Power usage effectiveness (PUE) is useful, but it doesn’t tell the whole story. Efficiency today also depends on utilization, storage design, cooling strategy, and hardware lifecycle. Our new article explores what reduces energy consumption in modern data centers.
For years, green computing sat near the back of the deck as a compliance footnote most people skipped. What changed is fairly simple: AI workloads pushed power bills up hard, and regulators began asking operators to report the numbers behind them.
Today, efficiency influences budgets, site selection, and every infrastructure refresh cycle. The awkward part is that plenty of what gets labeled green is real engineering that saves money, and some of it is closer to a paint job. Sorting one from the other before you commit budget is what this article is for.
Why this stopped being optional?
The forecasts are what shifted the conversation. Data center electricity consumption is expected to grow 26% in 2026 to roughly 565 TWh, and the IEA sees it doubling to around 950 TWh by 2030, which would put it near 3% of the world’s electricity. At that scale, power stops being an operations detail and turns into something finance watches line by line.
Three pressures now land on the same desk. Power is among the largest recurring costs in the building, so efficiency maps straight onto margin. Grid connections and rack space are finite, which means wasted watts quietly cap how far you can grow. On top of that, regulation has caught up, as the compliance section explains below. Taken together, efficiency has moved from a reputational nicety to an operational constraint.
The AI energy bill nobody expected
AI is where the old capacity assumptions come apart. AI-optimized servers are forecast to reach 175 TWh in 2026 and to overtake conventional servers in 2027, and by 2030 they could account for half of all data center power.
That growth brings a cooling problem as well as a power problem. Dense GPU racks exceed what traditional air cooling can handle, and Goldman Sachs expects 76% of AI servers to be liquid-cooled by the end of 2026. Simply adding more hardware is the most expensive way out of that. In many environments, you can achieve meaningful savings by right-sizing inference, scheduling flexible jobs for hours when the grid is cleaner, and avoiding oversized general-purpose clusters for AI workloads.
Where the savings actually are?
Consolidation is the biggest lever, and one many organizations still underuse. Virtualization and containers pack workloads onto fewer physical servers, lifting utilization from the usual 10-20% toward 50% or more and often halving the box count. What often gets overlooked is that an idle server still consumes a large share of its peak power. Ten machines running at 15% utilization waste considerably more energy than four running at 50%. Fewer machines mean less power drawn, less heat to remove, and less gear to recycle later. It rarely makes a roadmap look exciting, yet it tends to pay back faster than anything else on the list.
Before consolidating, it pays to find the servers already doing nothing. Studies of large fleets keep landing on the same figure: around 30% of physical servers, and a similar share of VMs, are comatose, powered on but delivering no useful compute. They still consume power, hold software licenses, take up rack space, and usually sit several patches behind, which quietly makes them a security problem as well as an energy one. A few weeks of tracking CPU, network, and storage activity is generally enough to identify them, and decommissioning is the rare change that cuts cost, energy, and attack surface in a single move.
Cloud can help here too, though it also blurs the footprint. Hyperscalers run at a power usage effectiveness (PUE) of around 1.09 against an industry average closer to 1.56, so many workloads run more efficiently once they move there. The footprint, however, is relocated rather than eliminated. Egress fees and vendor lock-in still matter, so workload placement should ultimately follow where the data genuinely needs to live.
The storage angle
Storage design quietly decides how much hardware ends up drawing power, which is why it deserves more attention in this conversation than it usually gets. Software-defined storage allows commodity hardware to do work that once required dedicated appliances, reducing both the operational energy footprint and the amount of hardware that eventually reaches end of life.
Two-node hyperconverged clusters are a clear case. They deliver high availability without a separate SAN, and solutions such as StarWind VSAN and StarWind HCI Appliance go a step further by eliminating the need for a dedicated witness node. That keeps the hardware footprint at edge sites as small as practical. Nutanix AOS and VMware vSAN address the same hyperconverged architecture, primarily in larger-scale deployments.
The energy math behind this is worth looking at because it shows where software-defined storage delivers measurable efficiency gains. A classic three-tier setup runs compute servers, a dedicated storage array, and the SAN fabric between them as three separate sets of hardware, each with its own power supplies, fans, and cooling load. The storage array is the least efficient part of that design: its controllers, cache, and fans consume close to baseline power whether the array is 10% utilized or 80% utilized. In other words, a significant portion of the electricity goes toward keeping the array operational rather than serving data.
Hyperconverged infrastructure places storage on the same servers already running the workloads, eliminating the separate storage fabric and much of that fixed overhead. As a result, the power you consume is spread across more useful work. The outcome is lower energy consumption per usable terabyte, not simply a smaller rack footprint – and that is the metric that actually affects your carbon footprint.
Tiering is the other half of the equation. Most data becomes cold within weeks of being written, yet it often remains on the same fast, power-hungry storage tier as active workloads. Keeping hot data on NVMe while moving cold data onto dense, lower-power storage tiers reduces ongoing energy consumption. It also extends the life of storage arrays you already own instead of replacing them prematurely. That matters more than it might seem because the carbon cost of manufacturing a drive has already been paid whether it remains in service for three years or seven. DataCore SANsymphony pools mixed arrays under a single policy engine, and for archives, dense object storage such as DataCore Swarm parks cold data at low watts per terabyte on standard nodes.
Cooling, and the metric everyone games
Cooling is typically 30-40% of data center energy, which makes it the second-biggest lever after consolidation. Free and air-side cooling can pull PUE below 1.2 in the right climate, while aisle containment, liquid cooling for dense racks, and waste-heat reuse account for much of the remaining gains.
One of the cheapest wins is also one of the most ignored, which is raising the temperature setpoint. The American Society of Heating, Refrigerating and Air-Conditioning Engineers (ASHRAE) recommends a server inlet temperature between 18 and 27°C, yet many facilities still run colder than necessary out of habit. Every degree higher within that range can reduce cooling energy by roughly 4–5%, and containment makes those higher temperatures practical by preventing hot and cold air from mixing. Go beyond the ASHRAE recommendations, however, and server fans begin working harder to compensate, offsetting much of the savings. In other words, the efficiency gains come from operating within the recommended range, not outside it.
That naturally raises another question: how much does PUE actually tell you?
Whether PUE still carries the same weight is an active debate across the industry. Equinix and others have questioned whether PUE has outlived its usefulness, pointing out that a facility can report an impressive ratio while still consuming enormous amounts of electricity in absolute terms. PUE also says nothing about water consumption or carbon emissions. The criticism is valid. PUE remains a useful operational metric, but it should be interpreted alongside other measurements instead of on its own.
| Metric | What it measures | Better value | The catch |
|---|---|---|---|
| PUE | Total power / IT power | Closer to 1.0 | Ignores water, carbon, and absolute load |
| WUE | Water per unit of IT energy | Lower | Can rise while you chase low PUE |
| CUE | Carbon per unit of IT energy | Lower | Depends on your grid mix |
| ERF | Share of energy reused as heat | Higher | Needs someone to take the heat |
The ticking compliance clock
Regulation is turning efficiency from a best practice into a reporting requirement. Under the EU Energy Efficiency Directive, data centers with an installed IT power demand of 500 kW or more already submit annual reports to the European Database on Data Centers. The European Commission has published a draft regulation for an EU-wide data center rating scheme as part of its forthcoming Data Center Energy Efficiency Package, which is expected to be adopted later in 2026. In Germany, new data centers entering operation from July 2026 must reuse a minimum share of their waste heat. Alongside these mandatory requirements, Energy Star, EPEAT, and ISO 50001 remain widely used voluntary frameworks.
It is also worth deciding early who owns these metrics, because they do not collect themselves. Facilities or data center operations teams typically manage physical measurements such as PUE and water usage. Platform engineering or SRE teams own utilization and infrastructure telemetry, while FinOps increasingly connects those metrics to cost and carbon reporting. Establishing ownership before reporting deadlines arrive is the difference between a routine submission and a last-minute scramble.
The part that ends up in landfill
Power gets measured almost everywhere, and disposal rarely does. The UN counted 62 million tonnes of e-waste in 2022, heading for 82 million tonnes by 2030, with only about 22% formally recycled, which puts e-waste on track to grow roughly five times faster than the recycling that is supposed to catch it.
Infrastructure decisions influence that trend more than many organizations realize. Extending hardware life through software-defined pooling, decommissioning equipment responsibly, favoring refurbished hardware where appropriate, and purchasing EPEAT-registered systems all help reduce waste. Every storage array you avoid replacing prematurely also avoids the embodied carbon associated with manufacturing its replacement.
What it looks like by industry
The underlying efficiency principles stay the same, but the priorities vary by industry.
Retail and other multi-site operators feel the pressure at the edge. A chain running a few hundred stores does not want a full rack and a SAN behind every checkout counter. Two-node hyperconverged deployments, using solutions such as StarWind HCI Appliance, provide local high availability with only two compact servers, allowing power and cooling requirements to scale much more efficiently across hundreds of sites.
Healthcare providers and media organizations face the opposite challenge: enormous volumes of data that are written once and rarely accessed again. Medical imaging archives and video libraries routinely grow into the petabyte range, with most of that data becoming cold within weeks. Tiering it from NVMe onto dense object storage such as DataCore Swarm while continuing to use existing arrays under a policy engine like DataCore SANsymphony reduces both power consumption and the need to purchase additional storage prematurely.
Financial services, AI platforms, and HPC environments typically encounter power density limits first. Their GPU-heavy racks make liquid cooling and grid-aware workload scheduling practical necessities, and many of these facilities already fall under EU reporting requirements. As a result, energy efficiency and regulatory compliance often become part of the same operational discussion.
Manufacturing and logistics frequently operate infrastructure at the industrial edge, sometimes in spaces with little or no dedicated data center cooling. In those environments, airflow management, temperature setpoints, and minimizing the hardware footprint become operational constraints instead of optimization opportunities. A network closet on a factory floor simply cannot behave like a purpose-built cold aisle.
Green or greenwashed?
Start with a baseline. If you cannot tell someone your PUE, WUE, and average infrastructure utilization, it becomes difficult to justify any efficiency investment that follows. From there, the priorities generally sort themselves by impact:
- Consolidate first, because idle servers are where much of the waste hides.
- Tier storage and extend the useful life of existing hardware.
- Right-size AI workloads and schedule them intelligently instead of overprovisioning clusters.
- Improve airflow and recover waste heat wherever there is a practical use for it.
- Buy efficient hardware backed by a documented decommissioning process.
What separates the two is evidence. If the work shows up on your utility bill, the efficiency gains are real. If the sustainability story depends entirely on renewable-energy announcements while electricity consumption remains unchanged, the improvement is largely an accounting exercise.
The encouraging part is that the changes producing the biggest efficiency gains also tend to deliver the fastest financial returns. That makes establishing a baseline and tackling your first consolidation project a sensible place to start.
FAQ
Is green computing just about saving electricity?
No. It also includes water consumption, embodied carbon in hardware, and electronic waste. In many cases, the most sustainable storage system is the one you never have to replace.
Why does it matter more in 2026?
Because of AI. Data center power is up 26% this year, and EU reporting rules make efficiency a compliance question, not a preference.
What is the single biggest lever?
For most organizations, consolidation delivers the largest gains. Higher utilization means fewer servers, lower power consumption, reduced cooling requirements, and less hardware to replace over time.
What is a good PUE?
Leading hyperscalers typically operate around 1.09, while the industry average remains close to 1.56. PUE is still useful, but it should always be considered alongside metrics such as WUE and CUE.
How does storage tiering cut carbon?
Moving cold data onto lower-power storage extends the life of existing arrays and delays new hardware purchases. Since manufacturing accounts for a significant share of a storage system’s lifetime carbon footprint, extending its useful life reduces overall emissions.
