Diesel generators are the part of a data centre that neighbours actually experience: the monthly test runs, the fuel deliveries and the stacks along the fence line. They are also the part of the industry changing fastest and most unevenly. The hyperscalers are pursuing different pathways beyond conventional diesel; much of the rest of the market has yet to articulate one. For assessors, the first question is no longer “how many generators?” but “what actually needs backup, for how long and why?”
Why this matters
Backup generators are the part of a data centre neighbours hear and breathe, through test runs rather than blackouts. The world’s largest operators are pursuing different pathways beyond diesel while most of the market has yet to articulate one. Assessments that count generators instead of asking which loads genuinely need to survive an outage are approving 1985 defaults for 2030 facilities.
Google has estimated that more than 20 gigawatts of diesel generator capacity stands behind the world’s data centres, most of it idle for all but a few hours a year. Every megawatt of new data centre capacity has traditionally brought a little more than a megawatt of diesel with it, sized for the day the grid fails and tested weekly or monthly whether it fails or not. In Australian development applications the generator yard is routinely the largest source of the three impacts communities raise first: noise, diesel exhaust and fuel storage risk.
Diesel has survived every attempt to retire it for three unglamorous reasons. Its energy density is exceptional: a litre of diesel holds roughly 36 megajoules, which is why multi-day autonomy is physically possible at all. The volumes are still substantial. Generators carrying 100 megawatts burn roughly 25,000 to 30,000 litres an hour, depending on set efficiency and loading, so two days of autonomy is well over a million litres. A 300 megawatt campus at the same duration is a fuel farm rather than a tank compound. Nothing else on offer comes close, which is the reason it endures, though it also explains why fuel storage sits alongside noise and exhaust in the objections communities raise. Days of autonomy are a choice rather than a universal standard: Uptime Institute sets a minimum of twelve hours of fuel storage at N load for all Tiers, with anything beyond that requiring a risk-based assessment of the facility’s energy supply. The equipment is proven across half a century and every failure mode is understood. It is cheap per kilowatt of standby capacity, precisely because it does almost nothing for decades. Any replacement has to beat that combination or change the question.
Who is moving, to what, by when
The largest operators are pursuing different pathways away from conventional diesel, which tells you something important: the successor has not been settled.
Microsoft has committed to eliminating its dependence on diesel fuel for data centre backup by 2030. It demonstrated a 3 megawatt hydrogen fuel cell system with Plug in 2022 and in January 2024 completed a separate 1.5 megawatt demonstration with Caterpillar and Ballard at Cheyenne, Wyoming, running a simulated 48-hour outage at 1,855 metres in below-freezing conditions. It also uses renewable diesel in its Swedish region. Google is working to 24/7 carbon-free energy by 2030 and has replaced a diesel generator with a lithium-ion battery system backing 3 megawatts of live production load at St. Ghislain in Belgium, arguing that batteries earn their keep by supporting the grid while diesel only ever waits. AWS, under Amazon’s net-zero 2040 Climate Pledge, is converting its European generator fleet to hydrotreated vegetable oil (HVO), beginning with Ireland, Sweden and Oregon in 2023: a drop-in fuel change rather than a technology change. Meta, targeting net zero across its value chain by 2030, began piloting HVO at Clonee in Ireland in 2024.
Two things are worth noticing about those timelines. The near-term moves (HVO) change the fuel, not the machine; the structural moves (batteries, hydrogen) are still at single-site scale a decade after the commitments were made. Both facts deserve to be part of every assessment conversation.
Meanwhile the AI build-out has scrambled the picture from the other direction. In the United States, grid connection queues have pushed AI campuses into generating on site as primary power: xAI has deployed hundreds of megawatts of gas turbines, OpenAI’s Stargate site ordered 29 turbines of 34 megawatts each and solid oxide fuel cell suppliers signed billions of dollars in data centre contracts in early 2026 alone. When a facility brings its own power station, the backup question inverts: the grid becomes the backup and the generator yard becomes the front of house, with an entirely different emissions, noise and assessment profile.
Backup is not one question
The biggest analytical error in assessing backup power is assuming that megawatts determine resilience requirements. They do not. Workloads do. A facility carrying financial transactions or emergency communications has a fundamentally different continuity requirement from an AI training campus whose jobs can checkpoint and restart, whatever their relative size. Facility class is useful shorthand because it bundles a typical workload mix with a typical set of contractual constraints, so it is where an assessment can reasonably start. Four classes, four different answers.
Colocation carries the strictest and slowest-moving requirements: tenant contracts specify redundancy and autonomy, so the operator cannot unilaterally swap the philosophy. HVO is colo’s realistic near-term path because it changes nothing the contract cares about. Hyperscale cloud holds the opposite position: workloads and services can be distributed across availability zones and regions, so resilience can be designed partly at the software and network layer rather than entirely within the facility. That does not automatically mean lighter backup, since hyperscale facilities still carry substantial systems, but it creates architectural options that conventional colocation contracts do not permit. That is exactly why the battery and hydrogen experiments are happening at hyperscale first. Edge facilities are small enough that for some of them batteries plus network rerouting may be the whole answer; their backup is architectural, not mechanical. AI facilities split in two: training loads can checkpoint and restart, which argues for less backup, yet the US experience shows AI campuses installing more generation than anyone, as primary supply. That contradiction exposes the question sitting underneath the whole section. Should backup requirements follow the size of the facility or the criticality of the workloads inside it? A 300 megawatt training campus whose jobs can restart does not carry the same continuity requirement as a far smaller facility supporting payments, hospitals or emergency communications. From an assessment perspective, very large AI campuses increasingly resemble power stations with attached computing infrastructure, rather than conventional data centres.
At the far end of the spectrum sits the counterexample. Cryptocurrency mining carries no backup on the mining load itself, because that load is interruptible by design and earns demand-response revenue by switching off when the grid is stressed. In Texas, flexible loads of this kind now represent around a tenth of forecast consumption. That is proof that a very large computational load does not inherently require backup generation. What decides the answer is what the load is for.
The reversal is no longer theoretical. During grid emergencies through 2026 the US Department of Energy repeatedly authorised PJM and other grid operators to call on backup generation at data centres and other large-load industrial and commercial sites, as a last resort before firm load shedding. Equipment installed to protect the data centre was being called on to help protect the grid. Backup fleets are becoming grid assets, which is an argument communities and assessors should hear as a benefit when the technology is clean and a cost when it is not.
The technologies compared
| Technology | Autonomy and storage challenge | Local emissions | Lifecycle and operational characteristics | Standing in backup duty |
|---|---|---|---|---|
| Diesel generator | Days of autonomy, though at campus scale that means bulk fuel storage on site; deliverable anywhere | NOx and particulates at the fence line; ~2.7 kg CO2 per litre | Engines last 20-30 years at standby duty | The incumbent default |
| HVO (renewable diesel) | Identical to diesel; same tanks, same engines | Combustion emissions including NOx remain; local air-quality benefit depends on engine and operating conditions | Up to about 90% lower lifecycle greenhouse gas, though feedstock, certification and land-use impacts decide whether that claim holds | The pragmatic near-term move (AWS, Meta, Microsoft Sweden); supply chains still thin |
| Natural gas engines and turbines | Piped supply, no onsite storage; autonomy depends on the pipeline staying up | Lower particulates than diesel; NOx remains | Roughly a quarter less CO2 per unit of energy; upstream methane risk; mature, 20-30 years | Growing fast, but mostly as primary power for AI campuses rather than standby |
| Hydrogen fuel cells (PEM) | Very low volumetric energy density against liquid fuels; substantial onsite storage or delivery infrastructure required | Water vapour at point of use | Lifecycle depends entirely on how the hydrogen is made; stack life suits standby duty; refuelling logistics unproven at fleet scale | Demonstrated at 3 MW (Microsoft/Plug) and 1.5 MW (Caterpillar/Ballard); the long-term bet |
| Solid oxide fuel cells (gas-fed) | Piped supply | Lower NOx, no particulates | CO2 remains; high electrical efficiency; designed for continuous duty, stacks replaced within engine-lifetime timeframes | Primary power economics, not standby economics (Bloom’s 2026 contract surge) |
| Lithium-ion battery systems | Hours, not days; autonomy is the binding constraint | Zero onsite | Lifecycle follows the grid mix that charges them; 10-15 years with augmentation as cells fade | Proven at single-facility scale (Google St. Ghislain) |
| Long-duration storage (flow batteries, emerging chemistries) | Designed for many hours; footprint and cost rise with duration | Zero onsite | Long cycle life, low degradation; limited operating history at data centre scale | Not a diesel replacement yet, but the reason “hours, not days” should not be read as permanent |
| Linear generators | Piped or stored fuel; fuel-flexible including hydrogen and ammonia | Ultra-low NOx | CO2 follows the fuel; new technology, early fleet | A watching brief |
The table’s honest summary: nothing on it beats diesel on all three of density, cost and proof at once. Everything on it beats diesel on emissions somewhere. The choice is therefore a design decision that should be argued in an application, not a default inherited from 1985.
What an assessor should ask
An assessor should be able to get clear answers to six questions.
- What loads are genuinely critical? Which workloads must continue through an outage? Which can stop, shift or restart?
- What autonomy is required and can it actually be delivered? How many hours or days, on what evidence, given the facility’s redundancy architecture. Raw tank volume is not the same as usable autonomy: a fuel system is only as good as its least redundant pump, day tank or transfer path, which is why the Uptime minimum is twelve hours while still meeting the facility’s topology objective. Where the figure assumes tanker resupply, that assumption should be tested against the events most likely to cause the outage in the first place, since bushfire and flood close roads at exactly the moment the fuel is needed.
- Why this technology? What alternatives were considered, on what grounds was this fuel and backup architecture chosen over them?
- What will neighbours experience? What are the noise, NOx, particulate, fuel-storage and testing impacts at the nearest receivers, given that test runs rather than outages are what neighbours actually experience?
- Can the asset support the grid? Can storage, generation or flexible load provide demand response or other grid services rather than wait idle for an outage?
- What is the transition pathway? If diesel or another transitional technology is installed at opening, what happens across the facility’s 20 to 30 year life?
The March national expectations put efficiency and community impact firmly into Commonwealth assessment policy and forthcoming standards are set to turn those expectations into obligations. Backup power intersects directly with both: the efficiency with which a facility uses and manages energy, together with what its operation imposes on the surrounding community. It is also one of the few parts of a data centre where practice is being rewritten in public, with named dates, by the largest operators on earth. Australian assessments can simply ask for it.
The bottom line
Diesel earned its place through density, cost and proof and it will not vanish by decree. But the era of backup-by-default is ending: the largest operators are pursuing different pathways beyond conventional diesel, the AI build-out has turned generators into power stations and interruptible designs have shown that some facilities need no backup at all. The question for an Australian assessment table is no longer how many generators a facility has. It is which loads genuinely need to survive an outage, what the neighbours will hear and breathe while the backup is tested and when diesel’s replacement arrives on site. Applicants should be prepared to justify all three.
A topical commentary from the Digital Infrastructure Institute. For assessment practice on backup power, noise and air quality, see our State & Territory Frameworks and community information series.
Sources: Data Center Knowledge — Google: batteries can replace generators; Baxtel — Google replaces diesel at St. Ghislain; DCD — Microsoft 3MW hydrogen fuel cell backup; Amazon — renewable diesel (HVO) for European generator fleets; Data Center Knowledge — replacing diesel in AI-scale data centres; US DOE §202(c) emergency orders, 2026