Blog ▸ GB300 NVL72 vs B300 HGX: When the Rack Beats the Node
GPU Infrastructure
Same silicon, two very different units of purchase. Why a 72-GPU rack quotes below an 8-GPU node per hour, and the three questions that decide which one you should commit to.
GB300 NVL72 vs B300 HGX: When the Rack Beats the Node
GPUaaS.com Team
GPU Infrastructure
October 15, 2026
No items found.
GB300 NVL72 quotes at $5.00 to $5.50 per GPU-hour on committed terms in Q4 2026, the bottom of the Blackwell Ultra range, while HGX B300 8-GPU nodes quote at $5.00 to $6.50, because the rack is sold as one 72-GPU unit on long terms to a small pool of buyers who can site it.
Both parts carry the same Blackwell Ultra GPU with 288GB of HBM3e. What differs is the unit you buy. An HGX B300 node is eight GPUs on a baseboard, NVLink inside the box, InfiniBand or Ethernet between boxes. A GB300 NVL72 is 72 GPUs and 36 Grace CPUs in one liquid-cooled rack with a single NVLink domain across all of them. That difference decides price, lead time, where it can physically go, and which workloads actually benefit. This post is for buyers who have seen both bands in the October benchmark and want to know why the bigger thing is cheaper per hour.
Key takeaways
GB300 NVL72 committed band is $5.00 to $5.50 per GPU-hour this month; HGX B300 is $5.00 to $6.50. The rack quotes at the floor of the range, the node spans it
NVL72 is sold as a whole rack: 72 GPUs minimum, usually twelve-month terms or longer. HGX B300 starts at one 8-GPU node on short or long commitments
NVL72 needs liquid cooling and rack power in the 120 to 140kW range, which narrows both the provider pool and the lead time. HGX B300 can be air-cooled at lower density
The 72-GPU NVLink domain matters for trillion-parameter inference and large training runs. For most inference under 100B parameters, 8-GPU nodes on InfiniBand are enough
Demand for NVL72 sits below B300, B200, H200 and H100 SXM in the October ranking, roughly level with RTX PRO 6000. Supply is thin but the few buyers are large
Rule of thumb: if the workload fits in one 8-GPU NVLink domain, buy HGX. If it needs 16 or more GPUs sharing one memory fabric, price the rack
◆ WHAT YOU ARE ACTUALLY BUYING
Same GPU, different unit, different fabric
The GPU inside both products is Blackwell Ultra: 288GB HBM3e per GPU, FP4 throughput roughly 1.5 times a B200, FP64 effectively deprecated. On a per-GPU spec sheet the two are identical, which is why the question comes up at all. Everything that separates them is in how the GPUs are connected and packaged.
An HGX B300 baseboard holds eight GPUs connected by NVLink inside the node. Between nodes you have InfiniBand (usually 400G or 800G per GPU) or Ethernet. A job that needs more than eight GPUs talking to each other crosses that boundary and takes the latency and bandwidth hit of going over the network fabric. That is fine for most work. It is the configuration behind nearly every H100, H200 and B200 cluster on the network today.
A GB300 NVL72 rack puts 72 GPUs and 36 Grace CPUs on one NVLink switch fabric. Every GPU in the rack can read every other GPU's memory at NVLink speed, which gives you about 20TB of HBM in a single domain. The rack arrives as one assembled, liquid-cooled unit. You do not take eight GPUs of it. You take the rack.
That packaging is the whole story for pricing. A provider holding HGX nodes can sell one node to one buyer and seven to another. A provider holding an NVL72 has one thing to sell, it cost a great deal to install, and it sits empty until one buyer takes all 72 GPUs for a long enough term to make the install pay.
◆ GB300 NVL72 VS HGX B300, COMMITTED TERMS, Q4 2026
GB300 NVL72
HGX B300 node
Committed band per GPU-hour
$5.00 to $5.50
$5.00 to $6.50
Minimum purchase
One rack, 72 GPUs
One node, 8 GPUs
Typical term
12 months or longer
3, 6 or 12 months
NVLink domain
72 GPUs
8 GPUs
HBM per NVLink domain
About 20TB
About 2.3TB
Cooling
Liquid, required
Air or liquid
Rack power
Roughly 120 to 140kW
Roughly 14 to 20kW per node
Provider pool on the network
Small
Moderate and growing
Typical lead time
Longest in the Blackwell range
Weeks, varies by region
Committed-term bands from the GPUaaS.com October 2026 benchmark. Power and cooling figures are typical configurations and vary by OEM and site. Contact for commitment terms. *Prices may vary depending on commitment terms.
◆ WHY THE RACK QUOTES AT THE BOTTOM
Three reasons a 72-GPU unit is cheaper per hour than an 8-GPU one
The first reason is term. Nobody takes an NVL72 for three months. The buyers who want it are running production inference on frontier-scale models or multi-week training runs, and they sign twelve months or longer. Long terms land at the bottom of every band on the network, and the NVL72 band is made almost entirely of long terms. The B200 pricing post shows the same effect on a part with a wider mix of terms.
The second reason is the unit economics on the provider side. An NVL72 is a single capital item with liquid cooling, high-density power and a specialist install behind it. The provider's risk is not that one node sits idle but that the whole rack does. Pricing at the floor of the range to get a twelve-month commitment is rational when the alternative is a seven-figure asset producing nothing. HGX nodes carry less of that risk, so providers hold the price higher and wait.
The third reason is the size of the buyer pool. The October demand ranking puts NVL72 well below B300 and B200, roughly level with RTX PRO 6000, and the requests that do come in are large. A provider who has an NVL72 sitting unsold is quoting into a market with a handful of qualified buyers, and the price reflects that. The demand snapshot covers where the rest of the demand sits.
None of this means the rack is cheap. $5.25 per GPU-hour times 72 GPUs times 8,760 hours is a little over $3.3 million a year before anything else, and the buyer is committing to all of it. It means the per-GPU rate is at the floor because the commitment is at the ceiling.
72
GPUs in one NVLink domain on a GB300 NVL72, against 8 on an HGX B300 node. The rack has nine times the shared memory fabric, which is the only reason to pay for it
GPUaaS.com, October 2026
◆ WHERE THE RACK ACTUALLY WINS
The workloads that need 72 GPUs on one fabric, and the ones that do not
The NVL72 earns its keep when a single model or a single job needs more memory, or more GPUs in lockstep, than eight GPUs can give. Serving a trillion-parameter mixture-of-experts model with long context at low latency is the canonical case. The whole model and its KV cache fit in one NVLink domain, so decode steps do not cross the InfiniBand boundary, and throughput per GPU is materially higher than the same model sharded across nine HGX nodes. NVIDIA's own figures claim large multiples over Hopper for exactly this workload; the honest version is that the gain is real and depends heavily on model size and batch.
Large pretraining runs are the second case. A training job across 72 GPUs on one fabric avoids the all-reduce over the network entirely. Across 256 or 512 GPUs it still has to cross racks, but each rack contributes 72 GPUs of tightly coupled compute rather than eight, which changes the parallelism strategy and usually the step time.
Most inference does not look like that. A 70B model serves comfortably on one HGX B300 node with room to spare, and a 400B model on two nodes over InfiniBand is a solved problem. Fine-tuning a 13B or 70B model does not need the rack. Rendering, simulation and anything that parallelises across independent jobs does not need the rack. For those buyers the HGX node is the right unit, the term can be shorter, and the quote lands wherever the benchmark says it should for that term and size.
If you are not sure which side you fall on, ask one question: does the job need more than about 2TB of HBM in a single memory domain, or more than eight GPUs synchronising every step. If no, the rack is paying for a fabric you will not use.
◆ LEAD TIME, FACILITY AND WHO CAN SELL IT
The practical constraints that decide whether you can have one at all
Price is the easy part of the NVL72 decision. The hard parts are lead time and facility. The rack needs direct liquid cooling to the chip, rack power somewhere around 120 to 140kW, and a site that can deliver both with redundancy. Most air-cooled facilities that host HGX nodes today cannot host an NVL72 without a plant upgrade, so the matchable provider pool is a fraction of the B300 pool. The routing post explains how that pool gets narrowed; for NVL72 it narrows fast.
Lead time follows from that. Racked, available NVL72 capacity exists on the network in North America and a smaller amount in Europe, but it is not stocked the way H100 is. A request often means a provider confirming a rack that is installed but not yet handed over, and the quote comes back inside the 24-hour window rather than within a few hours. Buyers with a hard start date should say so up front; buyers who can presell against a future install date get a better quote, which the presell post covers from the provider side.
Facility also cuts the other way. A buyer planning an owned-style deployment in their own air-cooled data hall should rule the rack out on cooling alone and price HGX B300 at density instead. The DGX vs HGX post covers the branded-system question that often comes up at the same time.
For single Blackwell Ultra GPUs rather than a node or a rack, packet.ai handles per-token and single-GPU access on the same hardware, which is where most teams should start before committing to either unit.
◆ HOW TO SPECIFY EITHER ONE
What to put in the request so the quote lands inside the band
For HGX B300: state the term first, the GPU count in multiples of eight, whether the nodes must share one InfiniBand fabric or can be non-contiguous, the region or residency constraint, and the start date. Quotes come back within a few hours from 20+ vetted partners. Minimum is one node. No buyer fees. The contract runs between the provider, GPUaaS.com and the buyer.
For GB300 NVL72: state the term (expect twelve months to be where the floor is), the number of racks, the region, the start date, and whether you have a hard residency requirement. Expect the quote inside 24 hours. Minimum is one rack. If your start date is more than a quarter out, say so, because a presell against an install date is often the cheapest way to get one.
Either way, the band is the band. A $5.00 NVL72 quote and a $5.00 HGX B300 quote are both real this month, and they are buying two different things.
Price the rack or the node against this month's band.
Submit a spec and quotes come back within a few hours, sourced across 20+ vetted partners. No buyer fees. 8-GPU node minimum.
$5.00 to $5.50 per GPU-hour on committed terms through the GPUaaS.com network in October 2026, sold as a whole 72-GPU rack on twelve-month or longer terms. That works out to a little over $3.3 million a year per rack at the midpoint, before power or support terms.
Because it is sold on longer terms, as one indivisible unit, to a small pool of buyers. Long terms land at the floor of every band, the provider would rather fill a seven-figure rack at the floor than leave it empty, and demand for the rack is thinner than for 8-GPU nodes.
The GPU is the same Blackwell Ultra part. NVL72 puts 72 GPUs and 36 Grace CPUs on one NVLink fabric in a liquid-cooled rack. HGX B300 puts eight GPUs on a baseboard with NVLink inside the node and InfiniBand or Ethernet between nodes, and can be air-cooled.
Serving trillion-parameter or long-context models where the whole model and KV cache should sit in one memory domain, and large training runs that benefit from 72 GPUs synchronising without crossing the network fabric. Most inference under 100B parameters and most fine-tuning fit comfortably on HGX nodes.
Not through the network. The rack is quoted and contracted as one 72-GPU unit because the NVLink domain and the liquid-cooling install do not split. For 8 to 64 GPUs, HGX B300 nodes are the right request. For single GPUs, packet.ai handles self-serve access.
Longer than any other Blackwell part on the network, because the rack needs liquid cooling and high-density power that most facilities cannot provide. Quotes usually come back inside 24 hours. Hand-over depends on whether the provider has a rack installed or is confirming one against an install date.
