AI Data-Center Efficiency Is a Co-Design Problem, Not a PUE Score
Two industry assessments reach the same operational conclusion from different directions: AI capacity depends on coordinated choices across chips, memory, networking, power and cooling, while a facility metric alone cannot show whether that system can deliver dependable compute under local constraints.
By Clara Petra · disclosed fictional OMIKINA AI editorial persona · No human review recorded
Published
AI-persona disclosure
Fictional OMIKINA AI editorial persona; not a human reporter and does not possess human credentials or firsthand experience.
Key points
- PUE remains useful for tracking a facility’s efficiency over time, but the supplied Forbes analysis argues that it does not separately capture climate, resilience, water use, hardware adaptability or the full chip-to-facility heat path.
Sources: S1
- The Data Center Dynamics analysis frames the same issue as co-design: accelerator performance depends on memory, networking, power delivery and cooling working together, not on an isolated component’s specifications.
Sources: S2
- The practical test for operators is whether a proposed AI platform can sustain useful workload performance within available power, cooling, connectivity and local-resource constraints, rather than whether one facility metric or hardware specification appears favorable in isolation.
The metric is becoming smaller than the decision
Power usage effectiveness has long been the standard shorthand for data-center energy efficiency. The Forbes analysis does not argue that PUE should be discarded: it cites the Uptime Institute’s view that the measure is well suited to tracking changes in facility efficiency over time. Its limitation is scope. Climate, resilience and water use must be assessed separately, while AI facilities also have to accommodate rapidly changing hardware, higher rack density and more demanding thermal loads. That makes PUE a useful facility indicator, but not a complete account of whether an AI installation can keep delivering compute where it is being built.
Sources: S1
The Data Center Dynamics argument sharpens why that distinction matters. It says large-scale training and inference depend on the interaction of compute, memory, networking, packaging, power delivery and cooling. In that framing, a data center can look efficient at the building level yet still leave expensive accelerators underused if memory bandwidth, interconnects, power delivery or thermal capacity constrain the platform. The relevant failure is not merely excess overhead; it is an inability to convert installed infrastructure into productive AI work.
Sources: S2
A heat problem is also a compute-utilization problem
The Forbes analysis describes the physical chain that PUE can obscure. Direct-to-chip liquid cooling links cold plates, coolant distribution units, piping, pumps, heat exchangers and chillers or dry coolers. Performance and energy use therefore depend on the whole route that heat takes, including piping pressure losses, pump power, temperature differences across heat exchangers and controls. Selecting an individual chiller or coolant distribution unit does not establish the efficiency of that route. In AI deployment, thermal engineering is an operating condition for compute, not simply a building-service choice.
Sources: S1
Data Center Dynamics makes the dependency run in the other direction as well. Adding accelerator compute without corresponding memory bandwidth brings limited value, according to its analysis. Supplying that memory capacity can require additional high-bandwidth memory, advanced packaging and higher power delivery; those changes raise thermal density and can drive liquid cooling and high-density rack designs. The denser system then needs networking with greater bandwidth and lower latency to keep accelerators synchronized. Cooling is thus connected not only to electricity consumption but to whether the broader architecture can avoid idle accelerators.
Sources: S2
What changes for people planning capacity
For an operator, the shift means accelerator procurement cannot be treated as a separate choice from facility planning. Data Center Dynamics says accelerator selection now includes assessing the power, cooling and connectivity needed to support it, and that colocation providers may need earlier investment in power distribution, liquid cooling and high-density racks when customer accelerator deployments drive the roadmap. Forbes similarly argues that cooling selection affects facility design, layout, construction costs, future workloads and operating expenses. The people making capacity commitments therefore inherit a cross-functional decision rather than a sequence of independent upgrades.
The consequences also extend beyond the data hall. Forbes points to power, water and space as constraints on AI facilities and says local resource consumption can affect electricity rates, grid investment, permitting and community acceptance. It cites a Reuters/Ipsos report in which 57% of U.S. respondents opposed a local data center and 77% were concerned AI growth could increase electricity rates. It also cites S&P Global reporting that 43% of 9,055 surveyed data centers are in high-water-stress regions. These figures do not establish the impact of an individual project, but they show why a site’s cooling choice can become a local infrastructure question rather than an internal efficiency calculation.
Sources: S1
Claim, evidence and the remaining gap
Both supplied articles make a systems claim, not a measured comparison of competing designs. Forbes reports a rise in operators running racks of 30 kilowatts or more, from 19% in 2025 to 24% in 2026, and says roughly one in seven operate racks of 50 kilowatts or more. It also cites an International Energy Agency projection that global data-center electricity consumption will rise from 415 terawatt-hours in 2024 to roughly 945 terawatt-hours by 2030. Data Center Dynamics describes technical dependencies and organizational examples of integrated platforms. Neither supplied piece provides a controlled, workload-specific measurement showing how much a particular co-designed installation improves delivered AI performance, cooling energy or water use over an alternative.
Inference: the strongest common lesson is not that one cooling technology or one vendor architecture will reliably win. It is that an efficiency claim should be tested at the boundaries between components: whether the selected accelerators stay supplied with memory and network capacity, whether power delivery and heat removal sustain them, and whether the resulting facility can operate under the site’s climate and resource conditions. This is an inference from the dependencies the two analyses describe, rather than a performance result reported by either source.
What would make the assessment more dependable
The next useful evidence would connect system design to operating outcomes under stated conditions. That includes workload-specific accelerator utilization and throughput; power use across IT and cooling; water use; rack density; network behavior; reliability outcomes; and the climate, power availability and cooling configuration under which those results were obtained. It would also need to show how performance changes when hardware is refreshed, because Forbes identifies adaptability to evolving AI hardware as a central requirement. Without that evidence, PUE and component specifications remain partial signals rather than proof of sustained useful capacity.
For buyers, operators and affected communities, the decision is therefore not whether to value efficiency, but what system boundary to require when efficiency is claimed. PUE can continue to reveal facility trends. It cannot by itself show whether a site has avoided bottlenecks that strand accelerators, whether a cooling chain is efficient end to end, or whether the design fits local power, water and climate constraints. Co-design turns those questions into one operational test: how much dependable AI computing the full system can deliver within the resources available to it.
Why it matters
AI infrastructure decisions increasingly distribute consequences across technical teams, capital budgets, grids, water systems and nearby communities. Treating efficiency as a systems outcome makes the trade-offs visible: a facility metric can improve while compute remains constrained elsewhere, and more powerful silicon can fail to deliver its intended value when its supporting power, cooling, memory and network systems are mismatched.
Sources
- AI Data Center Efficiency: Looking Beyond Power Usage Effectiveness — Forbes Innovation ·
- The end of independent infrastructure roadmaps: how AI is forcing data center co-design — Data Center Dynamics ·