MoE training run backs this up outside NVIDIA's own marketing, too: Reflection AI pretrained its 501B-parameter MoE model, Beam, on 6,144 GB300 NVL72 GPUs in under four weeks at 92.3% goodput, roughly 85 full 72-GPU domains' worth of hardware stitched together, a workload our Beam deployment walkthrough covers in more depth. A dense model at that same GPU count would care far less about where the domain boundaries fall.

The Domain Size Jump Nobody's Renting Yet: Rubin Ultra's Kyber NVL144

If the goal is a bigger NVLink domain, not just more HBM inside the same 72-GPU domain, the system to watch isn't base Vera Rubin. It's Rubin Ultra, which ships in NVIDIA's new Kyber rack at NVL144 scale, doubling the GPU-package domain to 144 versus Vera Rubin NVL72's 72. NVL576 goes a step further, linking eight Kyber racks together over co-packaged optics into one larger pod.

That's the domain-size jump that actually matters for MoE expert-parallel scale, and it's not available to rent anywhere yet. Kyber slipped more than 12 months, from an original 2027 target to 2028, reportedly over manufacturing problems with the rack's 78-layer PCB midplane. Our Kyber delay writeup goes through what NVIDIA has and hasn't confirmed about that slip. Until Kyber actually ships, GB300 NVL72 and Vera Rubin NVL72 are the ceiling on domain size that any cloud customer can book, and that ceiling has been 72 GPUs since Blackwell launched.

When a Bigger Domain Actually Changes Your Rental Decision

The memory increases from GB200 to GB300 to Vera Rubin are real and useful on their own: more HBM per rack means a bigger model, or more KV cache headroom, fits inside one domain before you need to cross a domain boundary at all. But if your MoE model's expert-parallel group already fits inside a 72-GPU domain, GB300's extra 6.6TB over GB200, or Vera Rubin's extra bandwidth over GB300, doesn't change your scale-out ceiling. It just means you hit that ceiling with more memory to spare. The domain-size decision and the HBM-capacity decision are two separate questions, and conflating them is the mistake most rack-spec comparisons make.

Spheron reserves GB200 NVL72 and GB300 NVL72 capacity from a single node up to a full 72-GPU rack, sourced from its data center partner network, with Vera Rubin open for pre-order ahead of its H2 2026 ship date. None of the three carry a published hourly rate; reservation requests typically come back with a sourced quote within 24 to 48 hours. If your model fits inside a single 8-GPU node's worth of domain, a non-rack B200 or B300 cluster is cheaper and faster to provision than reserving rack-scale capacity for headroom you won't use, and both rent on-demand today at $11.12/hr and $12.42/hr respectively. And if what you actually want is a domain bigger than 72 GPUs, there's no honest answer to give you yet. Rubin Ultra's 144-GPU Kyber domain isn't reservable from any provider, Spheron included, and isn't expected before 2028.

Pricing fluctuates based on GPU availability. Spheron rates above are live as of 27 Oct 2026; other providers reflect their most recent published rates and may have changed. Check current GPU pricing → for live rates.

Rack-scale NVLink domain stays at 72 GPUs through GB200, GB300, and Vera Rubin, so the HBM capacity gains are worth having, but they won't change your expert-parallel ceiling. Reserve GB200 NVL72 or GB300 NVL72 capacity, or rent B200 and B300 on-demand today.

Get started on Spheron →