Liquid Cooling for Blackwell: What B200 and GB200 Require

For most of the GPU era, air cooling was the default and liquid was the exception. Blackwell changes that calculus. As per-GPU power climbs and racks grow denser, air cooling runs out of headroom — and for the densest Blackwell systems, direct-to-chip liquid cooling stops being an optimization and becomes a requirement. If you are planning B200 or GB200 deployments, cooling is no longer an afterthought to the GPU choice; it is part of it.

Why air runs out of room

Air has a fixed ability to carry heat away from a chip, and that ceiling is reached well before Blackwell's densest configurations stop generating it. The B200 carries a higher TDP than Hopper parts, and packing eight into an HGX node or 72 into a GB200 NVL72 concentrates heat beyond what airflow can practically remove. Past a certain rack power, liquid is not a preference — it is the only way to keep the silicon in its thermal envelope without throttling.

How direct-to-chip cooling works

Direct-to-chip (DLC) liquid cooling routes coolant through cold plates mounted directly on the GPUs and other hot components. Heat transfers into the liquid and is carried out of the rack to a coolant distribution unit (CDU), which exchanges it into a facility loop. Liquid moves far more heat per unit volume than air, so DLC sustains high-power GPUs at full clocks without the acoustic and airflow limits of fans — and it does so more energy-efficiently, improving data-center PUE.

What your facility must provide

• Coolant distribution units sized to the rack's heat load, with redundancy appropriate to your uptime needs.
• A facility water loop or heat-rejection path, and the plumbing to connect it to the CDU.
• Leak detection, serviceable quick-disconnect fittings, and trained staff for liquid-cooled hardware.
• Power delivery and structural capacity matched to dense racks that can reach 120 kilowatts or more.

Air-cooled Blackwell still exists

Not every Blackwell deployment demands liquid. Air-cooled HGX B200 configurations exist for facilities not ready for DLC, typically at lower density or with power-capped operation that trades some performance for compatibility with an existing hall. That can be a sensible bridge. But GB200 NVL72 is liquid-cooled by design, and high-density B200 increasingly assumes DLC — so if your roadmap points at scale, plan the cooling investment alongside the compute.

Design cooling and compute together

The expensive mistake is buying GPUs first and discovering the facility cannot cool them. Cooling type, rack power, CDU capacity, and GPU selection are one decision, not four. Getting them right together is what lets Blackwell run at the performance you paid for, reliably, over its service life.

Nexus Compute configures both air-cooled and direct-liquid-cooled Blackwell systems, validates them against your facility's real thermal and power capacity, and warranty-backs the result. Share your data-center constraints and we will return a tested configuration and quote within 48 business hours.