Air versus direct-liquid cooling for Blackwell systems.

The cooling decision on a Blackwell buildout is not a thermal engineering footnote. It is the first fork in the road, and it determines almost everything that follows: which OEM systems you can buy, how dense your racks can run, what your facility has to be able to do on the day of delivery, and how long the whole thing takes to land. Get this fork wrong and you discover, weeks into a procurement cycle, that the system you contracted for cannot be cooled in the hall you reserved, or that the colo you signed with cannot accept the system you actually need.

This is a facilities-first explainer. The thesis is simple. Air-cooled HGX B200 and B300 systems are real, deliverable, and appropriate for a large class of deployments, but they cap practical rack density. The moment you move to NVL36 or NVL72 rack-scale Blackwell, direct-liquid cooling stops being a choice and becomes a precondition, because the power per rack leaves the air-cooling envelope entirely. Understanding where that line sits, and what crossing it costs in OEM options and delivery time, is the difference between a clean buildout and a stalled one.

The physics that draws the line

Air cooling works by moving heat into a fluid (air) and carrying it away. The limit is not subtle. Air is a poor heat-transfer medium, and a rack can only accept so much airflow before fan power, acoustic limits, and hot-aisle recirculation defeat the design. The practical ceiling for air cooling sits around 20 kW per rack, and most legacy halls were optimized for 8 to 12 kW racks. Above roughly 20 kW, air has no viable answer for sustained AI density.

Direct-liquid cooling replaces air at the hottest surfaces with a coolant that contacts cold plates mounted directly on the GPUs and other high-power components. Water in motion across a chip surface conducts heat on the order of 300 times faster than air. That single fact is why the industry crossed to liquid at the top of the Blackwell line, and why it had no choice.

Now place the systems on that scale. An air-cooled HGX B200 or B300 node draws on the order of 10 to 14 kW for the GPU complex, and a hall can run one such node per rack, sometimes two with careful airflow design, while staying inside the air envelope. A GB200 NVL72 rack is a different animal. NVIDIA presents the GB200 NVL72 as a single rack-scale, liquid-cooled design: 72 Blackwell GPUs joined into one 72-GPU NVLink domain that behaves as a single massive GPU, delivering 130 TB/s of low-latency GPU-to-GPU communication. That rack pulls 120 to 130 kW total. There is no air-cooling capacity that removes 120 kW from a single rack. NVIDIA ships no air-cooled variant of the NVL72, and that is not a market-segmentation decision. It is thermodynamics.

So the line is bright. Below roughly 20 kW of rack power, air is on the table. Above it, you are in liquid territory, and the NVL rack-scale systems sit far above it.

What is available air-cooled, and what is not

The cooling choice is not uniform across the Blackwell line. It varies system by system, and the variation is exactly what gates your OEM short list.

HGX B200 and B300: available both ways

The HGX 8-GPU server is the workhorse, and it is sold in both air-cooled and liquid-cooled forms. Supermicro's Blackwell catalog lists air-cooled and liquid-cooled HGX B200 and B300 systems side by side, alongside the liquid-cooled-only GB200 NVL72. Concretely, an air-cooled HGX B200 ships as systems like the Supermicro SYS-A22GA-NBRT, the Gigabyte G894-AD1-AAX5, the Dell PowerEdge XE9680L, the HPE ProLiant Compute XD685, and the Lenovo SR680a V3, typically a 180 GB-per-GPU HGX B200 NVL8 board paired with dual Intel Xeon 6980P (Granite Rapids AP) or AMD EPYC Turin head nodes, ConnectX-7 (400G NDR) or ConnectX-8 (800G) NICs, and Micron 9550 PRO NVMe.

The B300 generation ships the same two ways but pushes the density much further on the liquid side. Supermicro offers the B300 as air-cooled HGX 8-GPU servers and as liquid-cooled ORV3 racks, with a 48U ORV3 rack accommodating up to 144 B300 GPUs. Its DLC-2 liquid-cooled HGX B300 line spans a 4U system for standard 19-inch racks (up to 64 GPUs per rack) and a 2-OU OCP system for 21-inch Open Rack V3 (up to 144 GPUs per rack), with DLC-2 capturing up to 98 percent of system heat and delivering up to 40 percent data center power savings. The takeaway for a facilities team: you can buy a B300 air-cooled, but you leave most of its density on the table. The high-power B300 envelope only opens up under liquid.

NVL36 and NVL72: direct-liquid only

The rack-scale systems are not optional on cooling. The GB200 NVL72 and GB300 NVL72, and their half-rack NVL36 siblings, are direct-liquid-cooled designs with no air-cooled variant. The whole point of an NVL rack is to fuse dozens of GPUs into one NVLink5 domain at a power density that only liquid can serve. If your workload wants a single coherent 36-GPU or 72-GPU domain, you are committing to liquid the moment you choose the topology. There is no air-cooled fallback to retreat to.

This is the cascade in one sentence: the cooling method is set by the system class, and the system class is set by the density and domain size your workload needs.

How cooling gates the OEM short list

Here is the part facilities teams underweight. The cooling decision does not just constrain the racks. It constrains who you can buy from at all, because it runs straight through your colo or hall.

A datacenter that can only reject heat to air can host air-cooled HGX systems and nothing else. The instant you need a B300 at full density or any NVL rack, you need a facility that can accept liquid: a secondary coolant loop, a coolant distribution unit, in-rack manifolds, and a way to move the captured heat out of the building. If your provider cannot do that, every DLC and NVL system disappears from your menu. A provider that cannot accept DLC effectively halves your OEM short list and removes the entire rack-scale tier.

The practical reseller-grade Blackwell market splits cleanly along this line:

Read that table as a procurement filter. If your facility is air-only, you are confined to the top row. Confirm your facility's liquid readiness before you shortlist OEMs, not after, because the order is the reverse of how most teams run it. They pick the system, then discover the facility cannot host it. The catalog of standardized Rillor SKUs at the SKU index is organized so you can filter by exactly this constraint, and live contract activity across both air and DLC systems is visible on the marketplace.

The facility checklist for a DLC system

If you are contracting a DLC or NVL system, there is a specific set of facility-side requirements you must confirm before you sign, not after delivery. These are the items that straddle the IT and facilities boundary, and the coordination across that boundary is precisely what makes DLC builds take longer.

The CDU and the secondary loop

A coolant distribution unit circulates coolant through a closed secondary loop that contacts the cold plates, isolated from the facility's chilled-water supply by a heat exchanger. The choice of CDU is the first decision, and it has two shapes. A liquid-to-liquid CDU rejects heat into facility water and therefore requires facility water infrastructure: supply and return pipes, pumps, and the chilled-water capacity to absorb the load. A liquid-to-air CDU needs no facility water connection but is limited in cooling capacity, which makes it suitable for smaller DLC deployments or rooms that cannot get water plumbed in. For an NVL72 at 120-plus kW, you are almost certainly in liquid-to-liquid territory, which means your facility water plant is on the critical path.

Manifolds and in-rack plumbing

Below the CDU sit the rack manifolds and the in-rack plumbing: the quick-disconnects, the drip-management, the flow balancing across cold plates, and the leak detection. These components straddle the boundary between the IT vendor and the facilities team, and that is exactly why DLC blurs the old separation between the two groups. Someone has to own the secondary loop from the CDU through the manifold to the cold plate and back, and on a multi-rack build that ownership has to be agreed before hardware arrives.

The confirmation list, before you contract

Before executing a forward on a DLC system, confirm with your facility or colo:

• CDU type and capacity (liquid-to-liquid versus liquid-to-air), and whether it is provided by you, the OEM, or the colo.
• Facility water supply and return: flow rate, temperature, and available chilled-water capacity for a liquid-to-liquid CDU.
• Rack format compatibility (19-inch versus 21-inch Open Rack V3) for the system you are buying.
• Manifold and quick-disconnect standard, plus who owns the in-rack plumbing and leak detection.
• Power: a 120-130 kW NVL72 rack needs the electrical feed and busway to match, which is its own long-lead facility item.

If any of these is unconfirmed, the contract delivery month is not yet meaningful, because the facility, not the boards, will gate when the system can actually go live.

Why DLC adds four to six weeks, and what to do about it

DLC builds typically run four to six weeks longer to deliver than air-cooled equivalents. The instinct is to blame the hardware. That is wrong. The boards and cold plates are not the long pole. The delay is facility-side coordination: standing up or validating the CDU and secondary loop, plumbing manifolds, balancing flow, commissioning leak detection, and getting the facilities and IT teams aligned on a loop that neither group historically owned alone. Liquid cooling has moved the IT-facilities boundary, and that organizational seam, not the silicon, is where the weeks go.

This is exactly the kind of timing risk a forward contract is built to absorb. A Rillor contract is a bilateral OTC forward with intent of physical delivery, which the CFTC defines as an agreement between a commercial buyer and seller on delivery of a specified quality and quantity of goods at a specified future date. The operative word is the date. You set the delivery month. For an air-cooled HGX system landing in a ready hall, you might pull the month in. For an NVL72 that depends on a new CDU, a chilled-water tie-in, and a busway upgrade, you set the delivery month to land after those facility milestones, not before, so you are not taking custody of a 120 kW rack into a hall that cannot yet cool or power it.

That alignment is the whole argument for buying DLC capacity forward rather than chasing it on spot. Spot delivery arrives when it arrives. A forward delivery arrives when your facility is ready, against a price you locked when you contracted, with a 10 percent deposit at execution and the balance at delivery, an independent escrow agent holding funds until the system actually changes hands, and a seller performance bond standing behind the delivery promise. The contrast between locking forward and waiting on a queue is laid out in why serious GPU buyers need a forward market, not a waitlist, and the full per-rack power and thermal budgets that drive these milestones are worked through in power and thermal budgets per rack for Blackwell. If you want the system-level differences that ride alongside the cooling choice, B200 versus B300: what actually changes at the system level covers the silicon side of the same fork.

Putting it together for a facilities team

The sequence that keeps a Blackwell buildout clean runs in this order. First, decide the workload's density and domain requirements, because that sets the system class. Second, let the system class set the cooling method: air for standalone HGX at modest density, DLC for high-density B300 and for every NVL rack. Third, confirm your facility can host that cooling method, because if it cannot, your OEM short list collapses to the air-cooled top row. Fourth, run the facility readiness items (CDU, secondary loop, manifolds, water, power) and put real dates on them. Only then, fifth, set the forward delivery month to match those dates.

Done in that order, the cooling decision stops being a surprise and becomes a planning input. The four-to-six-week DLC premium becomes a scheduled, funded part of the project rather than a slip discovered at delivery. And the forward contract does the job it is built for: it holds the price and the system steady while your facility catches up to the rack you are about to put in it.

Buyers planning a DLC or NVL build can map their facility milestones to a contract delivery month with our team. Access to the marketplace is invite only.