NVIDIA's Certified CDU List Decoded: The Qualified Map of High-Density Liquid Cooling

Single-GPU power has crossed 1kW and single-rack heat density can exceed 100kW, exhausting air cooling headroom and making liquid cooling the default for high-density AI data centers. In this architecture, the data center liquid cooling CDU (Coolant Distribution Unit) moves chip heat from the secondary loop to the primary cooling system, using pumps, heat exchangers, valves and controls to keep circulation stable while monitoring leaks and temperature in real time. Its reliability decides whether compute runs steady and efficient. Since NVIDIA's influence largely defines the core supply base, this article decodes its officially recognized CDU vendor and model list (September 2, 2026 version): 19 vendors, 28 models, plus selection guidance for 100MW-class GB200/GB300 data centers.

Above 100kW per rack, no air-side supplement, including the liquid cooling CDU rear door heat exchanger, can bridge the gap, so direct-to-chip (D2C) cooling is the only path. A D2C system has three parts: the cold plate on the chip, the manifold distributing coolant to each rack, and the CDU handling heat exchange and circulation. Inside the CDU sit a heat exchanger, pump, filter, valves and controls that pass secondary-side heat to the primary side, drive circulation at precise flow and pressure, and monitor leaks, pressure and temperature to protect expensive equipment. The CDU holds about 25% of a liquid cooling system's BOM cost, and NVIDIA-class certification takes over 12 months, a double barrier that makes it the highest-value, least-settled link in the chain.

In a liquid-cooled build, heat flows from the cold plate through the data center liquid cooling manifold into the CDU, so qualification validates the whole loop, not a single box. NVIDIA's system is multi-layered: NPN signals compatibility, RVL means the product passed performance tests and can be selected by ecosystem players, and AVL, the highest threshold, means it passed strict quality and capacity audits and can take orders directly. Above all sits product-level "NVIDIA Product Qualified." The self-certification guide (DA-12515) requires vendors to pass standardized tests covering hydraulic performance, sensor accuracy, cold start, thermal loads, pump and fan switchover, group control and pumping capability, and every model on the list carries this validation status.

The list is a test-verified roster, not a cooperation slide, and it is the de facto benchmark of AI data center thermal management. Of the 28 models, 24 (86%) are liquid-to-liquid (L2L) and only four are liquid-to-air (L2A), all capped at 45kW to 130kW, confirming that high-density AI has fully shifted to L2L. Cooling capacity spans 45kW to 3.6MW, roughly an 80x range from rack level to facility level. Nautilus leads at 3.6MW, LiquidStack (2.5MW) and Envicool (2.4MW) follow, Supermicro, Delta and Envicool hold the 2MW tier, and the 1MW to 1.6MW band is the most crowded, with CoolIT, Motivair, Nidec, MetaGreen, Vertiv, Mitsubishi Electric and MHI. Notably, the highest power comes from newer specialists like Nautilus and LiquidStack, while incumbents enter steadily at 1MW to 1.3MW, so the power ceiling does not belong to the old names.

The 19 vendors span industrial giants and liquid cooling specialists, all chasing the same prize: a proven, high efficiency CDU for data center scale. In North America and Europe, Vertiv, NVIDIA's system-level partner, covers the full XDU spectrum around GB200/GB300; Supermicro offers a 2.2MW row-level CDU; CoolIT, Motivair, LiquidStack and Nautilus push performance; Boyd brings a full cold plate, manifold and CDU stack; and Carrier, JCI and Stulz enter from HVAC. From Japan and Korea, Nidec, MHI, Mitsubishi Electric and LGE extend from motors and heavy industry. Taiwan contributes AVC, Delta (2MW), LITEON and MetaGreen, whose TC1288 sits in the GB200 NVL72 reference design. Envicool is the only mainland China vendor, certified at 2MW and 2.4MW.

Three cross-sections deserve attention. First, the route has tilted decisively: 24 L2L against 4 L2A, all capped at 130kW, so new high-density builds choose L2L while server air liquid CDU unit products serve mainly air-cooled retrofits. Second, 14 models are flagged "Wetted Materials Compatibility: Compatible," meaning the materials contacting coolant have been validated with mainstream fluids, a hidden but decisive factor for long-term reliability and maintenance. Third, test depth varies: AVC covers 10 items including group control, nine vendors cover 9, and some only 5 to 6, with gaps in cold start, group control and pumping, a useful proxy for engineering maturity.

For a 100MW-class GB200/GB300 build, six points apply. Calculate heat load first: at about 130kW per GB200 NVL72 rack and 140kW per GB300 rack, 100MW means roughly 700 to 770 racks, with CDU capacity derived backward from PUE and redundancy. Prefer row-level L2L above 2MW, such as Envicool XCRow (2MW/2.4MW), Delta CDU3000, LiquidStack GigaModular (2.5MW) and Supermicro, with Vertiv XDU1350, CoolIT CHx1500, Motivair MCDU50 and Nidec as mid-band options; AI server air liquid cooling CDU products only matter for retrofits. Deploy N+1 redundancy (two units per eight racks in NVL72 logic) and value group control for smooth failover. Check wetted materials compatibility and secondary-side temperature, since GB300 favors higher supply temperatures. Select within the full cold plate, manifold, CDU, UQD and primary-side chain, preferring full-stack vendors like Envicool, Vertiv and Boyd. Finally, pair a primary supplier with a backup: Vertiv and Motivair for overseas and hyperscale, Envicool, the only mainland option with NPN Tier 1 certification, for domestic substitution. Always recheck against the latest NVIDIA list.

NVIDIA's certified CDU list is a test-verified map of the global high-density liquid cooling supply base: 19 vendors, 28 models, from 45kW rack-level to 3.6MW facility-level, across the US, Europe, Japan, Korea, Taiwan and mainland China. It shows the modular air liquid cooling CDU scaling from a rack accessory into the backbone of an entire facility. For professionals it is a selection map; for investors it marks who can enter the next-generation AI supply chain. Liquid cooling is settled; the question is who converts certification into scale orders that pass the triple test of power, reliability and cost.

To turn certification credentials into real-world large-scale deployment, operators need solution providers capable of balancing power performance, reliable operation and cost competitiveness at scale. AI won't slow down—and neither should your cooling. EXTRCOOL delivers a fully integrated liquid-cooling package that combines CDUs, cold plates, manifolds, and piping into one seamless solution. You can choose from our ready-to-ship standard range or collaborate with us on a custom build for your exact site requirements. The result: superior heat removal, lower power costs, and a smaller environmental footprint—all supported by a track record at gigawatt-scale data centres worldwide. For operations where downtime is not an option, EXTRCOOL is the partner you can depend on. Explore our complete product line or request a tailored quote at: https://www.extrcool.com/