ASHRAE TC 9.9 Thermal Guidelines for AI Data Center Cooling

Liquid cooling is no longer an edge case in data center design—it is the dominant thermal strategy for every hyperscale rack that holds a B200, GB200, or MI300X. Yet the governing standards document most operators reference, ASHRAE TC 9.9's Thermal Guidelines for Data Processing Environments, was finalized in 2021, before the current generation of AI silicon existed in production deployments. The result is a practitioner community running hardware that materially outpaces its own reference standard. Understanding exactly where TC 9.9 is still authoritative, where it has been quietly superseded by OEM cold-plate specifications, and where the next revision needs to go is no longer an academic question—it is a prerequisite for sound facility design and responsible coolant procurement.

ASHRAE TC 9.9 in 60 Seconds

ASHRAE Technical Committee 9.9, Mission Critical Facilities, Technology Spaces and Electronic Equipment, was established to develop and maintain thermal and environmental guidance for data processing facilities. The committee draws its membership from facility engineers, server OEMs, chipmakers, colocation operators, and cooling system vendors. It does not write building codes. It writes reference documents—principally the Datacom Series of handbooks—that the industry treats as de facto engineering baselines even in the absence of a legal mandate to follow them.

The Datacom Series now spans eight volumes covering thermal guidelines, liquid cooling, best practices, energy efficiency, UPS considerations, and fire suppression. Of these, Thermal Guidelines for Data Processing Environments is the volume that determines equipment class, supply temperature bands, and humidity envelopes. The fifth edition, published in 2021, is the current authoritative release. It expanded the liquid-cooling envelope with the H1 class, updated psychrometric guidance to reflect a broader acceptable humidity range, and absorbed lessons from the hyperscale builds of the late 2010s.

Why do hyperscale operators converge on TC 9.9 guidance even when it carries no legal force? The answer is straightforward: server OEMs write their own thermal specifications by referencing the ASHRAE classes. When a procurement team evaluates a server's thermal compliance, when a colocation operator negotiates SLA temperature bounds with a tenant, when a cooling vendor rates its in-row units, all of those conversations start from the ASHRAE class taxonomy. Diverging from TC 9.9 without documented engineering justification creates warranty exposure, complicates insurance underwriting, and introduces ambiguity in multi-vendor environments where compatibility depends on shared thermal vocabulary. TC 9.9 is the lingua franca of data center thermal engineering, and fluency is not optional for serious practitioners.

The committee itself meets multiple times per year and maintains a continuous balloting process for revisions. Working-group participation is open to dues-paying ASHRAE members and typically requires a 6–12 month onboarding commitment before voting rights are granted. Given that the revision cycle for a full document edition runs three to five years, the gap between cutting-edge silicon and published guidance is structural—not a failure of the committee, but a function of how consensus standards bodies operate.

The Thermal Classes Explained

TC 9.9 organizes IT equipment into environmental classes that define the operating envelope a facility must maintain and the equipment must tolerate. The classification system serves two audiences simultaneously: equipment designers who build to a class, and facility engineers who provision to that class. When both sides use the same taxonomy, interoperability is predictable and contractual disputes have a reference point.

The air-cooled progression runs from A1 through A4. A1 represents traditional enterprise server environments—narrow temperature bands, tight humidity control, and conservative change rates. A4 represents the permissive end of air-cooled design, enabling free-air economization, wider humidity swings, and higher supply temperatures that improve PUE at the cost of reduced thermal margin. The H1 class was introduced to address liquid-cooled equipment, where the thermal transfer mechanism is fundamentally different and the relevant metrics shift from air dry-bulb temperature to coolant supply temperature and flow rate.

The A1-to-A4 progression reflects the industry's gradual acceptance of higher inlet temperatures as economizer hours improved PUE economics. Each step up the ladder trades thermal headroom for energy efficiency and infrastructure simplicity. A4 facilities running 40–45°C supply can use outside air economization in most North American climates for the majority of the year, dramatically reducing mechanical cooling load. The tradeoff is that transient thermal events—a brief spike in compute density, a CRAC failure, a hot aisle containment breach—have less margin before equipment hits throttle thresholds.

H1 does not fit cleanly into the A-class progression because liquid cooling changes the physics. The relevant thermal interface is now the cold plate, not the server inlet. Ambient air conditions matter for the facility mechanical plant and for residual air-cooled components, but the primary thermal pathway is the coolant loop. This distinction has significant implications for how H1 should be specified—and where the 2021 document leaves meaningful gaps that practitioners are filling with OEM guidance.

The H1 Class and AI Cooling Reality

The H1 class as written in the 2021 guidelines uses a nominal coolant supply temperature range of 18–22°C as the baseline design point. This range was calibrated for the liquid-cooled HPC and early AI accelerator workloads of the late 2010s, where conservative coolant temperatures provided comfortable margin against chip junction limits and where energy efficiency optimization of the coolant loop was secondary to compute reliability.

Production deployments of current-generation AI clusters tell a different story. Hyperscale operators running NVIDIA GB200 NVL72 racks are converging on coolant supply temperatures of 28–30°C as the operational steady state. This is not a violation of the cold-plate specification—NVIDIA's engineering documents for the GB200 cold-plate system explicitly bless supply temperatures up to approximately 35°C under controlled flow conditions, with the understanding that the cold plate's internal geometry is optimized for the higher delta-T that results from warmer supply. The 28–30°C operational band that is emerging in production represents a deliberate optimization: warmer coolant supply enables the facility to run water-side economizers or cooling towers without mechanical refrigeration for a larger fraction of the year, improving annualized PUE by measurable amounts in temperate climates.

Engineering Note: Moving coolant supply from 20°C to 28°C can shift a facility's mechanical cooling crossover point by 6–10°C of ambient dry-bulb temperature. In a climate like Dallas or Phoenix, that translates to several hundred additional economizer hours per year—a non-trivial operational cost reduction at scale.

Not all operators are running at the warm end of the permissible band, and that conservatism has legitimate engineering justification. AI training workloads are bursty. A GB200 system running a large language model training job will experience significant thermal transients as batch boundaries pass, as gradient checkpointing kicks in, and as interconnect traffic patterns shift. An operator running coolant supply at 22°C has approximately 8–10°C more margin before a thermal transient drives junction temperatures toward throttle thresholds. For operators prioritizing maximum sustained throughput with zero throttle events—particularly in financial modeling or pharmaceutical simulation workloads where interrupted training runs carry direct cost—the conservative coolant band is a reasonable choice even when it carries a PUE penalty.

The practical outcome is a bifurcated market. Hyperscale cloud operators and colocation providers serving AI training workloads are pushing toward the warm end of the coolant envelope, running 28–32°C supply with active monitoring and coolant chemistry programs that account for the higher operating temperatures. Enterprise and research operators with smaller deployments and lower risk tolerance are staying closer to the 2021 guideline's nominal 18–22°C range. Both approaches are defensible, but the 2021 document does not adequately describe the tradeoffs between them—leaving facilities teams to make those decisions based on OEM application notes and peer-operator experience rather than a single authoritative reference.

Operations Warning: Running coolant supply above 28°C without verified corrosion inhibitor reserve alkalinity and conductivity monitoring creates compounding risk. Higher temperatures accelerate inhibitor depletion and galvanic corrosion rates, particularly in mixed-metal systems with aluminum cold plates and copper manifolds. Chemistry monitoring cadence should increase proportionally with coolant temperature.

Where the 2021 Guidelines Fall Short

The fifth edition of the TC 9.9 thermal guidelines was finalized and published before B200 and GB200 systems existed as production silicon. This is not a criticism of the committee—it is a statement of timing. The ASHRAE balloting and publication process for a full document revision takes three to five years, and the H1 class as drafted in 2021 was forward-looking relative to the hardware available at the time. It anticipated liquid-cooled high-density environments with substantially higher power densities than the A-class envelope could support. What it could not anticipate was the specific thermal characteristics, cold-plate geometries, and system-level flow requirements of a 120 kW NVL72 pod.

"The 2021 H1 class was a meaningful step forward for the industry," said Andre Taki, Lead Product Specialist and Practice Leader, Cooling Chemistry at Alliance Chemical. "But the thermal density assumptions baked into that document were roughly 40–60 kW per rack. We are now deploying production AI infrastructure at two to three times that density, and the coolant-side specifications that the H1 class either leaves unspecified or underspecifies—flow rates, reserve alkalinity, particulate contamination limits—are exactly the parameters that determine whether a liquid cooling system survives a multi-year operational life in an AI cluster. The gap is real, it is material, and operators should not mistake the absence of a specification for the absence of a requirement."

The H1 shortfall is not evenly distributed across all parameters. The supply temperature band, while calibrated conservatively, at least acknowledges that liquid-cooled systems have fundamentally different thermal dynamics than air-cooled equipment. The gap is most acute on the coolant-side chemistry and flow specifications. TC 9.9 does not specify a minimum flow rate per unit of heat load, does not set a reserve alkalinity floor, does not define a conductivity ceiling, and does not specify particulate filtration requirements for the coolant stream. These omissions were acceptable in 2021 when liquid cooling was a specialty application with dedicated engineering support. They are less acceptable in 2026 when liquid cooling is the baseline architecture for every AI GPU deployment at scale.

The standards body is aware of the gap. Working-group activity within TC 9.9 has accelerated since the GB200 platform was announced, and informal coordination with NVIDIA and other OEMs has increased. The expected timeline for a sixth edition that addresses AI-specific thermal density is approximately 2026–2027, depending on ballot cycles. In the interim, the industry is operating on a patchwork of OEM cold-plate specifications, colocation operator design guides, and peer-operator experience shared at venues like AFCOM and the Open Compute Project Summit. That patchwork is not an adequate substitute for a revised authoritative standard, but it is the reality of the current transition period.

The Coolant-Side Spec Gap

A serious next revision of TC 9.9 needs to add four coolant-side parameters that are currently either absent or underspecified in the H1 class. Each of these parameters has a direct link to system reliability and operational longevity in production AI cooling environments. The de facto interim guidance for most of these parameters comes from NVIDIA's cold-plate specification sheets for the GB200 platform and from the CDU (Coolant Distribution Unit) qualification requirements published by major CDU vendors—a situation that works adequately when all systems in a facility share a single OEM's architecture but breaks down in heterogeneous deployments.

Flow rate floor, expressed in LPM per kW of heat load. Cold-plate thermal resistance is not a fixed value—it is a function of flow velocity across the heat exchanger surface. Below a minimum flow rate, thermal resistance rises steeply and junction temperatures increase faster than coolant supply temperature reductions can compensate. NVIDIA's GB200 cold-plate documentation specifies minimum flow rates at the system level, but TC 9.9 contains no flow rate specification. A revised standard should establish a floor expressed as LPM per kW, with a reference curve showing how thermal resistance varies with flow for a representative cold-plate geometry. This gives facilities teams a facility-side design target that is independent of any single OEM's hardware revision.

Reserve alkalinity floor, referenced against ASTM D1121. Coolant formulations for direct liquid cooling systems rely on corrosion inhibitors that are consumed over time by reaction with metal surfaces, atmospheric CO₂ ingress, and thermal degradation. Reserve alkalinity, measured by the titration method defined in ASTM D1121, is the primary indicator of inhibitor reserve—the buffer that prevents pH from dropping to values where galvanic and crevice corrosion accelerate. TC 9.9 does not specify a reserve alkalinity floor. Facilities running warm coolant loops without a defined alkalinity minimum are flying blind on corrosion risk. A practical floor is a baseline reserve alkalinity ratio of no less than 0.7 relative to fresh fluid specification, with a defined recharge or replacement protocol below that threshold.

Conductivity ceiling in µS/cm. Conductivity is the fastest and cheapest real-time indicator of coolant contamination. For deionized water-based systems, a conductivity ceiling of 100 µS/cm is the practical threshold above which galvanic corrosion rates at dissimilar-metal interfaces begin to accelerate meaningfully. For propylene glycol-based systems, the threshold is higher—typically 2,000–2,500 µS/cm—because the glycol carrier has its own ionic contribution. In-line conductivity sensors are inexpensive, reliable, and already present in most CDU systems; adding a TC 9.9 ceiling would give operators a contractual basis for the alarm setpoints that most CDU vendors already recommend informally.

Particulate filtration specification, sub-50 µm absolute. Cold-plate internal flow channels in current-generation GPU coolers are engineered with tight tolerances to maximize turbulence and heat transfer coefficient. Particulate contamination above approximately 50 µm in dimension can partially obstruct these channels, create localized hot spots, and introduce erosion-corrosion downstream of the obstruction site. NVIDIA's GB200 system documentation recommends sub-50 µm filtration at the system level. A TC 9.9 standard should codify this as a class requirement, and should additionally specify 10 µm bypass filtration at cold-plate inlet manifolds in systems with mixed-metal architectures where corrosion products may be locally generated.

ASHRAE TC 9.9 vs. OCP Cooling Environments

The Open Compute Project's Cooling Environments project operates in a different lane than ASHRAE TC 9.9, and understanding the complementary roles of the two efforts is essential context for any engineer tracking where authoritative liquid-cooling guidance is heading. OCP is a practitioner-driven open hardware consortium. Its working groups move quickly because they operate by contributor consensus rather than formal standards balloting, and because the contributors are often the same hyperscale operators who are deploying the hardware in question. OCP Cooling Environments has published design specifications for data center cooling infrastructure—including CDU performance metrics, coolant chemistry recommendations, and rack-level flow distribution guidance—at a cadence that outpaces the ASHRAE revision cycle by a factor of two to three.

The relationship between TC 9.9 and OCP is not competitive—it is staged. OCP captures fast-moving practitioner consensus and publishes it in a form that is immediately useful to engineers designing and operating current-generation infrastructure. ASHRAE TC 9.9 absorbs validated OCP practices through its working-group process, codifies them in the formal standards language that makes them suitable for reference in contracts, insurance policies, and building specifications, and publishes them with the institutional backing that makes them defensible in regulatory and legal contexts. The two bodies serve different needs in the same ecosystem.

Working-group overlap between TC 9.9 and OCP is well-documented and intentional. Several engineers hold simultaneous memberships in both bodies and actively work to align terminology, test methods, and threshold values. The expected trajectory for the sixth edition of the TC 9.9 thermal guidelines is direct incorporation of OCP Cooling Environments' validated guidance on coolant chemistry, flow specification, and CDU performance requirements. This is good news for the industry. It means that the practitioner-level knowledge being generated in hyperscale deployments today will find its way into the authoritative reference document within a reasonable timeframe, rather than remaining siloed in OCP white papers that lack the contractual reach of an ASHRAE publication.

Standards Tracking Tip: OCP Cooling Environments publishes draft specifications and meeting notes at opencompute.org. Tracking the Cooling Environments project wiki is the fastest way to anticipate where TC 9.9 guidance is heading before the next formal revision is published. When OCP and NVIDIA cold-plate specs converge on the same parameter value, that value is very likely to appear in the next ASHRAE edition.

The practical implication for facilities teams is that neither document should be read in isolation. TC 9.9 provides the class framework, the contractual language, and the baseline thermal envelope. OCP Cooling Environments provides the engineering detail on coolant system design, and NVIDIA's application engineering notes provide the system-specific implementation guidance. A competent liquid-cooling design uses all three, cross-referencing where they align and flagging where they diverge for documented engineering justification. The divergences are fewer than they used to be, and the trend is toward convergence as the standards bodies catch up with operational practice.

How Operators Should Engage

Standards engagement is a small operational investment with a return that most facilities teams underestimate. The mechanics of engagement with TC 9.9 and OCP Cooling Environments are accessible, and the cost of participation—in money and time—is modest relative to the capital exposure of a multi-rack AI cooling deployment.

The first and lowest-barrier step is acquiring the relevant documents. The Datacom Series volumes are available for purchase through ASHRAE's online store, with individual volumes priced at approximately $80–150 depending on format and membership status. ASHRAE members receive discounts, and corporate memberships are available for facilities teams that need multiple copies. The Thermal Guidelines for Data Processing Environments fifth edition and the Liquid Cooling Guidelines volume are the two highest-priority purchases for an AI data center practitioner. Together, they represent less than $300 in reference material for decisions that involve tens or hundreds of millions of dollars in capital infrastructure.

Formal participation in TC 9.9 working groups requires ASHRAE membership at the technical committee level. The onboarding process typically takes 6–12 months, during which a new member attends committee meetings as an observer, builds relationships with existing working-group members, and develops familiarity with the committee's current work items. Voting rights and formal contributor status follow after a period of active participation. The time commitment for an active working-group member runs approximately 2–4 hours per week during active balloting periods and lower between cycles. For a facilities engineering lead at a hyperscale operator, that time investment is defensible as part of a standards compliance function—and the influence that comes from active participation shapes the documents that govern the industry for the next five years.

OCP Cooling Environments engagement has a lower barrier to entry. The project operates through publicly accessible mailing lists, bi-weekly working-group calls, and an open wiki. Participation does not require formal membership fees, though OCP corporate membership ($5,000–$50,000 per year depending on tier) provides additional access to design reviews and advance draft documents. For operators who cannot commit to full TC 9.9 working-group membership, tracking OCP Cooling Environments as an observer is the next best option for staying ahead of where the standards are heading.

Procurement Note: When evaluating CDU vendors, colocation operators, or cooling system integrators, ask explicitly whether their engineering teams participate in TC 9.9 or OCP Cooling Environments. Vendors with active standards participation are more likely to be designing to the actual current and future state of practice rather than to a frozen internal specification. This is a meaningful differentiator for long-term support quality.

For coolant chemistry procurement specifically, standards engagement has a direct operational return. Understanding where TC 9.9 is heading on reserve alkalinity, conductivity, and filtration specifications allows procurement teams to write fluid service contracts that anticipate the next edition rather than lagging it. A coolant program written today to the proposed thresholds discussed in this article—0.7× reserve alkalinity floor, 100 µS/cm conductivity ceiling for DI systems, sub-50 µm particulate filtration—will be compliant with the sixth edition of TC 9.9 on day one of its publication, rather than requiring a costly program overhaul to achieve compliance retroactively.

The broader case for standards engagement is straightforward: the engineers who participate in writing the guidelines are the ones who understand how and why they are written the way they are. That understanding translates directly into better facility design, more defensible procurement decisions, and stronger position in conversations with OEMs, colocation operators, and insurers who reference TC 9.9 as a basis for their own requirements. In a capital environment where a single AI cluster build represents hundreds of millions of dollars of infrastructure investment, spending two hours per week on standards participation is not a luxury—it is due diligence.

For a deeper look at the specific coolant formulations, inhibitor chemistries, and fluid qualification protocols that underpin production AI liquid cooling deployments, see the related sub-pillars on Corrosion Inhibitor Chemistry for Data Center Liquid Cooling, Selecting and Qualifying Coolants for Direct Liquid Cooling Systems, and Conductivity and Chemistry Monitoring for AI Cooling Infrastructure. All of these resources are grounded in the TC 9.9 framework discussed here and align with the OCP Cooling Environments guidance that is shaping the next revision cycle. For Alliance Chemical's full product portfolio serving the data center cooling market—including inhibited propylene glycol formulations, deionized water blends, and coolant additives qualified against NVIDIA cold-plate specifications—visit the AI & Data Center Cooling Chemicals pillar page.