GB300 is NVIDIA's current top system. Vera Rubin is the generation that follows it. For the building that hosts them, two things change: Rubin packs more power into the same space, and it can't be cooled with air at any size.
The short version
- Same shape. Both put 72 GPUs in a single rack.
- More power. A GB300 rack draws up to 142 kW. NVIDIA hasn't published the figure for a Vera Rubin rack. Industry reporting puts it between 190 and 230 kW.
- Liquid everywhere. A GB300 rack is liquid cooled, with air for its smaller parts. Rubin is liquid cooled in every form, including its 8-GPU server.
- Warmer water. The Vera Rubin rack runs on warm water, which lets a building do without chillers.
Power: more in the same space
A GB300 rack draws up to 142 kW by NVIDIA's reference design. For the Vera Rubin rack, the reported range of 190 to 230 kW is about a third more at the low end and about 60 percent more at the high end.
In round numbers: ten GB300 racks draw up to about 1.4 MW. Ten Vera Rubin racks, on the reported range, would draw 1.9 to 2.3 MW. Same floor space, same number of GPUs.
Those figures are the racks alone. The building's own cooling and power equipment comes on top.
Cooling: liquid everywhere, and warmer
GB300 is cooled with liquid on its hottest parts, the GPUs and CPUs. The rest of the rack is still cooled with air, so a GB300 hall needs both.
Rubin goes further in two ways.
First, there is no air-cooled Rubin. Even the 8-GPU server, DGX Rubin NVL8, is what NVIDIA calls "a liquid-cooled AI system." Its Blackwell counterpart, the B300 server, runs on air.
Second, the Vera Rubin rack is designed for warm water: 45°C going in. That lets a building do without chillers, the large refrigeration machines that colder systems need.
The smaller option: Rubin NVL8
Not every deployment is a full rack. NVIDIA also announced Rubin as an 8-GPU server, the NVL8 family.
NVIDIA's early figure for DGX Rubin NVL8 is about 24 kW per server. A B300 server with the same number of GPUs draws about 14 to 15 kW. So ten Rubin servers need about 240 kW, against 140 to 150 kW for ten B300 servers, and they need liquid cooling.
What it means when you choose a site
- A site that hosts GB300 today isn't automatically ready for Rubin. Ask how much power each rack position can take, and whether that can grow.
- A site that hosts B300 servers on air can't host Rubin without adding liquid cooling.
- Ask about the water. Rubin is designed for warm water. Ask what temperature the site supplies and whether liquid cooling is in place or planned.
- Think past the first hardware. Minimum terms on the capacity we broker run 36 to 60 months. Ask now what the site can take next.
What NVIDIA has not published
Be careful with numbers for Vera Rubin. NVIDIA has not published the power of the Vera Rubin rack, the power of a single Rubin GPU, or a list price. The figures you see for those come from industry reporting. Ask for the rated power in writing before you size a site around it.
One naming note: you may still see the Vera Rubin rack called NVL144. NVIDIA renamed it NVL72 in December 2025. It's the same rack.
Questions to ask a data center
- How many kW can each rack position take today, and how far can that be raised?
- Is liquid cooling in place, and what water temperature does it supply?
- For 8-GPU servers, can you host liquid-cooled servers, or only air-cooled ones?
- Is the site built to Tier III standards? NVIDIA's reference design for GB300 clusters calls for Tier 3 design standards.
- What's the minimum term?
Where XIRR fits
XIRR Advisors brokers two things: Tier III colocation space and power for hardware you own, and reserved GPU compute capacity (H100/H200/B200/B300/GB200/GB300/Vera Rubin) from neocloud operators. We do not sell GPUs or hardware.
If you're planning for Rubin, tell us the cluster size, the market and the timing, and whether the site has to take GB300 first.
For more on cooling, see Liquid vs Air Cooling for AI GPUs: Rack Density Explained. For the chips themselves, see From Hopper to Rubin: A Plain-English Guide to NVIDIA Data Center GPUs.
Sources
NVIDIA product pages for GB300 NVL72, Vera Rubin NVL72, DGX B300 and DGX Rubin NVL8. NVIDIA Enterprise Reference Architecture for GB300 NVL72. NVIDIA DGX SuperPOD reference architecture for DGX GB300. The Vera Rubin rack power range is from industry reporting, not from NVIDIA.
Share your requirements. We'll canvas the market.
Tell us what you need (for colocation space and power: market, kW or MW, cooling and timing; for reserved GPU capacity: region, GPU type, cluster size and timing) and we'll canvas colocation and neocloud operators in parallel. Shortlist in 48 hours. We don't sell hardware. We broker colocation space and power, and reserved GPU compute capacity.
Earlier conversations get better terms. When you engage early, we have time to negotiate with vendors before you need to commit. We're paid only when a deal closes, and the fee structure is agreed in writing before any introduction.