GB200 and GB300 NVL72
Price, Specs, Rack Power, and How to Rent or Host NVIDIA's Rack-Scale Blackwell
A GB200 NVL72 rents from about $10.50 per GPU-hour, over half a million dollars a month for a full rack, and GB300 capacity is sold by reservation only. We find who can deliver, or the liquid-cooled space to host your own.
Price My NVL72 RequirementGB200 NVL72 and GB300 NVL72 are NVIDIA's rack-scale Grace Blackwell systems: 72 Blackwell GPUs and 36 Grace CPUs in one liquid-cooled rack, joined by NVLink so the whole rack works as a single giant GPU. The GB200 uses the original Blackwell GPU; the GB300 uses Blackwell Ultra, with 288 GB of memory per GPU and about one and a half times the FP4 compute. They are built for training and serving the largest models, and they are priced accordingly: published on-demand rates for a GB200 NVL72 start around $10.50 per GPU-hour at the main clouds, with some specialist listings from about $8, and GB300 NVL72 capacity is sold by quote and reservation only. A rack also draws 120 to 160 kW, which most data centers cannot host, so buyers either rent it through GPU cluster rental or place owned racks through AI and GPU colocation.
Prices checked October 2026, from published price lists and the IntuitionLabs pricing index. GB300 NVL72 has no published hourly rate. Free to cite with a link to this page.
Once you have contacted a provider directly, it usually will not work through an advisor on your deal. Send the requirement to us first and you'll hear back within 24 hours. The provider you choose pays us, so it costs you nothing.
GB200 and GB300 NVL72 Price
| Figure | GB200 NVL72 | GB300 NVL72 |
|---|---|---|
| Per GPU-hour, on demand | About $10.50 to $27, depending on the provider | Not offered by the hour; quote and reservation only |
| Cheapest listings | From about $8 per GPU-hour at some specialist clouds | Quoted case by case |
| Full rack, per hour | About $756 to $1,944 for 72 GPUs | Quoted case by case |
| Full rack, per month | About $552,000 at $10.50, or $841,000 at $16 | Typically above GB200 per GPU, reflecting the extra memory and compute |
| Reserved | Typically 25 to 45 percent below on demand at cluster scale | Reserved capacity is the main way to get it |
| Smallest unit to rent | Instances of four GPUs at some clouds; full racks for training | Generally full racks or multi-rack reservations |
| Buying the rack | About $3.1 million bare and $3.9 million all-in (SemiAnalysis); other estimates $2 million to $3 million | About $3.7 million to $4 million in most estimates; some above $5 million |
Prices checked October 2026. Rack purchase prices are analyst estimates, since NVIDIA publishes none: Tech Insider, IntuitionLabs and SemiAnalysis as widely reported. Monthly figures are MCA calculations at 730 hours.
GB200 vs GB300 NVL72: Specs
The GB300 NVL72 keeps the GB200's rack, NVLink fabric and Grace CPUs, and upgrades the GPUs. Most of the gain is memory: half as much again per GPU, which is what long-context inference and the largest models run out of first.
| Specification | GB200 NVL72 | GB300 NVL72 |
|---|---|---|
| GPUs and CPUs | 72 Blackwell GPUs and 36 Grace CPUs | 72 Blackwell Ultra GPUs and 36 Grace CPUs |
| Memory per GPU | About 186 to 192 GB HBM3e | 288 GB HBM3e |
| Memory per rack | About 13.4 TB HBM3e | About 20.7 TB HBM3e, plus 17 TB of Grace CPU memory |
| NVLink fabric | 130 TB/s across all 72 GPUs | 130 TB/s across all 72 GPUs |
| FP4 compute per GPU | Baseline | About 1.5 times GB200 |
| Rack power | About 120 to 130 kW | About 135 kW typical, peaks near 155 kW; plan for 140 to 160 kW |
| Cooling | Liquid, direct to chip | Liquid, direct to chip |
| Availability | On demand at a few clouds; reserved widely | Quote and reservation only |
From NVIDIA's published specifications and system vendor documentation, October 2026. Power varies with configuration and workload.
Choose GB200 NVL72 when
You need rack-scale training or inference now, at the lowest available rate, and the model fits in about 13 TB of GPU memory. It is the only NVL72 with on-demand pricing.
Choose GB300 NVL72 when
Memory is the limit: very long contexts, the largest reasoning models, or inference that needs the whole model in one rack. Expect a reservation and a longer lead time.
Choose an eight-GPU server when
The workload fits in eight GPUs. An HGX B200 or B300 server costs less per GPU-hour and needs no liquid cooling. See B200 and B300 rental.
GB200 vs B200: What NVL72 Adds
A B200 and a GB200 NVL72 use the same Blackwell GPU. The difference is how many work together. An HGX B200 server links eight GPUs with NVLink; an NVL72 links 72, with Grace CPUs alongside, so a model can spread across the whole rack as if it were one GPU. The same applies to the B300 and GB300.
| GB200 NVL72 | HGX B200 server | |
|---|---|---|
| GPUs in one NVLink domain | 72 | 8 |
| CPU | NVIDIA Grace, linked directly to the GPUs | Standard x86 server processors |
| Power and cooling | About 120 to 130 kW per rack, liquid only | About 14 kW per server, air-cooled |
| On-demand rate per GPU-hour | From about $10.50 | Median $6.49 |
| Best for | Models too large for eight GPUs; frontier training and inference | Most training, fine-tuning and inference |
NVL72 Rack Power: Why Few Data Centers Can Host It
An NVL72 rack draws more power than a row of ordinary racks, and it needs liquid cooling piped to every GPU. That is why NVL72 capacity is scarce whether you rent it or own it.
From NVIDIA and system vendor specifications; the HGX figure is an MCA estimate from GPU power ratings plus CPUs, memory, networking and fans. Free to cite with a link to this page.
Liquid to every rack
Direct-to-chip cooling with a coolant distribution unit sized for 150 kW or more, and facility water to feed it. Air-cooled halls cannot be retrofitted quickly.
Power per rack, not per floor
A data center with megawatts available may still not deliver 130 kW to a single rack position. Ask for the per-rack figure in writing.
Where it fits
Purpose-built AI halls and liquid-ready colocation. The AI and GPU colocation guide covers which facilities can host 120 kW and up.
Liquid-cooled space for 130 kW racks is the scarcest capacity in the market. Send your rack count, power per rack and date, and you'll hear back within 24 hours.
Rent or Own an NVL72 Rack?
A GB200 NVL72 rack costs about $3.9 million to buy, all-in. Rented on demand at $10.50 per GPU-hour it runs about $552,000 a month, so at full use, owning pays for itself in well under a year before hosting. The catch is hosting: you need a data center that can deliver 130 kW and liquid cooling to a single rack.
MCA calculation from the figures on this page. Excludes power consumption, staff, networking and the residual value of the hardware; liquid-cooled space can price above the average rate used here. Free to cite with a link to this page.
Rent when
The project is under two years, demand is uncertain, or you need capacity before you could find a hall to host it. Renting also leaves you free to move to Vera Rubin without owning a depreciating rack.
Own and colocate when
The racks will run for three years at high use and you can secure liquid-cooled space for 130 kW-plus racks. That space is what we find for clients, along with the terms.
Mix the two
Rent capacity now to start, and own the racks that will run steadily once a hall is ready. Many large AI deployments take this path.
Send the rack count, GB200 or GB300, term and start date. You'll hear back within 24 hours with rented and owned options, and the provider you choose pays us.
How We Get You NVL72 Capacity
You send the requirement
GB200 or GB300, rack count, rent or own, term, start date, and networking and storage. You'll hear back within 24 hours.
We confirm what exists
We check which providers have NVL72 racks with a delivery date, or which data centers can host 130 kW liquid-cooled racks on your timeline, before you commit to anything.
We negotiate the terms
Rate, delivery date, term, upgrade path and exit, benchmarked against the figures on this page. The provider you choose pays us, so it costs you nothing.
GB200 and GB300 NVL72: Frequently Asked Questions
How much does a GB200 NVL72 cost to rent?
The lowest published on-demand rate is about $10.50 per GPU-hour, with Oracle Cloud around $16 and Azure around $27. A full 72-GPU rack works out to about $756 to $1,944 an hour, or roughly $552,000 a month at the lowest rate. Reserved terms typically run 25 to 45 percent below on demand.
How much does a GB200 NVL72 rack cost to buy?
NVIDIA does not publish a price. SemiAnalysis has estimated about $3.1 million bare and $3.9 million all-in, and other estimates run $2 million to $3 million, depending on configuration, networking and volume.
How much does a GB300 NVL72 cost?
There is no published price. Rack estimates mostly fall between $3.7 million and $4 million, with some above $5 million. To rent, GB300 NVL72 capacity is sold by quote and reservation only at every major provider.
Can you rent a GB300 NVL72 by the hour?
Not as of October 2026. GB300 NVL72 capacity is offered through reservations and custom quotes, typically for full racks or multiple racks on multi-month or multi-year terms.
What is the difference between GB200 and GB300?
The GB300 NVL72 uses Blackwell Ultra GPUs with 288 GB of memory each, against about 186 to 192 GB in the GB200, and about one and a half times the FP4 compute per GPU. The rack, NVLink fabric and Grace CPUs are the same, and power rises from about 120 to 130 kW to about 135 kW, with peaks near 155 kW.
GB200 vs B200: what is the difference?
They use the same Blackwell GPU. A B200 server links eight GPUs and is air-cooled; a GB200 NVL72 links 72 GPUs and 36 Grace CPUs in one liquid-cooled rack, so a model can run across the whole rack as if it were one GPU. The NVL72 costs more per GPU-hour and is only worth it for models too large for eight GPUs.
What is NVL72?
NVL72 is NVIDIA's rack-scale design that connects 72 GPUs with fifth-generation NVLink at 130 TB/s, so they share memory and work as a single system. It is used in the GB200 NVL72 and GB300 NVL72.
How much power does a GB200 NVL72 rack use?
About 120 to 130 kW per rack. A GB300 NVL72 draws about 135 kW typically, with peaks near 155 kW depending on workload, so facilities plan for 140 to 160 kW. Both require direct-to-chip liquid cooling.
Can I rent part of an NVL72 rack?
At some clouds, yes: GB200 NVL72 capacity is offered in instances of four GPUs. Training at full scale normally uses whole racks, and GB300 NVL72 is generally sold by the rack.
Is it cheaper to rent or own an NVL72 rack?
At high utilization over three years, owning is far cheaper: about $4.9 million to own and colocate a GB200 NVL72, against about $12.9 million to rent it reserved and about $19.9 million on demand. Renting wins for shorter projects, uncertain demand, or when you cannot find a data center to host it.
Which data centers can host a GB200 or GB300 NVL72?
Facilities with direct-to-chip liquid cooling and at least 130 kW available to a single rack position. Most existing colocation was built for 5 to 15 kW per rack, so suitable space is scarce and often has to be reserved ahead of delivery.
How long does it take to get NVL72 capacity?
It varies widely by provider and quarter. Reserved racks and liquid-cooled colocation are often committed months ahead, so ask any provider for a named delivery date in writing before you sign.
What comes after the GB300 NVL72?
NVIDIA's next generation, Vera Rubin, follows Blackwell Ultra. That is a reason to negotiate an upgrade path, or a shorter term with a renewal option, on any long NVL72 commitment signed now.
Do you help with buying and hosting NVL72 racks, not just renting?
Yes. We find liquid-cooled colocation for owned NVL72 racks, as well as rented capacity, and negotiate the terms either way. The provider you choose pays us.
What does it cost to use Metro Colo Advisory?
Nothing. The provider you choose pays us from its channel budget, and the rate is not marked up. If the better answer is an eight-GPU server or the provider you already use, we say so.
Price Your NVL72 Requirement
Tell us GB200 or GB300, how many racks, whether you want to rent or own, the term and the start date. You'll hear back within 24 hours. If the better answer is an eight-GPU server or the provider you already use, we say so. No cost, no obligation, and the provider you choose pays us.
Once you have contacted a provider directly, it usually will not work through an advisor on your deal.
GPU rental: the GPU rental hub, GPU cluster rental, B200 and B300 rental, H200 rental, H100 rental, neocloud providers, CoreWeave competitors and bare metal.
Owning and hosting: AI and GPU colocation, wholesale colocation and data center cost.