Insights
Power planning for modern servers has become a critical consideration as organisations adopt higher-density virtualisation, GPU acceleration and AI workloads. Many existing server rooms may no longer have the power, cooling or rack capacity to support new infrastructure without upgrades.
That assumption is rapidly becoming outdated.
Five years ago, a typical virtualisation host might have looked something like this:
A rack containing ten servers of this type would typically consume between 4kW and 6kW. Add some switches and supporting devices, and most server rooms could
Today's virtualisation environments often consolidate significantly more workloads onto fewer hosts.
While these systems deliver dramatically higher performance, they can consume nearly double the power of similar systems deployed only a few years ago.
The largest shift in infrastructure power planning is the adoption of GPUs.
Many organisations are investigating:
The following example demonstrates how power planning for modern servers has evolved as virtualisation platforms become more powerful and workloads continue to grow.
Example Configuration
Typical Power Consumption
In many environments this remains manageable.
Example Configuration
Power Requirements:
A single server can now consume as much power as three or four traditional virtualisation hosts from only a few years ago.
Five years ago, many mid-market racks operated comfortably at:
Today we regularly see designs targeting
AI and GPU environments may exceed:
This creates challenges beyond simply supplying electricity.
Higher rack densities increase:
A server room that comfortably supported ten virtualisation hosts in 2021 may require significant upgrades before supporting modern AI infrastructure.
Many Australian server rooms were originally designed around standard 10A circuits.
10A Circuit @ 240V has a theoretical maximum: 2.4kW
Recommended continuous load (80% rule): Approximately 1.9kW
15A Circuit @ 240V has a theoretical maximum: 3.6kW
Recommended continuous load (80% rule): Approximately 2.9kW
What does this mean in practice?
A dual-GPU AI server consuming 1.8kW may almost fully utilise an entire 10A circuit by itself. Adding a second similar server could require a dedicated 15A circuit or larger power infrastructure.
Before purchasing new servers, organisations should evaluate:
Before investing in new infrastructure, let our specialists assess your power, cooling and rack capacity to identify potential constraints and recommend the right solution.
Talk to a Power SpecialistCPU performance continues to increase, but the real change in infrastructure planning is the rise of GPU-accelerated workloads.
Many server rooms were designed around workloads that consumed a few hundred watts per server. Modern virtualisation hosts often exceed 1kW, while AI-focused systems can approach 2kW or more.
Power, cooling, and rack capacity are rapidly becoming critical design considerations rather than afterthoughts.
Organisations considering server refreshes, AI initiatives, VDI projects, or GPU deployments should evaluate their power and environmental infrastructure before new equipment arrives.
The question is no longer whether the server will fit in the rack.
Effective power planning for modern servers should now be considered a core part of every server refresh and infrastructure modernisation project.
The question is whether the rack, power, and cooling infrastructure can support the server.
Modern servers consume significantly more power
GPUs dramatically increase infrastructure demands
Power planning extends beyond the server
Plan for future growth, not just today’s requirements
The answer depends heavily on the server configuration and GPU selection.
A traditional virtualisation server may consume between 500W and 1,000W under normal load. An AI server equipped with one or more GPUs can consume anywhere from 1,000W to over 3,000W.
Understanding the actual workload is critical when planning power, cooling, UPS capacity, and rack space.
Sometimes, but not always.
Many mid-market server rooms were originally designed around infrastructure consuming 3kW–5kW per rack. Modern AI workloads can quickly push rack densities beyond 10kW and, in some cases, beyond 20kW.
Common limitations include:
Before deploying AI infrastructure, organisations should conduct a power, cooling, and capacity assessment.
Rack power density refers to the amount of electrical power consumed within a single rack, typically measured in kilowatts (kW).
Examples include:
Rack Type Typical Power Density
Traditional server rack (2020) 3–5kW
Modern virtualisation rack 5–10kW
GPU-enabled rack 1 0–20kW
Large AI training rack 20kW+
Higher rack density usually requires enhanced cooling, larger UPS systems, and more robust electrical infrastructure.
A standard Australian 10A circuit provides approximately 2.4kW of total power capacity.
For continuous operation, most designers limit utilisation to around 80%, leaving approximately 1.9kW of usable capacity.
Examples:
Actual capacity should always be validated against real-world power measurements and local electrical standards.
Yes. AI and GPU servers can consume several kilowatts each, significantly increasing UPS sizing requirements and reducing available battery runtime. UPS capacity should always be reassessed before deploying high-density AI infrastructure.
Organisations should consider 15A circuits when:
A 15A circuit provides approximately 50% more capacity than a standard 10A circuit and is increasingly common in modern server room deployments.
Almost all power consumed by a server ultimately becomes heat.
As a rough guide:
As power consumption increases, cooling requirements increase proportionally. Many organisations discover that cooling becomes the primary constraint before power capacity is exhausted.
A basic AI readiness assessment should review:
Many organisations discover that infrastructure upgrades are required before deploying AI workloads at scale.
The most common mistake is focusing exclusively on GPU specifications while overlooking the supporting infrastructure.
AI projects often begin with questions such as:
“Which GPU should we buy?”
In reality, successful deployments require consideration of:
The GPU is often the easiest part of the project. Supporting it reliably over the next three to five years is where most organisations face challenges.
There is no single answer.
Cloud AI platforms can offer rapid deployment and lower upfront costs, while on-premises infrastructure may provide advantages in
Many mid-market organisations are adopting a hybrid approach, using cloud-based AI services where appropriate while deploying selected workloads on-premises where data sensitivity, performance, or economics justify the investment.