For years, data center efficiency was judged mostly by one number: PUE.
Power Usage Effectiveness gave the industry a simple way to measure how much total facility energy was required to support the energy actually consumed by IT equipment. If a data center had a PUE of 2.0, that meant the facility used roughly one watt of overhead power for every watt delivered to IT equipment. If the PUE moved closer to 1.0, the facility was becoming more efficient.
That metric was useful. It forced the industry to get serious about cooling efficiency, airflow management, containment, power distribution losses, lighting, controls, and facility overhead. It also gave owners, engineers, and operators a common language for comparing facility performance.
But AI is exposing the limits of that old scorecard.
A data center can have a strong PUE and still waste a fortune if the GPUs are underutilized, if electrical losses are quietly compounding upstream, if thermal constraints force throttling, if harmonics increase heat, or if the facility can deliver power but not enough useful compute from that power. In the AI era, the winning question is no longer only, “How efficient is the building?”
The better question is, “How much useful compute do we get from every watt we consume?”
That is the shift from PUE to performance per watt.
And it changes where companies like Moonshot fit into the conversation.
PUE still matters, but it no longer tells the whole story
PUE is not obsolete. It is still one of the most important facility-level efficiency metrics in the data center industry. The problem is that PUE measures the efficiency of the facility overhead around IT equipment. It does not tell you whether the IT equipment is being used well, whether the workload is valuable, or whether the power delivered into the rack is being converted into useful AI output.
That distinction matters more now because AI workloads consume power differently than traditional enterprise computing. AI training and inference can create dense, sustained, high-value loads. The U.S. Department of Energy reported that data center load growth has tripled over the past decade and is projected to double or triple by 2028, driven in part by AI applications, domestic manufacturing, and broader electrification. (energy.gov)
Globally, the International Energy Agency estimated that data centers consumed about 415 TWh in 2024, or roughly 1.5% of global electricity consumption, and that this demand grew at about 12% per year over the prior five years. IEA also notes that accelerated servers are increasing power density in data centers, which makes accelerator adoption a key driver of future electricity demand. (iea.org)
So, yes, facility overhead still matters. But when the load itself becomes enormous, the value created by each watt matters even more.
A traditional data center with a PUE of 1.3 might be considered efficient. But if much of the compute is idle, poorly scheduled, thermally constrained, or waiting on network and storage bottlenecks, the facility is not truly efficient in business terms. It is simply good at delivering power to equipment that may not be producing enough value.
That is why performance per watt is becoming the more strategic metric.
CoreSite’s 2026 data center outlook captures the mood clearly. It notes that the “idle GPU epidemic” is expected to create an industry-wide awakening, where the question becomes less about how much compute a company owns and more about how intelligently it orchestrates every watt, workload, and chip. The same article points to performance per watt as the next competitive barometer for AI infrastructure. (coresite.com)
That is the right framing. AI infrastructure is not rewarded for simply existing. It is rewarded for converting energy into output.
The old efficiency wins are not enough anymore
The data center industry has already made enormous progress on efficiency. PUE has improved dramatically over the past two decades. Global Electricity summarizes the historical shift by noting that average PUE has moved from roughly 2.0 in the early 2000s to about 1.5 to 1.6 for typical facilities today, while leading hyperscale operators often report PUEs closer to 1.1 to 1.3. That improvement came from better cooling, hot and cold aisle containment, optimized airflow, more efficient power distribution, and better building design. (globalelectricity.org)
Those improvements matter. They should not be dismissed.
The problem is that efficiency gains are now being overwhelmed by demand growth. The same Global Electricity analysis notes that despite major efficiency improvements, absolute energy consumption continues rising because AI demand is growing so fast. It also points out that modern processors deliver much higher performance per watt than older generations, but those gains are partly offset because chips are becoming more powerful and consuming more total power. (globalelectricity.org)
That is the infrastructure version of running faster on a treadmill.
The industry gets better at using energy, then immediately deploys more compute, denser racks, larger models, more inference workloads, and bigger clusters. This does not mean efficiency is failing. It means the efficiency target has moved.
In the old model, the focus was on reducing facility overhead. In the AI model, the focus expands to the full chain: utility power, transformers, switchboards, UPS systems, PDUs, bus systems, rack power, cooling, workload scheduling, thermal behavior, monitoring, and compute utilization.
A watt lost in the electrical path is wasted. A watt converted into unnecessary heat is wasted. A watt delivered to an idle accelerator is wasted. A watt that produces compute but forces throttling because the thermal system is constrained is partially wasted.
The new KPI is not just low overhead.
It is useful work per watt.
Performance per watt is a power architecture problem, not just a chip problem
It is tempting to think performance per watt is mostly a semiconductor issue. Better GPUs, better CPUs, better accelerators, better memory, and better software all matter. That is true.
But it is incomplete.
A chip does not operate in isolation. It sits inside a rack. The rack sits inside a power and cooling architecture. The facility receives power from the grid or on-site sources. That power moves through transformers, switchboards, UPS systems, breakers, busway, PDUs, and conductors before it becomes useful compute. Every piece of that chain affects real efficiency.
Electrical design drives efficiency in several practical ways.
First, voltage architecture matters. Higher voltage can reduce current for the same power delivery, which reduces I²R losses and conductor heating. That is one reason 800 VDC and higher-voltage architectures are getting attention for high-density AI environments.
Second, distribution losses matter. Transformers, UPS systems, PDUs, bus systems, cable lengths, connections, and conversion stages all create loss. A single small inefficiency may look tolerable. Across a 50 MW, 100 MW, or 300 MW facility, small losses become large operating costs.
Third, harmonics matter. AI clusters depend on power electronics. Nonlinear loads can distort current waveforms, increase heating, stress transformers, reduce power quality, and narrow thermal margin. A system that looks efficient on a basic power model may perform differently when harmonic content and sustained utilization are included.
Fourth, thermal margin matters. Electrical losses become heat. Heat increases stress. Higher temperatures reduce equipment life, increase cooling demand, and can push systems toward derating or throttling. Efficiency is not just an energy bill issue. It is a reliability issue.
Finally, monitoring matters. You cannot optimize what you cannot see. If the only data available is at the facility meter, operators may know the total energy draw but not where waste is occurring. The best AI facilities will need much more granular visibility into power, energy, current, harmonics, thermal conditions, and load behavior.
This is where Moonshot becomes part of the efficiency equation.
Moonshot’s PDU offering is not just a passive distribution product. Moonshot describes its PDUs as designed and engineered in the USA, assembled in Texas, and available with dumb and smart monitoring capabilities for data centers and Bitcoin mining. The Leviathan V2 Smart PDU includes individual outlet monitoring, aggregate harmonic detection across lines and neutral, voltage monitoring, current monitoring, outlet control, power measurement, energy-used measurement, and harmonic current measurement. (moonshotus.com)
That matters because performance per watt requires visibility. If power distribution equipment can help operators understand energy used, current behavior, line conditions, and harmonic current, it becomes part of the optimization loop rather than just a cost line.
Moonshot’s modular data center systems also include integrated options such as smart PDUs, interior and exterior temperature sensors, differential pressure sensors, VFD fan motor speed control, automated system monitoring, reporting, and power management. (moonshotus.com)
That is exactly the kind of infrastructure the performance-per-watt era requires: measured, controllable, and designed around real operating behavior.
Efficiency is now a reliability conversation
In AI infrastructure, efficiency and reliability are no longer separate goals.
A poorly designed electrical path does not only waste energy. It creates heat. Heat narrows margin. Narrow margin increases failure risk. Failure risk forces operators to overbuild, derate, slow deployment, add redundancy, or accept downtime exposure. All of those decisions affect cost and performance.
Likewise, a data hall that is thermally constrained may be forced to run below the theoretical capacity of its compute hardware. From a financial perspective, that is brutal. The owner paid for the GPUs, paid for the building, paid for the power infrastructure, and paid for the cooling system, but cannot extract the full value from the asset because one part of the infrastructure chain is limiting the rest.
CoreSite’s 2026 outlook points to this convergence. It notes that AI’s power and thermal requirements are making liquid cooling mainstream, and that modular liquid cooling systems starting around 2 MW are expected to become common for high-density builds. It also says new two-phase direct-to-chip cooling solutions are expected as rack densities climb toward and beyond one megawatt. (coresite.com)
That is not just a cooling story. It is a power story.
If racks are moving toward hundreds of kilowatts, then the electrical system has to deliver that power with minimal loss, stable quality, adequate protection, and sufficient monitoring. It also has to support the cooling infrastructure that keeps those racks alive. Pumps, CDUs, controls, sensors, fans, and monitoring systems all become part of the critical power chain.
This is why performance per watt is not a software metric alone. It is a system metric.
The system includes the building.
The Moonshot angle: from vendor to efficiency partner
Moonshot should not position itself as another electrical equipment vendor in this conversation.
That undersells the opportunity.
Moonshot’s UL891 switchboard page says its engineering team designs solutions based on supply chain constraints and maximum flexibility, can use major breaker manufacturers, and evaluates site design during the sales process to provide code-compliant, cost-effective, quick solutions. (moonshotus.com)
That is the right foundation, but the performance-per-watt story elevates it.
The message is no longer only, “Moonshot can build the gear faster.”
The message becomes, “Moonshot helps design the power infrastructure that determines how efficiently compute can run.”
That includes switchboards designed with realistic load profiles, PDUs that provide granular power visibility, modular systems that integrate monitoring and controls, harmonic awareness, thermal planning, expansion flexibility, and early coordination between engineering and manufacturing.
For a customer, that difference is important. A low-cost power distribution package can become very expensive if it creates avoidable losses, poor visibility, thermal stress, or expansion constraints. An engineering-led distribution strategy can reduce friction, preserve margin, and help the operator get more useful work from the same power envelope.
That is a direct business outcome.
In a power-constrained market, the facility that can produce more useful compute from a fixed megawatt allocation wins. If two operators both secure 50 MW, the better business is not the one with the prettier one-line. It is the one that converts more of those megawatts into revenue-producing AI output.
Moonshot belongs in that equation because every watt passes through electrical infrastructure before it becomes compute.
The real takeaway
The next phase of AI infrastructure will not be won by whoever builds the biggest data center or buys the most GPUs.
It will be won by whoever extracts the most useful compute from constrained power.
PUE still matters, but it is no longer enough. A facility can look efficient on paper and still underperform if its compute is idle, thermally constrained, poorly monitored, or fed by inefficient distribution. Performance per watt expands the conversation from facility overhead to the full infrastructure chain.
That is where electrical design becomes strategic.
Switchboards, PDUs, voltage strategy, harmonic management, monitoring, modular infrastructure, cooling integration, and expansion planning all affect how much useful output a facility can generate from its power allocation. These are not background details. They are part of the business model.
For Moonshot, this is the opportunity to move from vendor to strategic partner. The company is not just helping customers distribute power. It is helping them preserve efficiency, visibility, reliability, and future capacity in a market where every watt is becoming more valuable.
If you are planning a high-density AI deployment, the right question is not only how much power you can get.
The better question is how much useful compute your infrastructure can produce from it.
https://moonshotus.com/request-form/
Because in AI infrastructure, power is not just an input.
It is the scoreboard.
Sources:
https://www.globalelectricity.org/data-centers-energy-consumption/
https://www.coresite.com/blog/data-center-outlook-2026-power-and-cooling-challenges-and-solutions-are-top-of-mind
https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase-electricity-demand-data-centers
https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai
https://moonshotus.com/ul-891-switchboards/
https://moonshotus.com/power-distribution-units/
https://moonshotus.com/modular-datacenters/

