Blackwell Ultra servers are shipping fast. On 3 November 2026, NVIDIA said hyperscalers were installing about 1,000 NVL72 racks every week (Zouic, citing NVIDIA). But buying the hardware is only half the job. In early November, reports said some of Amazon Web Services’ new GB300 racks were waiting in warehouses because the data centers meant to host them did not yet have enough power (Crypto Briefing). The lesson for every buyer, large or small: check your site before you place the order.
Why power and cooling now come first
Each new GPU (graphics processing unit) generation uses more power, and all that power turns into heat. Older server rooms were built for racks of normal servers. Today’s AI systems pack far more power into the same space. If your room cannot deliver the electricity, or remove the heat, the servers cannot run at full speed, or cannot run at all. A server that sits switched off is the most expensive server you own.
Two ways to buy Blackwell Ultra
Most buyers choose between two formats. The first is an 8-GPU server, such as NVIDIA HGX B300 systems from server makers or NVIDIA’s own DGX B300. The second is a full rack system, the GB300 NVL72, which links 72 GPUs and 36 Grace CPUs (central processing units) into one liquid-cooled rack (Lenovo Press).
| Question | DGX B300 (8-GPU server) | GB300 NVL72 (full rack) |
|---|---|---|
| Power | About 14.5–15 kW per server, with estimated peaks near 19–20 kW | About 132–140 kW per rack, with peaks up to 155 kW |
| Cooling | Air-cooled (liquid-cooled HGX B300 options also exist) | Liquid cooling required; about 90% of heat goes to liquid |
| Size and weight | 10 rack units, about 168 kg with power supplies | A complete rack, far heavier than one server |
Sources: NVIDIA DGX B300 User Guide, NVIDIA data center best practices for DGX B300, Lenovo Press, HPE data sheet. kW means kilowatts.
In simple terms: a full NVL72 rack needs roughly ten times the power of one 8-GPU server.
What liquid cooling really means for your site
With direct liquid cooling, water-based coolant flows through metal “cold plates” that sit on the GPUs and CPUs and carry the heat away. In an NVL72 rack, a coolant distribution unit (CDU) moves the coolant and passes the heat to your building’s water system (Lenovo Press). That means pipes, pumps, leak detection and new maintenance routines. Cooling suppliers are moving quickly: in the first week of November 2026 alone, several companies announced new CDUs for high-density AI racks (The Data Center Engineer). Good options exist, but they take time to plan, install and test.
If your site has no liquid cooling yet, air-cooled 8-GPU servers are usually the faster path. Plan carefully, though: NVIDIA’s data center guide lists estimated peaks of about 19–20 kW per DGX B300, well above the average figure, so plan power and cooling for the peak (NVIDIA).
A simple checklist before you order
- Power per rack: Ask your facility team how many kW each rack can safely receive today, with redundancy. Compare it with the peak figure of the exact model you plan to buy.
- Cooling type: Confirm whether the exact model is air-cooled or liquid-cooled. The same GPU family can come in both versions.
- Liquid loop: If you need liquid cooling, confirm who supplies, installs and maintains the CDU and pipework.
- Floor and access: Check that floors, doors and lifts can take the weight and size on the way in.
- Timing: Make sure power and cooling are ready before the hardware arrives, not after.
- Network: Plan cabling and switches at the same time. Our guide to InfiniBand and Ethernet for AI clusters explains the choice.
Key takeaways
- New AI hardware only pays off once it is powered on and properly cooled.
- An 8-GPU Blackwell Ultra server uses about 14.5–15 kW; a GB300 NVL72 rack uses about 132–140 kW.
- Air-cooled 8-GPU servers suit most existing rooms; NVL72 racks need liquid cooling.
- Size power and cooling for peak use, and have them ready before delivery.
Q9 Technology supplies GPUs, servers and networking for AI projects, and can help you match the configuration to your site. Contact our team to discuss your requirements. For more background, read GPU Server Solutions for AI.
