Server Room Energy Costs: 5 Levers

In 2026, electricity prices are still a topic that many IT leads tackle backwards: they ask what new hardware to buy instead of looking at which existing power hogs they can reduce without investment. In our experience, 15–30 % of energy can be saved in most small and medium server rooms without replacing a single server.

This article describes five concrete levers we see at DATAZONE in almost every customer server room — sorted by effort / effect ratio. At the end comes the question: when does new, more efficient hardware actually pay off?

Important caveat: we don’t make up percentages. Where we name magnitudes, they refer to published vendor data (Intel / AMD power profiles, APC / Vertiv UPS efficiency curves, ASHRAE recommendations). Actual savings depend strongly on the starting point — real measurement with a power meter at the distribution panel is always the more honest answer than any spreadsheet.

Lever 1: Separate Hot Aisle From Cold Aisle

Hot-aisle / cold-aisle separation has been standard in large data centres for years. In small server rooms with one to three racks it’s almost always forgotten — with the result that cooling works against itself: cold air from the AC mixes with hot exhaust air from the servers before it even reaches the server inlet. The AC then has to blow colder than necessary.

What helps concretely

• Align servers consistently — fronts in one direction, exhausts in the other
• Blanking panels in empty rack units — prevent air short-circuits between hot and cold sides
• Seal floor and ceiling gaps — typical weak spots are cable pass-throughs and floor tiles
• Foil or curtain containment above the rack aisle — feasible even in small rooms with one or two racks (roughly €200–800 in material)

Effort and effect

• Effort: low to medium. Blanking panels and foil containment are doable inside a maintenance window
• Effect: noticeable — the AC cycles less and can run a higher setpoint, which directly cuts power consumption

ASHRAE recommends 18–27 °C inlet temperature for IT equipment (Class A1). If you cool to 17–18 °C today, after containment you can move to 23–25 °C — with no thermal problems on the servers.

Lever 2: Server Power Profiles on “Balanced”

This is the easiest win of all. In very many enterprise servers (Dell PowerEdge, HPE ProLiant, Lenovo ThinkSystem, Supermicro etc.) the BIOS power profile is set ex factory to “Maximum Performance” or “Performance” — which means: CPUs never clock down, even at idle.

For workloads that don’t need every microsecond of latency advantage (i.e. 95 % of all workloads), “Balanced” or “OS Control” is the much more efficient choice. The CPU clocks down at low load and jumps back to nominal frequency on demand.

Concrete settings at major vendors

• Dell iDRAC: System Profile → “Performance Per Watt (OS)” or “Performance Per Watt (DAPC)”
• HPE iLO: Workload Profile → “General Power Efficient Compute”
• Lenovo XCC: Operating Mode → “Efficiency – Favor Performance”
• Supermicro IPMI: Power Profile → “OS Controlled” or “Energy Efficient”

Effort and effect

• Effort: minimal. One BIOS setting per server, doable in a maintenance window, ideally via iDRAC / iLO / Redfish script for mass rollout
• Effect: on lightly loaded servers (typical mid-market hypervisor at ~30 % load), the power savings per server are measurable — own power measurements at the server show this within minutes

If you run performance tests, you should know the BIOS setting — some latency-sensitive applications (HFT databases, some real-time industrial applications) need “Maximum Performance”. But for file servers, AD domain controllers, normal web backends, and VM hosts, “Balanced” is the correct default.

Lever 3: Consolidate or Shut Down Idle VMs

In long-grown virtualization environments a variable accumulates that nobody likes to touch: zombie VMs. VMs that have been running for months or years without anyone knowing why. Test VMs from the last project. “Might-still-need” VMs. Migrated VMs whose successor has long since gone live.

Each of these VMs consumes hypervisor resources — CPU shares, RAM, storage IO, sometimes backup slots. The direct power bill is low per VM, but cumulative impact is significant.

Procedure

1. Inventory: list all VMs from Proxmox / VMware with last login date, 30-day CPU average, owner tag
2. VMs without owner tag → flag, discuss in IT steering meeting
3. VMs with CPU average < 5 % and no active service → snapshot, shut down, 30 days observation
4. After 30 days without complaint → archive status, after 90 days delete
5. VMs with similar profile → check consolidation (e.g. three small web servers into one with three reverse-proxy routes)

Effort and effect

• Effort: medium. This is more organisation than technology. Hardest part: the conversation with departments about “their” test system
• Effect: indirect but cumulative. Fewer VMs means: fewer hypervisor licences (VMware-relevant), less backup storage, less idle hypervisor CPU, lower patch load

Side effect: consolidated VMs often allow shutting down an entire hypervisor host — eliminating the biggest single power line item at one stroke.

Lever 4: UPS Efficiency

The UPS is the invisible component nobody measures — and depending on type it has significant losses. Three topologies with different efficiency:

UPS type	Operating principle	Typical efficiency	Typical loss in normal operation
Offline / standby	Mains passthrough, inverter on outage	very high (>95 %)	low — but switchover 4–10 ms
Line-interactive	Inverter in parallel, AVR	high (94–97 %)	low (~2–3 %)
Online / double-conversion	Permanent AC→DC→AC conversion	depends on eco mode	medium (~5–10 % without eco mode)

Online UPSes protect best — they filter mains disturbances, compensate voltage fluctuations, and have no switchover time. But they also have the highest loss in normal operation. Concrete steps:

• Activate eco mode / high-efficiency mode when mains quality permits. APC, Eaton, Vertiv, and Riello support this on larger models.
• Match UPS sizing to actual load — a UPS at 15 % load almost never runs in its most efficient point. Vendor efficiency curves typically show an optimum at 50–80 % load.
• Stick with line-interactive if no specific requirements (medical equipment, lab) call for double conversion. Often the right choice for SMB server rooms.

Effort and effect

• Effort: low (activate eco mode) to high (replace UPS). Eco mode activation is one click in the web front-end, UPS replacement is a capital project
• Effect: noticeable on the eco-mode switch — typical vendor figures show efficiency gain of 2–4 percentage points over double-conversion mode

Important: eco mode means higher switchover time on mains disturbance (milliseconds rather than uninterrupted). For an SMB server room with modern servers (all PSUs redundant), this is usually acceptable — for specialty installations or older hardware, take a closer look.

Lever 5: Right-Size Cooling

One of the most common bad investments in small server rooms is a significantly oversized AC. It was once ordered for worst case and has been continuously cooling a load only half that size. Result: frequent cycling, low efficiency in the part-load range.

What helps

• Measure actual thermal load — don’t extrapolate from PSU nameplate, measure actual power consumption of the server hardware (1 W electrical → ~1 W heat)
• Raise setpoint — once containment is in place (lever 1), the setpoint can move to ASHRAE A1 25–27 °C inlet — much higher than the AC setpoint
• Check free cooling — at outside temperatures below ~15 °C, a direct or indirect outside-air cooling system can pause compressor cooling
• Inverter AC instead of classic on/off — when the unit is end-of-life, the step to a continuously variable inverter is worthwhile

Effort and effect

• Effort: low (adjust setpoint) to high (replace unit). Setpoint adjustment costs nothing, new AC is a capital expense
• Effect: substantial. ACs are often 30–50 % of total power in the server room — every 1 K increase reduces compressor run

Caution: never raise setpoints without inlet temperature sensors at the servers. A cheap USB temperature box with five sensors (inlet, middle, outlet, rack-top, rack-bottom) is mandatory companion gear.

When New Hardware Is Worth It

At DATAZONE we rarely answer the question “should I replace?” with “yes, immediately”. More often with “do the math”. Concretely:

ARM servers (Ampere Altra, AWS Graviton class, Apple-silicon-like)

For the right workloads — web backends, container workloads, many small microservices — ARM servers are significantly more efficient than x86. Performance-per-watt is often 30–50 % better on modern ARM. But: licensing and software compatibility must fit. For classic Windows Server workloads, Oracle DB, SAP — forget ARM (as of 2026).

EPYC vs. Xeon

Since the Genoa / Bergamo / Turin generations, AMD EPYC is regularly ahead on performance-per-watt — Intel has closed in considerably with the Sierra Forest efficiency cores, but in many SMB scenarios is still half a generation behind EPYC. On a server refresh today, AMD should be seriously on the comparison list.

When to replace?

Rule of thumb: if the server is more than six years old and at the same time runs more than 50 % load, replacement almost always pays off. For younger servers or lower utilisation, levers 1–5 are almost always more economical than new acquisition.

What We Recommend at DATAZONE

A practical workflow for an “energy optimisation sprint”:

1. Measure — 1 week of power measurement at the main distribution, separately at AC and IT
2. Quick wins (1 day) — install blanking panels, BIOS to balanced, activate UPS eco mode, carefully raise setpoint
3. Mid-term measures (1 month) — idle VM inventory, hypervisor consolidation if applicable
4. Investment decision — for old servers or end-of-life ACs: technical and commercial assessment
5. Re-measure after 3 months — the before / after comparison turns the sprint into proof for management

Related DATAZONE articles:

• Scrappage premium for old servers
• Wortmann AG: sustainable IT vendor
• TrueNAS in the data center: future
• Proxmox hardware planning

Conclusion

Lowering server-room energy costs is not a hardware problem — it’s a measurement and configuration problem that in 80 % of cases can be solved without new acquisitions. The five levers in this article cost together a few hours of labour and a little material — and produce measurable savings without a single workload performance metric suffering. The harder part isn’t implementation — it’s the before-and-after measurement, so management sees the success.

Sources

• ASHRAE Thermal Guidelines for Data Processing Environments
• Dell PowerEdge System Profile Documentation
• HPE Power Regulator for ProLiant
• APC Eco Mode Application Note
• Uptime Institute: Annual Data Center Survey