AI Cooling Gets the Reference Design Treatment

On November 10, Daikin Applied published Reference Architecture D60-W, a 60-megawatt cooling blueprint for AI factories based on NVIDIA's DSX platform with Vera Rubin NVL72 S1. Cooling planning, Daikin notes, often means coordinating rack requirements, heat rejection and controls across multiple teams and suppliers. D60-W packages those elements into one documented design — a starting point for engineering cooling from rack to ambient — sold the way silicon vendors sell reference designs. Just note what it is not: a framework for project engineering, not shipping hardware.

What the Blueprint Specifies

The headline numbers: a 60 MW IT load across six deployable units, with liquid cooling handling 92% of the load and air the remaining 8% S1. Two temperature loops split the plumbing: a high-temperature liquid loop serving the coolant distribution units at the racks, and a low-temperature loop for air-cooled rack containment, a split Daikin credits with improving plant efficiency.

Operation is mode-based: D60-W switches between dry cooling, hybrid and full mechanical cooling by outdoor conditions; compressors are off in dry mode, while hybrid mode's waterside economizers precool return water and chillers cover the remainder. Dry coolers carry the load for most of the year in suitable climates S1 — the qualifier matters, because "compressors off most of the year" is a climate claim, not a universal one. Dry coolers instead of evaporative towers also eliminate cooling-tower make-up water and support low water-usage effectiveness.

The plant runs on Daikin's oil-free magnetic-bearing chillers, up to 10 MW each — an operational detail, not green theater: conventional chillers need oil management and periodic runs even in standby, and oil-related startup delays slow recovery from standby, which Daikin ties directly to protecting GPU availability S1. Recovery time from standby belongs on every cooling RFP; it shares a critical path with the GPUs it serves.

A Blueprint Today, Its Hardware in 2027

The honest caveat, in Daikin's own terms: the company's CDU and data-center dry cooler are currently under development, with availability targeted for the second half of 2027 S1. What customers get today is the design discipline — documented dependencies between rack water temperatures, flows, plant operation and heat rejection, plus a supervisory layer coordinating racks, chillers and dry coolers as one plant. It is the reference-design pattern applied to facilities: standardize interfaces first, let projects engineer against them, let productized hardware land on that foundation. Whether it compresses cooling timelines the way silicon reference designs did is the open question.

Demand Is Not Waiting

India's National Stock Exchange has deployed direct-to-chip liquid cooling in its on-premises data center — what it describes as the first such deployment in Indian capital markets — designed to accommodate up to 224 GPUs, with more than 80 planned this quarter and a colocation extension planned soon S2. The stated motive is sovereignty and control.

NVIDIA pitches its Vera CPU in the same per-watt terms: in a 100-megawatt data-center comparison, it claims 1.6x more agents per watt than the fastest x86 CPU and 2x faster agent speed than the most energy-efficient x86 part S3. These are vendor benchmarks, not field data — but the axis is the point. Cooling blueprints and agents-per-watt claims are two ends of one calculation: AI capacity is now planned as completed work per watt.

What Operators Should Take

• Read D60-W as a dependency checklist to adapt to your site, not a turnkey plant.
• Ask what ships today versus later: the chillers exist, the CDU and dry cooler target H2 2027.
• Treat economizer claims as climate-dependent, and demand workload-matched per-watt data before accepting vendor benchmarks.