Table of contents

A chiller platform takes an HVAC manufacturer three to five years to design, validate, certify, and bring to volume. An AI accelerator generation now lasts about eighteen months. Those two clocks stopped matching somewhere around the GB200 ramp, and every thermal OEM chasing the data center market has felt it since.

HVAC reference designs are the industry’s answer to that mismatch. A reference design is a pre-validated blueprint — thermal envelope, hydraulic definition, control logic, interface drawings, qualified bill of materials — published by a compute vendor, an infrastructure vendor, or an open consortium, that an OEM can build against instead of deriving from first principles. The engineering that would have consumed a year of a team’s capacity arrives as a document.

This guide covers what HVAC reference designs actually contain, which cost lines they remove, which risks they genuinely transfer, where they quietly fail, and how a cooling OEM adopts them without becoming an interchangeable box builder.

Why HVAC Reference Designs Became Essential in 2026

hvac reference designs data center cooling inline 01

Nothing about publishing a validated design is new. Semiconductor vendors have shipped reference boards for decades. What changed is that thermal equipment moved onto a compute release cadence it was never engineered to follow.

The load outran the product cycle

A conventional enterprise rack drew 5 to 15 kW and the cooling behind it changed slowly. An NVIDIA GB200 NVL72 rack draws well over 100 kW, and the GB300 NVL72 designs Schneider Electric released with NVIDIA in late 2025 target up to 142 kW per rack. Each step invalidates assumptions baked into an OEM’s existing hydraulic and heat-rejection platforms. HVAC reference designs exist because no manufacturer can re-derive those assumptions every eighteen months.

Two clock speeds that no longer match

The mismatch is structural rather than a failure of effort. Compute vendors publish a new rack architecture, hyperscalers commit capital against it within a quarter, and the thermal supply chain is expected to have qualified equipment ready. An OEM starting from a blank sheet arrives two generations late. Building against HVAC reference designs compresses the front half of that timeline to weeks.

What an OEM is actually being asked to deliver

The ask is no longer “a chiller.” It is a coolant distribution unit that interoperates with a specific rack manifold, holds a defined approach temperature under a defined flow curve, exposes telemetry a building management system can read, survives the compute vendor’s interoperability testing, and ships inside a lead time set by someone else’s construction schedule. That is a systems-integration brief, and HVAC reference designs are how it becomes tractable.

The market rewarded the shift immediately

Trane, Johnson Controls, Carrier, and Daikin — the four largest building-thermal manufacturers — have all pivoted into AI data center cooling. None of them got there by inventing a liquid cooling architecture in isolation. They got there by aligning to published designs and competing on manufacturing, service, and delivery.

What an HVAC Reference Design Actually Contains

hvac reference designs data center cooling inline 02

The term gets used loosely, so it is worth being precise about what arrives in the package. Serious HVAC reference designs are not marketing diagrams.

Layer What it specifies What it removes for the OEM
Thermal envelope Capacity, approach temperature, supply/return conditions Months of load-case modelling
Hydraulic definition Flow rate, pressure drop, pump curve, redundancy scheme Iterative pump and header sizing
Control strategy PLC logic, setpoints, failover, alarm behaviour Ground-up controls development
Interfaces Manifold, quick-disconnect, piping, electrical, telemetry Integration guesswork with the rack vendor
Bill of materials Qualified pumps, heat exchangers, sensors by part number Component qualification cycles
Compliance evidence Test methodology, rating conditions, safety basis Duplicated certification testing

The validated thermal envelope

This is the core of it. Google’s Project Deschutes CDU specification, contributed to the Open Compute Project as version 1.0 in February 2026, defines a 2 MW unit at a 3°C approach temperature with roughly 500 GPM at 80 to 90 psi. Those are not aspirations; they are validated operating points with test methodology attached. An OEM building to that envelope inherits a proven target.

Hydraulic and control definitions

The Deschutes specification also carries fully redundant power feeds per pump circuit and 0.2 micron side-stream filtration — two choices that look minor and are not. Side-stream filtration at that rating is what protects cold plate microchannels measured in tens of microns. HVAC reference designs encode that kind of hard-won detail, which is exactly the knowledge a first-time entrant lacks.

Interface and integration drawings

Most field failures in liquid cooling occur at interfaces, and interfaces are precisely what a single manufacturer cannot define alone. OCP’s Cooling Environments project frames the problem across five domains — cold plate, CDU, immersion, rear-door heat exchanger, and heat reuse — so the boundaries between them are specified rather than negotiated per project.

Qualified bills of material

A named pump, a named heat exchanger, a named sensor with known behaviour. This is the least glamorous part of HVAC reference designs and often the most valuable, because component qualification is slow, expensive, and produces no differentiation whatsoever when done independently by twelve manufacturers.

What HVAC reference designs deliberately leave open

A good one stops short of the product. Enclosure design, manufacturing method, serviceability, controls user experience, warranty structure, and regional compliance remain the OEM’s. That boundary is intentional and it is where competitive advantage now lives.

The Cost Structure HVAC Reference Designs Attack

hvac reference designs data center cooling inline 03

Cost reduction here is not about cheaper components. It is about deleting engineering hours that produce no customer-visible value.

Non-recurring engineering is the largest line

A new thermal platform carries NRE across concept modelling, prototype builds, test rig time, controls development, certification, and documentation. For a mid-size OEM this routinely runs into millions before a single unit ships. HVAC reference designs remove the largest single block of it — the architectural derivation — because the architecture arrives already argued and already tested.

Validation and test hours

Test capacity is the bottleneck most OEMs underestimate. Thermal validation needs load banks, instrumented loops, and calendar time nobody has. When HVAC reference designs supply test methodology and rating conditions, as OCP’s liquid-to-liquid CDU test methodology does, the OEM validates against a known procedure rather than inventing and defending one.

Rework discovered in the field

The expensive failures are the ones found after installation. A pressure drop assumption that was wrong, a control response that oscillates against the rack’s pump, telemetry the customer’s BMS cannot parse. Each generates a retrofit programme across an installed base. Designs that were pre-validated against the actual compute platform cut this class of rework sharply.

Sales and specification cost

There is a commercial line too. Selling a bespoke architecture requires convincing a consulting engineer, a colocation operator, and a compute vendor independently. Selling a unit built to recognised HVAC reference designs shortens all three conversations, because the reviewer is checking conformance rather than assessing novelty.

Risk Transfer: What HVAC Reference Designs Actually De-Risk

hvac reference designs data center cooling inline 04

Cost is the headline; risk is the substance. Four categories matter, and HVAC reference designs address them unevenly.

Interoperability risk with the compute vendor

This is the risk that most justifies adoption. NVIDIA’s reference architectures for Blackwell-class systems specify direct liquid cooling as a requirement, and the specificity has grown to cover rack dimensions, cooling conditions, power topology, and simulation-ready models. An OEM that designs outside that envelope has built equipment nobody can deploy. Following HVAC reference designs makes interoperability a checklist item instead of a gamble.

Thermal performance risk

Approach temperature under real load is where datasheets and reality diverge. A validated envelope with published test conditions gives the OEM a defensible performance claim and gives the buyer a comparable number. Both sides benefit from the comparison being possible at all.

Compliance and safety risk

The refrigerant transition made this acute. A2L refrigerants such as R-454B and R-1234ze(E) bring charge limits, leak detection requirements, and ventilation obligations under UL 60335-2-40 and ASHRAE 15. HVAC reference designs that already account for those constraints prevent an OEM from discovering a compliance blocker after tooling.

Supply chain risk

Dry coolers and cooling towers have run at 20 to 30 week lead times, and refrigerant reformulation has concentrated dependence on a small set of component suppliers. A qualified bill of materials that names second sources is a supply chain control, not just an engineering convenience. This is where HVAC reference designs overlap directly with disciplined vendor management.

The risk that does not transfer

Liability stays with the manufacturer. A reference design is guidance, not a warranty. If a unit fails in service, the OEM’s name is on the nameplate and the reference design’s publisher is not a party to the claim. Teams that misread this are the ones that skip their own validation entirely.

HVAC Reference Designs and the AI Compute Roadmap

Alignment to the roadmap is the whole point, and it creates a dependency worth understanding clearly.

The reference architecture cadence sets the pace

Schneider Electric’s EcoStruxure library illustrates the pattern. Reference Design 99 — a 3,818 kW, Tier III, chilled-water design covering both liquid-cooled and air-cooled AI clusters — was built with NVIDIA for GPU cluster deployment, and the library extended to GB300 NVL72 within roughly a year. An OEM tracking that cadence knows what to build before the demand signal arrives.

Retrofit is as important as greenfield

RD99 is instructive because it addresses retrofit scenarios alongside a purpose-built liquid-cooled room. Most cooling revenue over the next several years sits in halls that already exist, where new high-density clusters land beside traditional IT. HVAC reference designs that only describe greenfield builds miss where the work is.

Power and cooling are converging

The 800 VDC architectures NVIDIA is developing with ABB, Eaton, Schneider Electric, and Vertiv — with ecosystem readiness targeted around 2027 and Vertiv’s portfolio slated for the second half of 2026 — include CDUs powered directly from the DC bus. A cooling OEM that treats power distribution as somebody else’s problem will find its next platform stranded.

Chasing the roadmap alone is not a strategy

The dependency runs both ways. An OEM whose entire portfolio maps one-to-one onto one vendor’s reference architectures has outsourced its product strategy. The defensible position is to build to the published envelope while owning the layers the envelope leaves open.

Open Versus Vendor HVAC Reference Designs

Not all HVAC reference designs carry the same terms, and the difference matters commercially.

Dimension Open (OCP) Vendor-published
Access Public specification Often partner-gated
Scope Component and interface level Full facility system
Commercial tie None Aligned to a vendor ecosystem
Change control Community process Vendor roadmap
Best used for Interoperability and interfaces Whole-system deployment

The open path

Project Deschutes was the first CDU specification ever contributed to OCP, and the community publishes cold plate requirements and universal quick disconnect work alongside it. Open HVAC reference designs give an OEM interoperability without a commercial dependency, and they carry documentation from technical design through manufacturing quality and service procedure.

The vendor path

Vendor libraries cover more ground. A full facility reference design specifies power, cooling, white space, and lifecycle software together, which is what an operator actually buys. The trade is alignment to that vendor’s ecosystem and roadmap.

Most OEMs will use both

The practical pattern is open specifications at the interface layer — where standardisation lowers everyone’s cost — and vendor designs at the system layer, where the deployment story is sold. Treating the two as competing choices is a common and expensive error.

Where HVAC Reference Designs Compress Time to Market

Time is the scarcest input in this market, and it compresses in three distinct places.

Concept to prototype

Architectural derivation, load-case modelling, and hydraulic sizing are the slowest early activities and the least differentiating. Starting from a validated envelope moves an OEM to prototype in a fraction of the time, which matters more than the engineering cost saved.

Qualification and interoperability testing

When the target is a published specification, qualification becomes conformance testing. The OEM knows the test conditions in advance and designs to pass them, rather than discovering acceptance criteria during a customer’s evaluation.

The specification cycle

There is a commercial acceleration too. Consulting engineers and operators increasingly write HVAC reference designs into their basis of design. Equipment that conforms is specified in; equipment that requires a bespoke review is deferred. For an OEM, conformance to recognised HVAC reference designs is becoming a condition of being considered at all.

The Limits: What HVAC Reference Designs Do Not Solve

Honest assessment matters, because the failure modes here are predictable and repeated.

HVAC reference designs are not products

They define an architecture. They do not define enclosure, manufacturability, service access, spare parts strategy, firmware quality, or documentation. Teams that treat the document as a near-complete design consistently underestimate the remaining work by a wide margin.

Differentiation compresses

If every competitor builds to the same envelope, thermal performance stops being a selling point. Competition moves to price, lead time, service network, and controls quality. That is a healthier market for buyers and a harder one for manufacturers who have historically competed on efficiency curves.

Site conditions still vary

HVAC reference designs assume conditions. Real sites have different ambient extremes, water availability, electrical characteristics, seismic requirements, acoustic limits, and floor loading. The adaptation work is genuine engineering, and it is where a strong OEM earns its position rather than where it saves effort.

Field reliability data is still thin

High-density liquid-cooled deployments are young. The failure distributions needed for spares planning and warranty reserving do not yet exist at scale, and no set of HVAC reference designs supplies them. Operators running these estates increasingly cover the gap with structured data center operations practice rather than vendor documentation.

Publication lag is real

A specification captures a moment. Between publication and the next revision, the compute platform moves. An OEM that treats the document as static rather than as a tracked dependency will ship to a target that has already shifted.

How to Use HVAC Reference Designs Without Becoming a Box Builder

The strategic question is not whether to adopt them. It is what to own once you have.

Compete below the interface

HVAC reference designs define what crosses the boundary — flow, temperature, pressure, telemetry, physical connection. It rarely dictates how you achieve it. Heat exchanger selection, pump control strategy, acoustic design, and manufacturing cost are all still yours to win on.

Own the controls and telemetry layer

Control quality is the most under-exploited differentiator in liquid cooling. Failover that has actually been tested under load, alarm logic that distinguishes a sensor fault from a flow fault, and telemetry that integrates cleanly into building and IT monitoring are all high-value and only loosely specified. Pairing that data with predictive analytics turns a thermal box into a serviceable platform.

Make serviceability the moat

A CDU that can be maintained without draining a loop, with parts available regionally and a technician network trained on it, wins repeat business that a marginally better approach temperature never will. HVAC reference designs say little about this, which is exactly why it is defensible.

Treat conformance as table stakes, not achievement

Conformance gets an OEM onto the shortlist. Delivery, service, and total cost decide the award. Manufacturers that celebrate conformance as the finish line tend to lose the second order.

Metrics That Show HVAC Reference Designs Are Working

Metric What it tells you Target direction
Engineering hours per platform NRE actually avoided Falling generation over generation
Concept-to-prototype time Front-end compression Months, not quarters
First-pass qualification rate Design fidelity to the specification Toward 100 percent
Field rework rate per unit shipped Interoperability and validation quality Trending to zero
Specification-in rate Commercial effect of conformance Rising share of bids
Component second-source coverage Supply chain resilience Every critical part dual-sourced
Lead time variance Delivery credibility Narrowing

Measure engineering hours before and after

The claim that HVAC reference designs reduce cost is testable. Track engineering hours per platform across generations. If the number is not falling, the organisation is re-deriving work the specification already provided, usually because internal review culture does not trust external validation.

Watch first-pass qualification

A conformance failure at qualification means the design drifted from the specification somewhere. It is a cheap signal that catches expensive divergence early, and it costs nothing to collect.

Common Mistakes OEMs Make With HVAC Reference Designs

Treating the document as a complete design

It is an architecture and a set of constraints. Product engineering, manufacturability, and service design remain in full. This is the single most common misjudgement.

Skipping independent validation

Inherited validation covers the reference configuration, not your implementation of it. Liability does not transfer with the document, and no publisher will stand behind a unit they did not build.

Adopting one ecosystem exclusively

Building solely to a single vendor’s designs concentrates commercial risk in that vendor’s roadmap. Interface-level open specifications hedge it at low cost.

Ignoring the compliance overlay

A design validated in one region may not satisfy local refrigerant charge limits, leak detection rules, or pressure vessel codes. Compliance is regional and stays the manufacturer’s obligation.

Failing to track revisions

Specifications change. Without an owner assigned to monitoring them, an OEM discovers the delta during a customer’s technical review — the most expensive possible moment.

Letting engineering treat adoption as a threat

Resistance to external designs is often cultural rather than technical. The work does not disappear; it moves to the layers that actually differentiate the product. Leaders who fail to explain that shift get quiet non-adoption instead of open disagreement.

A Practical Adoption Path for HVAC OEMs

Days 1–30: establish the gap

Inventory current platforms against the HVAC reference designs relevant to your target segment. Record where each product already conforms, where it deviates, and whether the deviation is deliberate. Identify which compute and infrastructure ecosystems your customers actually buy. The output is a one-page conformance picture per product line.

Days 31–60: close the cheapest gaps

Adopt the interface specifications first — connections, telemetry, test methodology — because they cost least and unlock the most. Qualify second sources for components the reference bill of materials names. Run your existing platform against the published test methodology and record the honest result rather than the marketed one.

Days 61–90: restructure where value is created

Reassign engineering capacity from architectural derivation to controls, serviceability, and manufacturing cost. Assign a named owner for specification tracking. Put conformance evidence into the sales package. Where a platform cannot conform without a redesign, decide explicitly whether to invest or exit that segment, with the cost of continued non-conformance stated in the business case alongside broader cost optimization targets.

Frequently Asked Questions

What are HVAC reference designs in data center cooling?

They are pre-validated blueprints defining how a cooling system should be architected — thermal envelope, hydraulics, controls, interfaces, and qualified components — published by compute vendors, infrastructure vendors, or open consortia such as OCP, so manufacturers build against a proven target instead of deriving one.

How much engineering cost do they actually remove?

The largest saving is architectural derivation and validation, typically the biggest block of non-recurring engineering on a new thermal platform. Component qualification and test methodology development are the next largest. Product engineering, manufacturing, and certification costs remain.

Do HVAC reference designs transfer liability to the publisher?

No. They are guidance. The manufacturer remains responsible for its own product, its validation, and its regional compliance. Independent testing of your specific implementation is still required.

Should an OEM use open or vendor-published designs?

Most use both — open specifications at the interface and component layer for interoperability, vendor designs at the facility system level where the deployment is sold. Committing exclusively to one vendor’s ecosystem concentrates commercial risk.

Do HVAC reference designs eliminate product differentiation?

They compress it in thermal performance and move it elsewhere: controls quality, telemetry, serviceability, lead time, manufacturing cost, and service network. Those are harder to copy than an efficiency curve, so the shift favours operationally strong manufacturers.

How fast do these specifications change?

Roughly on the compute cadence — meaningful revisions every twelve to eighteen months, with the underlying rack architectures moving faster than traditional HVAC product cycles. Tracking revisions needs a named owner, not an occasional review.

Conclusion: Build the Architecture Once, Compete Everywhere Else

HVAC reference designs did not appear because thermal engineers stopped being capable. They appeared because the economics of re-deriving the same architecture twelve times across twelve manufacturers, every eighteen months, stopped making sense against a compute roadmap moving this quickly. The architecture became shared infrastructure, in the same way interface standards did a generation earlier.

For an OEM the decision is no longer whether to adopt them. Buyers are writing them into basis-of-design documents and compute vendors are treating conformance as an entry condition. The real decision is what you do with the engineering capacity that adoption frees, and the manufacturers gaining share are the ones redeploying it into controls, serviceability, supply resilience, and delivery rather than banking it as a cost saving.

The move to make this quarter

Take your highest-volume cooling platform and test it against the published methodology for its category. Record where it conforms and where it does not, without softening the result. That single exercise usually reveals more about competitive position than a year of market analysis, and it costs a test rig and two weeks. For context on why density forced this transition at all, see our analysis of liquid cooling versus air cooling for AI workloads, and for the open specifications themselves see the Open Compute Project’s coolant distribution unit work.