The New Rules of Generator Reliability
Part 1: The AI Paradox - Why Traditional Testing Is Failing Data Centres
How the Uptime Institute defines "primary power" - and why traditional generator testing is failing us
Imagine being told that the utility grid is merely an "economic alternative," and that your fleet of diesel generators is your facility's primary source of power. That is precisely the reality defined by the Uptime Institute, the global authority on data centre infrastructure reliability.
For Tier III and Tier IV data centres - the highest standards of data centre resilience - generators aren't a backup plan. They are the plan. The grid is viewed as inherently unreliable; convenient when it works, but never something to bet the business on.
This paradigm has driven the industry toward extreme caution. For decades, operators have run their backup generators bi-weekly to monthly, with additional extended tests yearly. But in an era defined by the explosive growth of AI, surging diesel prices, and tightening emissions regulations, this "better safe than sorry" approach is rapidly becoming a liability – both for reliability and the balance sheet.
The AI Factory Scale Problem: A Parallel Fossil Fleet
The rise of AI has fundamentally altered the scale of the data centre industry. A traditional large data centre might carry 20 to 100 MW of IT load. Today's AI campuses are typically designed for 100 to 300 MW of critical power, with several proposals exceeding 400 MW. Reliability codes require backup systems equal to full facility load, plus a 15–20% safety margin. A 200 MW AI cluster, therefore, needs roughly 230 MW of diesel capacity sitting in its backyard.
We are effectively building parallel fossil fleets alongside our digital infrastructure. The numbers are staggering:
The global data centre generator market is projected to grow from $6 billion in 2023 to $12 billion by 2030.
By 2026, the large-bore diesel engine market for data centres is expected to exceed $40 billion.
In China alone, demand reached 5,600 - 6,000 units in 2025, with projections of doubling within 2–3 years.
Supply chains cannot keep up. Major manufacturers have order backlogs extending into 2027–2028, with lead times stretching beyond 18 months. The market has shifted from "scrambling for electricity" to "scrambling for diesel generators."
With thousands of new units coming online, even small reductions in testing hours translate into massive, cumulative savings in fuel and emissions.
The Irony of Traditional Testing: Short Runs That Actually Hurt
The Cummins white paper (Minimizing Generator Set Exercising) identifies a critical flaw in conventional testing wisdom: frequent, short, low-load runs often do more harm than good.
Traditional regimes typically involve:
No-load or low-load short runs (weekly to monthly, 10–15 minutes)
Yearly full-load runs (1–2 hours at 70–100% output against a load bank)
But here's why the first category is problematic:
Wet Stacking is an Old Fear
Historically, loaded operation was required to prevent "wet stacking" – incomplete fuel combustion caused by low exhaust temperatures. However, newer engine controls maintain better combustion temperature control, minimising this risk. With modern engines, there is simply no requirement to run monthly.
The DPF Trap
Diesel Particulate Filters (DPFs) trap soot until enough exhaust heat burns it off (regeneration). During engine start, an excess of soot accumulates on the filters. If an engine is started frequently but not run long enough or hard enough to regenerate, the filter will eventually clog.
In a real utility outage, a clogged filter causes excessive backpressure, potentially shutting the generator down exactly when it is needed most. Frequent short runs aren't just ineffective – they are actively detrimental to DPF health.
The SCR Illusion
Selective Catalytic Reduction (SCR) systems reduce NOx emissions but require the catalyst temperature to exceed a threshold to start working. During a 10–15 minute low-load run, the exhaust temperature typically stays below this threshold. The urea injection system won't even be actuated. Consequently, short exercise runs produce unregulated NOx emissions and don't actually test the SCR system.
AI Workloads Raise the Stakes
AI workloads exacerbate these testing issues. GPU-based clusters can transition from idle to full load within seconds, creating rapid load transitions that traditional testing regimes rarely replicate. Wabtec, for example, has conducted extensive testing of Tier 4 diesel engines specifically "challenged to meet aggressive step loads and response tests, which included AI/Machine Learning scenarios."
In high-frequency trading or large language model training, the transition from grid to generator is a moment of extreme vulnerability. If the generator fails to start on the first crank – potentially due to a DPF clogged by unnecessary testing – the financial and operational consequences are catastrophic.
The Permitting Paradox
While AI drives unprecedented demand for diesel generators, it also makes emissions permitting more difficult. A recent analysis found that using diesel generators to power AI could cause approximately 500 more premature deaths per year in the U.S. due to fine particulate matter pollution. Outdated permitting rules make cleaner backup power too slow to deploy.
Diesel remains, in 2025, "the only backup power option that fits the speed, risk tolerance, and regulatory structure of the AI buildout." Developers can commission an AI data centre in 18–24 months; transmission interconnection queues are backed up for years.
That frictionless path for diesel is precisely why optimising generator testing is no longer a nice-to-have – it is a permitting and operational necessity.
In Part 2, we break down the economics of fuel price volatility, the supply chain squeeze, the rise of renewable alternatives like HVO, and the specific maintenance strategies that allow operators to safely extend testing intervals to 6–12 months – saving millions in fuel and emissions.