Blackwell deployments are systems, not loose GPUs. The cooling has to match the exact rack the server vendor ships, along with its cold plates, manifold, coolant distribution unit duty, water temperatures, and support agreement.
Reference-architecture suppliers
Vertiv has published a 142 kW GB300 NVL72 power-and-cooling architecture. Schneider Electric and Motivair, Trane and LiquidStack, and other integrated suppliers position portfolios around high-density AI factories.
Direct-liquid specialists
CoolIT publicly maps CHx2000 capacity to GB200 NVL72 racks. Accelsius and ZutaCore concentrate on two-phase cooling, while the specialist cold-plate companies reach customers by getting qualified with the server vendors.
Qualification rule
Treat the server vendor's own configuration and warranty as the deciding document. A cooling supplier saying it supports Blackwell does not mean it is approved for every rack, every GPU generation, or your building's water temperatures.
Method and inclusion criteria
A supplier appears here when it has a public Blackwell-era product, reference architecture, rack mapping, partner announcement, or qualified role in the thermal chain. Generic claims to cool artificial intelligence hardware are not enough. The evidence must connect the company to a named Blackwell platform or to the distribution and facility equipment explicitly designed for those loads. Inclusion is a research lead, not NVIDIA approval. The current qualified configuration from the server vendor overrides every landscape entry.
- Named GB200, GB300, or related Blackwell-era system evidence
- Clear role in server, rack, distribution, plant, or controls
- Public product, architecture, or partner source
- A configuration a buyer can verify with the server vendor
Do not rank unlike roles
A server integrator, cold-plate supplier, coolant distribution manufacturer, and chiller company are not substitutes. This page groups them by role and does not assign one numerical order. The buyer first selects the exact rack and its factory liquid configuration. That decision identifies the approved cold plates and manifolds. Distribution and facility suppliers can then be compared at the temperatures, flow, and heat load that configuration requires.
Lock the rack bill of materials
Record the server maker, rack version, accelerator count, networking, power shelf, firmware, cold plates, manifold, hoses, quick disconnects, fluid, and allowable inlet range. Similar Blackwell names can hide different thermal and hydraulic requirements. A supplier reference to one GB200 NVL72 design should not be carried to GB300 or another original-equipment manufacturer without written confirmation.
Translate rack load into facility duty
State total rack power, heat captured to liquid, residual air, secondary supply and return temperatures, flow, pressure, and maximum transient. Add the number of racks per phase and one-failure requirement. Use those values to rate the distribution units, headers, primary water, and heat rejection. A product mapping based only on dividing nameplate megawatts by nominal rack kilowatts ignores temperatures, redundancy, and pump limits.
Review the complete power and cooling architecture
Blackwell racks combine high electrical and thermal density, so busway, backup power, network cabling, controls, coolant distribution, and room air must be coordinated. Reference architectures such as Vertiv's 142 kilowatt GB300 design can shorten layout work. They remain templates. Each site must adapt utility voltage, redundancy, water temperature, climate, fire rules, and service access, then obtain approval for the assembled design.
Evidence hierarchy
A current server-vendor qualification is strongest for compatibility. A named operating deployment is strongest for field behavior. A witnessed full-load test supports capacity. A reference architecture supports coordination. A product launch or supplier statement supports availability and intent. Preserve those distinctions. No single announcement proves warranty, delivery, and operating performance at once.
Shortlist and acceptance
Mandatory gates are written server support, capacity at the project duty point, required redundancy, material and fluid compatibility, delivery before rack arrival, and service at the site. Score lifecycle energy, water, footprint, controls, expansion, references, and commercial terms after the gates. Acceptance should test full load, one failed component, facility-water drift, control loss, leak isolation, residual air, and the compute throttling sequence.
Keep the list current
Processor platforms and supplier qualifications change faster than facility plants. Recheck approved configurations before purchase and again before shipment. A vendor should move in the landscape when it gains a named qualification, ships a production product, publishes a reference, or retires support. Historic Blackwell marketing should not be treated as current support for a later rack revision.