NVIDIA NVL72 Explained: What Operators Need to Deploy and Finance a Rack

Summary

• NVL72 is NVIDIA's rack-scale system joining 72 GPUs and 36 Grace CPUs in one NVLink domain, so the rack is one machine and one purchase.
• Each rack draws roughly 120–140 kW in normal operation, with peaks up to about 155 kW, depending on generation and vendor. Liquid cooling makes the site a gating item.
• NVL72 capacity sells mostly on contract, so lenders finance it against the rack and the offtake contract. USD.AI's largest facility, $128.9 million (September 2026), supports 32 GB200 NVL72 systems.

Introduction

NVL72 has become the default way to buy NVIDIA's Blackwell-generation capacity at scale. It also changes the purchase from individual servers to a whole liquid-cooled rack. That rack needs far more power and cooling than a conventional one. For a GPU cloud operator, the rack is the unit you buy, power, cool, contract and finance.

We wrote this guide from the deals USD.AI reviews, so it follows the operator's side of the decision. It covers what NVL72 is, how GB200 and GB300 differ, what the site must provide and when the rack-scale design beats 8-GPU servers. It then turns to how NVL72 capacity is priced and sold, and how lenders look at an NVL72 deployment.

What NVL72 Is and Why the Rack Is the Unit You Buy

NVL72 is a rack-scale NVIDIA system in which 72 GPUs and 36 Grace CPUs share one NVLink domain. That domain is the group of GPUs joined by NVIDIA's direct GPU-to-GPU interconnect. Every GPU talks to every other GPU at NVLink speed, so the rack behaves as one large machine. NVIDIA ships it in GB200 and GB300 versions, and operators buy, deploy and finance it as a complete rack.

NVIDIA's GB300 NVL72 page describes a "fully liquid-cooled, rack-scale architecture that integrates 72 NVIDIA Blackwell Ultra GPUs and 36 Arm-based NVIDIA Grace CPUs." NVIDIA lists 130 TB/s of NVLink bandwidth and about 20 TB of GPU memory per rack (July 2026). Because all 72 GPUs share one NVLink domain, the rack is the unit of purchase, deployment and collateral. You can't split it into eight-GPU servers and sell or finance them separately.

GB200 vs GB300 NVL72

GB300 NVL72 keeps the GB200 rack format and 72-GPU domain and raises dense FP4 compute from 720 to 1,080 PFLOPS per rack. The GB300 figures below come from our GB300 NVL72 hardware reference, as published by NVIDIA and checked in September 2026. The GB200 figures come from the GB200 NVL72 hardware reference, checked the same month.

Both racks share the same format and domain size. GB300 adds memory and compute per GPU at higher power per GPU, so planning the site is the same exercise with bigger numbers.

NVL72 names the size of the NVLink domain, which is why the name carries across generations. NVIDIA's next generation, Vera Rubin NVL72, pairs 72 Rubin GPUs with 36 Vera CPUs, as our Vera Rubin NVL72 reference page shows.

NVL72 Power and Cooling: What the Site Must Provide

An NVL72 rack draws roughly 120–140 kW in normal operation, with peaks up to about 155 kW, depending on generation and vendor. Power and cooling therefore decide where you can deploy, since one NVL72 rack draws more than ten typical racks.

For an operator, the site is a gating item. For a lender, it is part of the collateral review, because the hardware has to be installed, powered and monitored in a qualifying data center. Financing through USD.AI requires a Tier III or Tier IV facility.

How much power an NVL72 rack draws

Published NVL72 power figures run from about 120 kW to a 155 kW peak, and each one measures something different. NVIDIA's DGX GB200 user guide states that "The rack power consumption is approximately 120kW" (September 2026). The same guide shows eight power shelves rated at 33 kW input each, and that rating describes what the shelves can deliver to the rack.

For GB300, Lenovo's GB300 NVL72 product guide says "Each rack consumes 135 kW TDP; up to 155 kW peak depending on workload and EDP behavior." TDP, or thermal design power, is the sustained load the system is designed around. Cisco's GB300 NVL72 datasheet gives "operating power approx. 132 to 140 kW" (August 2026).

In our view, an operator should provision power and cooling for the peak figure. A site sized to TDP has no headroom when a workload pushes the rack toward 155 kW.

Liquid cooling and the CDU

NVL72 moves most of the rack's heat into liquid, so the site needs a facility water loop and a coolant distribution unit (CDU). The CDU moves coolant between the rack and that loop. NVIDIA calls the GB300 NVL72 "fully liquid-cooled", while Lenovo puts the rack-level split at "about 10% to air and 90% to liquid". Both descriptions hold, and the practical result is that the site still needs some air handling for the components that stay air-cooled.

Lenovo says the water inlet "can be as high as 45°C (113°F)". Required rack flow rises with supply temperature: Lenovo's figures show 59 LPM at 25°C and 177 LPM at 45°C. Warmer supply water needs three times the flow, so the colocation agreement should state both the supply temperature and the flow the facility delivers.

Supermicro's SRS-GB300-NVL72 lists "1x in-rack Supermicro 250kW capacity CDU". Confirm whether your design uses an in-rack CDU like this one or a facility CDU, and who owns and maintains it under the colocation agreement.

NVL72 site checklist

Five site items decide whether an NVL72 rack can go live, and each one maps to a document in the financing file. Work through them with the colocation provider before the purchase order is final.

These items feed the documents an operator prepares for financing: the deployment site and colocation agreement, and evidence of power and cooling readiness. With financing through USD.AI, capital is held in escrow until the servers are installed, burned in and independently verified. A site that can't take the rack therefore holds up funding. Operators that host racks in their own facilities can see how this works in financing for data center operators.

NVL72 Vs 8-GPU HGX Servers: What the Rack-scale Design Buys

In Nebius's MLPerf Inference v6.1 results, 8 GPUs inside a GB300 NVL72 delivered 127,921 tokens per second on gpt-oss 120B, against 108,124 for an 8-GPU HGX B300 system (September 2026). HGX is NVIDIA's eight-GPU baseboard that system builders put into x86 servers. With the GPU count held equal, the gain of about 18% comes from the rack design.

Nebius's MLPerf Training v6.0 results point the same way (June 2026). On Llama 3.1 8B, the NVL72 configuration finished in 64.35 minutes against 72.01 minutes, which Nebius calls "approximately 12% faster".

At full scale, the 72-GPU GB300 NVL72 reached 1,122,490 tokens per second on the same gpt-oss 120B server test. These are one provider's published results on specific benchmarks, and other workloads may show a different gap.

For an operator, the choice follows the offtake customer's workload. NVL72 suits customers running large models across many GPUs. Eight-GPU servers such as NVIDIA's DGX B300 are air-cooled at about 14 kW, so they fit more existing sites; our HGX B300 hardware reference covers the specs. Scale-out networking between racks is its own decision and sits outside this comparison.

NVL72 Economics: Rental Prices, Power Cost and Who Is Deploying

In November 2026, one of the three NVL72 cloud providers we checked posted an on-demand NVL72 price, which is why the offtake contract carries these deals. An offtake contract is the agreement under which a customer buys the operator's compute. Its term, price and the customer's credit quality set the revenue a lender relies on.

What NVL72 capacity rents for

CoreWeave's pricing page lists GB200 NVL72 at $42.00 per hour on demand for an instance of "2 GB200 or GB300 Superchips", or four GPUs (November 2026). That works out to about $10.50 per GPU-hour, while GB300 NVL72 on the same page is "Contact sales". Nebius and Together AI list NVL72 capacity as "Contact us", with no posted price.

With so few posted prices, most NVL72 capacity is priced in negotiated contracts. Contracted revenue is what a lender underwrites, so the offtake contract's term, price and customer carry as much weight as the rack itself.

What a rack costs to power: an illustrative example

In this illustrative example, one GB300 NVL72 rack costs about $114,000 a year to power at the US industrial average, before cooling overhead. The example assumes Lenovo's 135 kW TDP running all year, or 8,760 hours, which comes to about 1,183 MWh.

The US Energy Information Administration puts the US industrial average at 9.64¢/kWh (August 2026). At that rate, 1,183 MWh costs about $114,000 a year in electricity for the IT load. Cooling overhead comes on top through PUE (power usage effectiveness), which is total facility power divided by IT power.

Colocation and power contracts differ from the national average, so run the same calculation with your contracted rate and your facility's PUE. A lender checks that the offtake revenue covers power alongside debt service.

Who is deploying NVL72

Operators beyond the hyperscalers are buying NVL72 racks. Nebul is deploying "more than 1,000 NVIDIA Blackwell Ultra GPUs in NVIDIA GB300 NVL72 systems" in Belgium (October 2026).

IREN delivered Horizon 1 to Microsoft and achieved NVIDIA Exemplar Cloud status on GB300 NVL72 (August 2026). Horizon 1 is the first of four 50 MW deployments, and IREN gave no rack count.

Large debt facilities are funding this build-out too. Nscale closed about $3 billion of delayed-draw term loans in August 2026. That total includes "up to $1.85 billion" to "fund the deployment of NVIDIA GB300 (Blackwell Ultra) and VR200 (Vera Rubin) systems" at about 200 MW. A delayed-draw term loan is one the borrower draws in stages up to a set total, which suits hardware that arrives in batches.

How Lenders Finance an NVL72 Rack

In September 2026, USD.AI announced an asset-backed $128.9 million GPU financing facility supporting "32 NVIDIA GB200 NVL72 systems in British Columbia, Canada", its largest to date. The capacity is contracted "under a multi-year agreement with a blue-chip, investment-grade counterparty". The deal shows the two things lenders finance in an NVL72 deployment: the rack, which is the collateral, and the offtake contract, which repays the loan.

In every financing through USD.AI, a special-purpose vehicle (SPV), a company set up only to hold the deal's assets, holds the GPUs, offtake contracts and revenue accounts. Lenders hold a first-priority claim on the GPUs, contracts and revenue. Capital is held in escrow until the servers are installed, burned in and independently verified. A dedicated servicer manages each loan for its lenders for the full term.

The loan terms for an NVL72 rack are the same as for any deployment:

• Up to 80% loan-to-value. Loans cover up to 80% of the deployment cost, with the operator contributing at least 20% as equity.
• Fixed or floating rates. Loans carry a fixed or a floating rate.
• 3–5-year terms. Terms run 3–5 years with straight-line monthly amortization.
• Debt-service reserve. A three-month debt-service reserve is funded at closing.
• Deal size. Financing starts from about $5 million, with larger facilities available.
• Qualifying sites. Collateral sits in Tier III or Tier IV data centers.

Compute IQ scoring rates every deal on a 100-point scale across the operator, the customer and the contract. The operator input covers experience, financial position and ability to deliver the deployment. The customer input is the credit quality of the business buying the compute, and the contract input is the terms behind the revenue. The score sets financeability, the pricing tier and the loan's size and structure.

For an NVL72 financing request, prepare these documents:

• System configuration and quantity
• Supplier and purchase order
• Deployment site and colocation agreement
• Power and cooling readiness
• Offtake contract or documented on-demand revenue

On-demand revenue doesn't disqualify a deal, and it is priced to reflect its shorter commitment. GPU financing through USD.AI covers new NVL72 deployments and refinances racks already in service.

Conclusion

NVL72 turns GPU buying into a rack decision, with a site that must deliver 120–155 kW per rack, liquid cooling, CDU capacity and the right coolant flow. The offtake contract matters as much as the spec sheet, because it is the revenue that repays the loan.

Financing through USD.AI covers up to 80% of an NVL72 deployment against the rack and the offtake contract, on the same structure as every deal. It already supports 32 GB200 NVL72 systems through USD.AI's largest facility to date. Our page on GPU financing for neoclouds sets out how the financing fits an operator's plan.

Bring your equipment plan and offtake contract to a financing conversation. Talk to USD.AI about your NVL72 deployment.

FAQ

What is NVL72?

NVL72 is NVIDIA's rack-scale system that joins 72 GPUs and 36 Grace CPUs in one NVLink domain, so the whole rack works as a single machine. It is sold and deployed as a complete rack, and GB200 NVL72 and GB300 NVL72 are the current versions.

How many GPUs are in an NVL72 rack?

An NVL72 rack holds 72 NVIDIA GPUs and 36 Grace CPUs. In the GB300 NVL72, the GPUs are Blackwell Ultra, arranged in 18 compute trays of four GPUs each.

How much power does an NVL72 rack use?

An NVL72 rack draws roughly 120–140 kW in normal operation, with peaks up to about 155 kW, depending on generation and vendor. Operators should plan power and cooling for the peak figure.

What is the difference between GB200 NVL72 and GB300 NVL72?

GB200 NVL72 and GB300 NVL72 share the same 72-GPU rack format and NVLink domain. GB300 NVL72 uses Blackwell Ultra GPUs with more memory (about 20 TB vs 13.4 TB per rack) and more dense FP4 compute (1,080 vs 720 PFLOPS). It also draws more power per GPU.

How much of an NVL72 deployment can USD.AI finance?

Financing through USD.AI covers up to 80% of an NVL72 deployment's cost (loan-to-value), with the operator contributing at least 20% as equity. The loan is secured by the rack and underwritten on the offtake contract.