Stephen Lafaille, VP Business Development, Tecogen, says a shift in how power is allocated within data centres could unlock significant AI capacity without waiting for grid expansion.
For the past two years, the data center industry has been locked in an arms race centred on cooling innovation.
Liquid-to-chip, immersion, rear-door heat exchangers, direct-to-chip loops. Every conference panel, every vendor roadmap, every headline seems to orbit one central idea: cooling is the bottleneck and better cooling is the breakthrough.
But there’s an uncomfortable truth hiding in plain sight: We’re not short on cooling innovation. We’re short on usable power and we’re wasting a massive portion of it.
Today, data center operators routinely allocate 25–40% of total facility power to cooling systems, with many legacy or air-cooled environments clustering around 30%. This range is supported by industry analyses from organisations like Uptime Institute and the International Energy Agency.
In a world where AI workloads are smashing into grid constraints, that’s not just inefficient, it’s strategically indefensible.
If you’re dedicating nearly a third of your available power to cooling infrastructure, you’re not optimising your data center. You’re cannibalising it.
The real AI bottleneck is within your facility
Let’s get clear about what’s actually happening using the standard metric for the energy efficiency of a data center – PUE. By today’s standards, a data center with a PUE of 1 has zero energy wasted on cooling, lighting or power distribution and all the incoming energy is used to power IT equipment.
AI workloads, particularly training clusters and high-density inference deployments, are pushing facility power densities to levels the grid was never designed to support. This density increase is caused not only by individual racks drawing >100kW of power, but by the subsequent need for more powerful chiller systems to drive the mechanical cooling.
In a newly built data center, the optimal PUE is 1, but average operational PUE often varies from 1.1 to 1.3 with peaks as high as 1.6. Of that, the chiller system represents a majority of this peak ancillary electrical draw with 60-70% of that energy being consumed by the chiller’s compressors alone.
At the same time, utilities across North America report multi-year interconnection delays, with grid upgrade timelines often stretching 3–7 years depending on region and load requirements.
Operators are being told to wait. Wait for capacity. Wait for approvals. Wait for infrastructure that may not arrive on time or at all.
The industry’s blind spot: Optimising for peak times
The current innovation cycle is centred on improving cooling efficiency under normal operating conditions. That means more efficient chillers, advanced liquid cooling and chip thermal optimisation.
These innovations are real, but they are largely optimised around average PUE performance, not peak conditions. And that distinction is critical.
Modern data center infrastructure should be designed around the peak conditions of high ambient temperatures or high user demand because grid congestion and system strain occur precisely during these situational stress events.
Even the highest efficiency electric cooling system draws power from the grid precisely when grid stress intensifies and energy prices spike.
In other words, designing your data center with a high peak PUE (and chances are it is) means that, at this very moment, every kilowatt dedicated to compressors and fans is a kilowatt unavailable for compute.
A contrarian shift: Remove cooling from the grid
If peak grid draw is the bottleneck, then optimising for peak PUE should be the true test of a data center’s efficiency.
And that requires a shift, not incremental improvements.
What if the fastest way to unlock major AI capacity isn’t by improving efficiencies but by reallocating the peak load from the cooling system altogether?
This is where the concept of a fuel swap becomes powerful.
If you remove the chiller’s compressor work from the grid during peak conditions, you can shift 60-70% of the ancillary load to alternative energy sources. While this can be done now with waste heat-driven absorption chillers or by utilising fuel swapping chillers, only one option is prepackaged as a drop-in replacement of the existing system.
The result is far from marginal: 10-30% more power for critical IT without waiting on the grid.
The 300 kW opportunity per megawatt
Let’s quantify the impact. For a 1MW IT load with a peak PUE of 1.4:
• Total facility load = 1.4 MW
• Cooling and auxiliaries = 0.4 MW
By shifting cooling off the grid and on to alternate energy sources, it becomes possible to reclaim approximately 0.3 MW – based on industry-average cooling loads cited above.
That translates into a total increase of approximately 30% in new power that can be reallocated for IT capacity.
For a 100 MW facility – that’s 30 MW of additional compute capacity unlocked without new substations, lengthy permitting and multi-year delays.
This is not a marginal efficiency difference between chiller COP’s. This is a significant power reallocation data center operators can benefit from now.
The International Energy Agency projects that global data center electricity demand will more than double by 2030, driven by AI and hyperscale expansion, putting increasing strain on power grids already struggling to keep pace. Yet, most of the grid stress boils down to a few hundred hours annually.
Operators stuck between demand growth and infrastructure limits can suddenly open up numerous potential power-constrained locations with the power flexibility granted by fuel-flexible cooling.
Peak demand: The moment that drives operating costs
Cooling demand peaks when temperatures rise and IT loads are fully utilised. These same periods are typically also when electricity prices spike and utilities assign their highest charges.
Those peak hours don’t just shape infrastructure; they disproportionately shape annual operating costs.
Electric bills are not driven by energy consumption alone. They are influenced by time-of-use pricing, 15-minute demand peaks, capacity tags tied to system stress and volatility during extreme weather.
For many utilities, a small number of peak intervals determines a significant share of total annual energy spending.
A cooling system reliant entirely on grid electricity amplifies exposure to those cost drivers precisely when rates are the highest. Compounding the issue, a chiller’s compressors and fans must ramp up during these most expensive hours, further increasing both total kWh charges and peak kW demand charges.
By shifting cooling off-grid during these periods, operators can:
• Reduce overall exposure to peak pricing
• Flatten electrical demand
• Stabilise operating expenses
From an operational cost standpoint, fuel flexibility again gives data center operators an edge. Using natural gas strategically to drive the chiller’s compressors and fans can be done without disruption and creates the ability to optimise energy sourcing based on real-time economics.
That’s not just resiliency but structural control over cooling related operating costs.
The Financial Case: millions left on the table
Every additional kilowatt of IT load that supports AI model training, inference workloads and high-value compute contracts is a kilowatt that’s creating revenue.
Reallocating 300 kW per MW away from cooling and towards computing can translate into millions in annual revenue, depending on workload value and utilisation rates.
Meanwhile, continuing to rely on grid-powered cooling means:
• Paying for non-revenue-generating power
• Delaying deployment timelines
• Underutilising facility potential
Start reclaiming power now
The industry’s focus on cooling innovation isn’t wrong, but it’s incomplete because the real opportunity isn’t just more efficient cooling, it’s a better allocation of power.
By removing cooling from the grid, operators can:
• Unlock stranded capacity
• Accelerate AI deployments
• Increase revenue per megawatt
• Reduce reliance on uncertain infrastructure timelines
And they can do it now.
If 30% of your power isn’t generating revenue, why are you still treating it as a fixed cost instead of an opportunity?
Because in the AI era, the fastest way to scale isn’t always building more. It’s reclaiming what you already have.
