AI workloads break data center assumptions that held for twenty years.
Traditional data center design was built around CPU-dominated workloads with predictable, relatively steady power draw. A standard server rack running at 5-10 kW is a solved infrastructure problem. A GPU cluster running an AI training job doesn’t work that way. Power draw spikes during compute- intensive forward and backward passes, drops during data loading, then spikes again. Cooling systems designed for steady-state thermal loads struggle with that variability. Electrical distribution designed for average draw, not peak burst, undersizes for the actual demand curve.
This isn’t a theoretical concern. It’s why data centers built before the AI infrastructure cycle are retrofitting at significant cost, and why colocation buyers are discovering that “AI-ready” is a marketing claim that needs engineering verification.
ASHRAE, NEMA, and PNNL’s jointly published AI Data Center Energy Performance Framework provides consolidated guidance across three application areas, new construction, retrofits, and ongoing operations, and two technical domains: cooling systems and electrical distribution. The regulation pillar has already addressed the compliance obligations this framework creates for operators. This brief focuses on the engineering decisions the framework is designed to inform.
What AI Workloads Actually Do to Data Center Design
Start with density. High-density AI compute clusters routinely operate at rack-level power densities of 30-100 kW per rack, an order of magnitude above traditional data center densities in many facilities. That density creates two simultaneous engineering challenges. First, cooling: removing heat from 100 kW in a rack footprint requires liquid cooling or high-density air cooling approaches that most legacy facilities weren’t designed to support. Second, electrical: peak power delivery at those densities requires electrical distribution infrastructure sized for burst capacity, not average load.
Add the variability dimension. GPU workloads don’t just draw more power, they draw it erratically. A training job running on a cluster of H100s will hit near-peak power consumption during forward and backward passes and drop during checkpoint saves and data preprocessing. Facilities management systems designed around stable load profiles generate false alerts or fail to trigger the right cooling responses when load profiles change rapidly within a training run.
Then add the grid coordination dimension. Data centers operating at AI-scale power draw have a materially different relationship with their utility than traditional facilities. Burst power requirements and variable load profiles affect grid planning, capacity reservation, and in some jurisdictions, demand response obligations. The framework addresses this explicitly through its electrical distribution guidance.
The Three Publishing Bodies and What Each Contributes
Understanding the framework is easier if you understand what each publishing body contributes.
ASHRAE, the American Society of Heating, Refrigerating and Air-Conditioning Engineers, owns the cooling domain. ASHRAE standards have governed data center thermal management for years. Its contribution to this framework addresses the cooling architecture questions: what cooling approaches are appropriate for different density profiles, how retrofits should be assessed for feasibility, and what operational protocols manage variable thermal loads. The framework builds on ASHRAE’s existing Thermal Guidelines for Data Processing Environments with additions specific to AI workload density.
NEMA, the National Electrical Manufacturers Association, owns the electrical domain. Its contribution addresses electrical distribution design: how to size distribution infrastructure for AI workload burst capacity, what power quality considerations apply at AI-scale loads, and how to plan electrical capacity for facilities running mixed traditional and AI compute workloads.
PNNL, Pacific Northwest National Laboratory, a Department of Energy national laboratory, contributes the energy efficiency research layer. PNNL’s role in this framework is providing the empirical research basis for the guidance: what the actual energy performance data says about AI workload efficiency at scale, what optimization approaches have demonstrated results, and how facilities can measure and report energy performance in ways that are meaningful for AI-specific workloads.
That combination of ASHRAE, NEMA, and PNNL isn’t a committee of bureaucrats, it’s the domain authority for cooling engineering, electrical standards, and energy efficiency research brought to bear on the same document. That’s why the framework carries more weight than a single organization’s guidance would.
Three Audiences, Three Implementation Starting Points
The framework addresses different decisions for engineers, operators, and utilities. The practical starting points differ accordingly.
For data center engineers: The design decisions that matter most are rack density planning and cooling architecture selection. Before specifying cooling systems for an AI-capable facility – new build or retrofit, engineers should use the framework’s density guidance to establish which cooling approaches are viable at their target rack density. Liquid cooling (direct liquid cooling, rear-door heat exchangers, immersion) becomes necessary above certain density thresholds. The framework provides that threshold guidance. Engineers designing electrical distribution should size for burst capacity at peak AI workload draw, not average load, the NEMA contribution to the framework addresses this specifically.
For operators: The operational protocol questions are different from the design questions. Running a facility with AI workloads requires cooling management systems that can respond to rapid load changes, not just steady-state thermal management. Operators should assess whether their building management systems and cooling controllers can handle variable load profiles at the cadence AI workloads create. The framework’s ongoing operations guidance addresses this.
For utilities and grid coordinators: The framework’s grid reliability content is relevant to utilities serving AI data center customers and to data center operators managing demand response obligations. Variable AI workload power draw creates grid coordination complexity that traditional load forecasting doesn’t capture well.
Implementation Starting Points for Engineering Teams
The practical question for engineering teams working with the framework now is where to start, particularly for facilities not in the new-build category.
Retrofit feasibility assessment comes first. Not every facility can support high-density AI compute without structural changes. The framework’s retrofit guidance provides an assessment methodology – engineering teams should run that assessment before committing to AI workload density targets. The assessment covers electrical capacity headroom, cooling system upgrade pathways, and floor loading constraints.
Current rack density benchmarking follows. Engineering teams should measure their current average and peak rack densities against the framework’s tiered guidance. Facilities operating below the framework’s lower density thresholds for AI workloads may need phased infrastructure upgrades before supporting production AI compute at scale.
Utility coordination is the step most engineering teams underweight. The framework’s grid reliability content is not just for utilities, it’s a prompt for data center engineers to have specific conversations with their utility providers about how AI workload variability affects capacity planning and whether demand response programs apply. Those conversations take time to work through procurement and facilities planning processes. Start them earlier than feels necessary.
The Infrastructure Investment Connection
This framework doesn’t exist in isolation from the capital markets activity in AI infrastructure. The gigawatt-scale power capacity race that’s driving hyperscaler and colocation buildout decisions is, at the engineering level, a problem this framework is designed to address. Enterprise buyers evaluating colocation providers for AI workloads now have a reference standard they can use to assess whether a provider’s facility is genuinely AI-ready, or whether the marketing claim runs ahead of the engineering reality.
Colocation RFPs that don’t ask for framework compliance documentation are leaving verification on the table.
TJS Synthesis
The ASHRAE/NEMA/PNNL framework gives engineering teams something that hasn’t existed before: a single document where the cooling, electrical, and energy efficiency standards for AI workload density are addressed together by the organizations that own each domain. The practical value is highest for teams doing retrofit feasibility assessments and for enterprise buyers evaluating colocation providers. Start with the retrofit assessment methodology and the electrical distribution guidance for burst capacity sizing. If you’re evaluating colocation options for AI compute, use the framework’s density and cooling requirements to build your RFP verification checklist, a provider that can’t answer to the framework’s guidance probably can’t support the workload.
The primary framework document is available through ASHRAE’s publications database and PNNL’s website, engineering teams should access the source document directly rather than relying on trade press characterizations, including this one.