Generator and Fuel Storage Sizing for AI Data Centers

Every AI data center has a number that never appears on a datasheet: how long it can run with the utility gone and the fuel trucks stuck in traffic. For most of the last decade that number was treated as a compliance checkbox - a day tank, a 24-hour runtime, done. AI has changed the economics behind it. A single hall can now draw 10 to 50 MW, grid interconnection queues run three to seven years in the busiest markets, and a generator that used to be an emergency asset is increasingly the thing that lets a project take load at all.

That shift makes fuel a capacity decision, not a facilities footnote. And it exposes a sizing method that is routinely wrong in the same three places: the load basis ignores PUE, the burn rate is taken at 100% load, and the tank is sized on nameplate instead of usable volume. This article walks through a five-step method that avoids all three, and finishes with a worked 10 MW example you can copy into a spreadsheet.

Why AI halls break the old sizing rules

Three properties of AI load change the answer.

Density. A traditional enterprise rack draws 5 to 15 kW. A liquid-cooled AI rack draws 60 to 130 kW, and an NVL72-class rack can approach 140 kW continuous with transient peaks well above that. The facility load concentrates into fewer, larger blocks, which means the generator fleet sees fewer steps and larger ones.

Load shape. Training load is a high plateau with a sawtooth on top; inference load follows a diurnal envelope and can swing between idle and near-peak within seconds. Neither is the steady 70% load that legacy sizing assumed. Generator step-load acceptance, alternator sizing and fuel burn at part load all depend on this shape.

Duration risk. Grid queues, wildfire-related public safety shutoffs, and storm events have stretched real outages beyond the 8-to-24-hour window that many plants were designed for. A 48-hour tank is no longer generous in a market where the utility cannot promise a firm return date.

The practical consequence: fuel autonomy has become a differentiator in leases and in project finance. A hall that can ride through 72 hours without a refuel is a different product from one that needs a truck at hour 12.

Step 1: Build the load basis, not the nameplate

Start from the IT load the customer actually contracted, then add the facility overhead. The conversion is PUE, and for a modern liquid-cooled AI hall it typically lands between 1.15 and 1.35 depending on climate, cooling architecture and part-load behavior.

Two adjustments are easy to skip and expensive to miss.

Future phases. A generator yard is civil work - pads, bunds, piping, cabling, permits - and it is far cheaper to build for phase 2 now than to retrofit later. Size the yard for the ultimate build, even if you only install phase 1 tanks.

Non-IT loads that ride the generator. Fire pumps, chillers on emergency power, UPS charging after a long outage, lighting, controls and security all sit on the standby bus. UPS recharge in particular can add 10 to 20% of nameplate for the first hours after a transfer, exactly when the fuel budget is tightest. Add these explicitly rather than burying them in a margin.

Step 2: Configure the generator fleet around the load step

Generator count is a redundancy decision before it is a fuel decision, but the two interact: the number of running units, their individual loading, and the spare capacity after one failure all change the burn rate.

Common configurations:

• N+1. One spare module beyond the peak requirement. Lowest cost that still tolerates a single failure; the standard for enterprise and most colocation.
• 2N. Two independent power paths, each capable of carrying the full load. Higher capital cost, but concurrent maintainability without derating the hall.
• N+2 or distributed N+1. Used where maintenance during peak load is a contractual requirement, or where the fleet is split across separate switchgear rooms.

Three engineering constraints shape the module size:

Part-load efficiency. Diesel engines are happiest between roughly 60% and 80% of rating. Below 30% for extended periods, wet stacking and cylinder glazing become real risks. If your fleet will run at 35% for most of an outage, the modules are too large.

Step-load acceptance. When the utility drops, the UPS carries the hall until the generators take load. The size of that first block - often 50% to 60% of the running fleet rating, and sometimes more in AI halls with large UPS systems - must be inside the manufacturer's published step-load capability. Ask for the transient response curve, not just the continuous rating. ISO 8528-5 performance class G3 is a reasonable floor for data center duty.

Alternator sizing for rectifier load. UPS and server rectifiers produce harmonic current that heats the alternator beyond what a linear-load rating assumes. Many specifications oversize the alternator by 20% to 30% relative to the engine for rectifier-heavy facilities; confirm the actual derating with the manufacturer rather than applying a rule of thumb blindly.

For the worked example, a 13 MW facility load with N+1 lands on five 4.5 MW units: 22.5 MW installed, 18 MW available with one unit out, and four running units at about 72% each.

Step 3: Get the burn rate from the engine curve

Fuel consumption is not linear with load. The engine's specific fuel consumption (SFC, in grams per kWh) rises at part load because parasitic losses - cooling fans, oil pumps, friction - stay roughly constant while useful output falls.

That 0.25 L/kWh figure is a full-load number. It is the one most spreadsheets use, and it is the one that understates consumption in real operation.

Table 1 - Representative part-load fuel penalty for modern industrial diesel engines. Use the manufacturer's curve for final sizing; older engines run 5% to 10% higher across the board.

Two site corrections matter as well. Altitude and high ambient temperature reduce generator output, so a unit that is rated 4.5 MW at sea level and 25 °C may carry meaningfully less on a hot plateau. The load stays the same, the derated unit runs at a higher percentage of its available capacity, and fuel per kWh moves accordingly. Apply the manufacturer's derating curve before you pick the module size - not after.

Finally, budget the fuel burned in testing. NFPA 110 requires routine exercise; a monthly 30-minute run plus an annual load test is normal. On a large fleet this is a real annual quantity, and it also means the fuel is being turned over, which is good for quality.

Step 4: Decide autonomy hours before tank size

Autonomy is a policy number, not an engineering output. It comes from code, the authority having jurisdiction (AHJ), the utility's restoration record, and the owner's risk appetite - in that order.

NFPA 110 expresses autonomy as a Class: Class 48 means 48 hours at the system's design load, Class 72 means 72 hours, and so on. The standard also defines Type (transfer time; Type 10 means power restored within 10 seconds) and Level (Level 1 for life-safety-critical systems). A data center is usually specified as Level 1, Type 10, with the Class set by the owner and accepted by the AHJ.

Table 2 - Autonomy bands and the logistical assumption behind each. The right answer is the band whose assumption your site can actually meet.

The refuel contract is part of the calculation, not a substitute for it. Ask three questions before you rely on it: How many trucks does the supplier have within 100 km? What is their guaranteed response time during a regional event, when every other site is calling? And can a truck physically reach the tank fill point with the yard flooded, snowed in, or blocked by emergency vehicles?

Step 5: Convert hours into usable tank volume

Nameplate tank volume is not usable volume. Tanks need ullage for expansion, a heel to keep water and sediment away from the suction, and space for the fill and vent. In practice, 85% to 90% of nameplate is a fair usable figure for a well-designed tank; 80% is conservative and safer for small tanks.

The 1.15 multiplier is the same reserve used in most fuel-system specifications: it covers the unusable heel, minor consumption variance and the fact that you never want to run the tank to the bottom during an event.

Day tanks are a separate requirement, not a substitute. NFPA 37 and NFPA 30 govern the arrangement; a day tank is typically 100 to 300 gallons per generator and exists to shorten the suction run and provide a local buffer. It does not reduce the bulk storage requirement.

Worked example: a 10 MW AI hall

Take a 10 MW IT load with a 1.30 PUE, N+1 generation, and a 48-hour autonomy target.

Table 3 - Worked sizing for a 10 MW IT hall at 1.30 PUE, N+1, 48-hour autonomy. Numbers assume 0.256 L/kWh at 72% load and 90% usable tank volume.

Two observations from the arithmetic. First, the 48-hour tank is about 185,000 liters - roughly two 25,000-gallon above-ground tanks, which is a serious piece of civil infrastructure, not a skid. Second, the difference between 48 and 72 hours is another 50% of tank farm, which is why autonomy decisions get made early.

For quick scoping before you have engine curves, the table below gives order-of-magnitude numbers at full load. Treat them as a starting point for tank farm area, not for procurement.

Table 4 - Scoping table at 0.25 L/kWh. Part-load operation and fleet configuration will move these numbers by 5% to 15%.

Fuel quality is a sizing variable

A tank sized for 72 hours is worthless if the fuel in it will not burn at hour 60. Diesel degrades: oxidation forms sediment, condensation adds water, and microbial growth produces sludge that blocks filters and injectors. In an outage, contaminated fuel is one of the most common reasons a generator starts and then trips under load.

Design for quality from day one:

• Specify ULSD to ASTM D975 and record the delivery certificate for every load.
• Slope the tank to a sump and draw from above the water layer, not the bottom.
• Install fuel polishing with filtration and water separation on a timed schedule; this extends practical storage life from 6 to 12 months toward 24 to 36 months.
• Test quarterly for water, microbial contamination, particulates and flash point. Test after any long idle period.
• Treat biodiesel as a different fuel. B20 has shorter storage life, worse cold-flow behavior and OEM-specific approval limits. If you use it, size the tank and the polishing regime for it.

Cold climates add a second layer: cloud point and cold filter plugging point must match the local winter design condition, and lines, filters and day tanks may need heat tracing. A tank that gels is a tank that does not exist.

Compliance checklist

Fuel storage triggers a different regulatory stack from the electrical side of the plant. For a US project, the short list is:

• EPA SPCC (40 CFR 112). Aggregate above-ground oil storage over 1,320 gallons, or any single container over 660 gallons, requires a Spill Prevention, Control and Countermeasure plan. Most data center tank farms are well past this threshold.
• Secondary containment. Sized to 110% of the largest tank, with an impermeable liner, a drain valve that stays closed, and a way to remove accumulated rainwater.
• NFPA 30. Tank listing, setbacks, venting, fill connections, and fire protection for the tank farm.
• NFPA 37. Installation of the stationary engines, including exhaust, combustion air and fuel piping.
• NFPA 110. Class, Type and Level for the emergency power supply system, plus the testing and maintenance regime.
• Local fire code and AHJ. Setbacks, permits, signage, and sometimes a maximum single-tank volume. Check before the yard layout is frozen.
• Leak detection and overfill protection. Level transmitters, high-high alarms and automatic fill shutoff are increasingly required and always worth specifying.

Seven sizing mistakes to avoid

1. Sizing on nameplate. Use actual load and the engine curve, not the generator rating plate.
2. Forgetting PUE. The IT number is not the facility number; cooling and losses are real load.
3. Using full-load L/kWh at part load. The penalty is 12% at 50% load and 43% at 25%.
4. Counting the spare as capacity. N+1 means the fleet survives one failure, not that the spare burns fuel. Size the tank for the running units carrying the real load.
5. Treating nameplate volume as usable. Plan on 85% to 90%, and keep the 15% reserve in the calculation.
6. Ignoring the refuel plan. A 24-hour tank plus a 36-hour delivery is a 24-hour site, not a 48-hour one.
7. Designing the yard for phase 1 only. Civil works are cheap now and expensive later.

Buyer's checklist

Before you issue a purchase order for generators or tanks, confirm you have answers to these:

• IT load, PUE and facility load, by phase, with the growth assumption stated.
• Redundancy architecture: N+1, 2N, or distributed, and whether maintenance at peak is required.
• Module size and count, checked against the 60% to 80% efficient band and the step-load curve.
• Derating for site altitude and design ambient temperature.
• Autonomy class in hours, agreed with the AHJ and the insurer.
• Engine-specific fuel curve at the expected load factor, in L/kWh and g/kWh.
• Usable tank volume, containment volume and the tank farm footprint.
• Refuel contract terms, including response time during a regional event.
• Fuel polishing, testing and water removal scope.
• SPCC applicability, NFPA 30/37/110 compliance and local permit path.

Bottom line

Generator and fuel sizing for AI halls is a chain of five conversions: IT to facility load, facility load to fleet configuration, fleet configuration to burn rate, burn rate to autonomy, and autonomy to tank volume. Each step has one common failure mode - using a nameplate number where an operating number belongs. Get the operating numbers right and the tank farm, the yard layout and the refuel contract all fall into place. Get them wrong and the site discovers the gap at the worst possible moment: hour 13 of an outage that was supposed to last a day.

Sizing a hall for AI density? Browse the power distribution catalog or send us the load basis - we will work the generator and fuel numbers with you, from the engine curve outward.

References

1. NFPA 110, Standard for Emergency and Standby Power Systems - Class, Type and Level definitions, fuel supply and routine testing requirements. nfpa.org
2. US EPA, Spill Prevention, Control, and Countermeasure (SPCC) Rule - 40 CFR Part 112; applicability at 1,320 gallons aggregate above-ground storage. epa.gov
3. eCFR, 40 CFR Part 112 - Oil Pollution Prevention. ecfr.gov
4. NFPA 30, Flammable and Combustible Liquids Code - tank storage, venting and containment. nfpa.org
5. NFPA 37, Standard for the Installation and Use of Stationary Combustion Engines and Gas Turbines. nfpa.org
6. ISO 8528-5, Reciprocating internal combustion engine driven alternating current generating sets - Part 5: Generating sets - performance classes G1 to G4, including step-load behavior.
7. Uptime Institute, Tier Standard: Topology - redundancy and concurrent maintainability requirements that drive N+1 and 2N configurations. uptimeinstitute.com
8. US Department of Energy, Data Centers and Servers - efficiency and load-factor context for facility planning. energy.gov