Generator and Transfer Switch for AI Labs

Most AI labs do not need a generator. If your site goes dark at midnight, a well-sized UPS gracefully shuts down your servers, preserves the training checkpoint, and lets you restore in the morning. That is the honest starting position, and it matters because generators are expensive to buy, expensive to maintain, and quietly useless when you need them most — if nobody ran the monthly exercise test.

This article is for the minority of Kentino-tier deployments where downtime genuinely has a measurable cost: 24/7 inference serving customers on SLAs, multi-day training runs that cannot afford a hard stop, labs in areas with poor grid reliability. If that is you, read on. If it is not — if you have one server running experimental workloads — skip to the last section, buy a 10 kVA UPS, and stop there.

Cross-references: W04, P01–P05, I04, L05.

When a generator is the right answer

The decision is about cost of downtime versus cost of the solution. Generator ownership for a medium lab is roughly €800–2,000/year in maintenance alone, before the €10,000–40,000 install. You need to be honest about whether your outage risk justifies that.

A generator makes sense when at least one of the following is true with no practical workaround:

• Training runs longer than UPS runtime. If you are running 48-hour distributed training jobs, a 15-minute UPS buys you a checkpoint and restart, not continuity. But if restoring from checkpoint takes two hours of GPU time that runs you €150, and outages happen quarterly, the math starts to move. A 30-server lab with a 5-figure weekly compute bill calculates this differently than a single-node hobby setup.
• Inference SLA commitments. If you are billing customers for uptime — and they can measure it — downtime has a direct monetary penalty plus customer confidence damage. A generator plus UPS is the standard answer for anything above a 99.5% SLA target.
• Grid reliability at the site is genuinely poor. Some EU industrial parks and rural business estates see three to five one-hour outages per year. A UPS for graceful shutdown still makes sense; a generator makes sense when that frequency multiplies against a meaningful cost per hour.
• Regulatory or contractual requirement. Some regulated environments (medical-adjacent, defence, critical infrastructure supply chain) require N+1 power. This is usually specified, not a judgment call.

When a generator is the wrong answer:

• One server, experimental workloads, in an office building in Prague or another well-gridded EU city. Just use the UPS.
• Any lab where the recovery procedure is under ten minutes and data is safe. Checkpoint restart is faster and cheaper than a generator at this scale.
• Labs below roughly 10 kW IT load where a 100–200 kWh lithium battery system is now competitive on TCO — covered later in this article.

Sizing math: IT load is only half the number

The most common generator sizing mistake is forgetting cooling.

A lab running 25 kW of IT load is not a 25 kW generator load. The room's cooling system — a precision AC unit, a DX split, an HVAC unit — is drawing additional power to reject that heat. At a typical data-room PUE of 1.4–1.6, a 25 kW IT load means 35–40 kW total facility load. Add another 20% headroom on top of that for inrush, load growth, and derating.

The formula:

P_genset = (P_IT + P_cooling) × 1.2
P_cooling ≈ P_IT × (PUE − 1)
PUE ranges 1.3 (well-cooled room) to 1.8 (small office retrofit)

A practical example for a medium lab with one 8-GPU K-AI server plus one 4-GPU server, cooling, and networking:

IT load:   8-GPU node 4.5 kW + 4-GPU node 2.4 kW + network/storage 0.6 kW = 7.5 kW
Cooling:   7.5 × (1.4 − 1) = 3.0 kW (well-managed rack room)
Total:     7.5 + 3.0 = 10.5 kW
Genset:    10.5 × 1.2 = 12.6 kW → buy a 15 kW unit

Never undersize the generator to cut capital cost. A generator running consistently above 85% load will overheat, wear faster, and fail during the outage you bought it to cover. The 20% headroom is not vanity — it is the margin between a generator that starts and a generator that starts and then stops two hours later.

Genset class by lab size

These are EU ex VAT approximate street prices for industrial-grade diesel gensets with canopy (acoustic enclosure). Prices vary by manufacturer tier, local distributor margins, and whether the site needs a special quiet enclosure. Budget another 30–60% on top for install labour, electrical connection, ATS, fuel tank, commissioning, and first-service.

Fuel type comparison

For European AI lab installs, diesel dominates — not because it is technically optimal in all cases, but because it is the most reliably serviceable, the most available, and the most predictable in regulatory terms. Natural gas and propane have genuine niches.

Diesel is the default for most EU lab installs because: tank on-site means no dependency on continuous gas supply; Euro Stage V-compliant diesels are well-distributed through Czech distributor networks (Cummins, MTU, and Kohler-SDMO all have CZ coverage); and diesel maintenance is well-understood by any industrial generator service firm.

Natural gas makes sense when the building already has a gas mains connection with adequate flow rate, when the site is urban enough that trucked diesel delivery is inconvenient, or when local regulations or landlord restrictions prohibit a diesel tank. The trade-off: a gas mains interruption simultaneously with a grid outage — not common, but real.

Propane is the niche answer for rural or isolated sites where grid reliability is poor and gas mains do not reach. LPG's indefinite shelf life beats diesel's ~12-month fuel rotation requirement. The tank permits are similar to diesel in most Czech and EU contexts.

Lithium battery alternative for sub-50 kW labs

In 2026 this is genuinely relevant. A 100–200 kWh lithium iron phosphate (LFP) BESS (Battery Energy Storage System) installed at a site with reasonable grid quality and some PV solar can now compete on 10-year TCO with a diesel generator + UPS combination for labs below 50 kW IT load and 4–8 hour runtime targets.

The BESS advantage is operational simplicity: no fuel rotation, no monthly exercise run, no exhaust permit, no noise, no diesel spill risk. The BESS disadvantage is that 200 kWh is a ceiling — if the outage runs 12 hours, you are done. A diesel generator runs as long as you have fuel.

The BESS route makes sense when: IT load is below 50 kW; target runtime is 4–8 hours (not "indefinite"); the site has or plans grid-tied solar; local noise or emissions restrictions make diesel unattractive; or the team does not want to manage a fuel and maintenance programme.

Diesel still wins when: runtime beyond 8 hours is required; IT load is above 50 kW; the site is rural with poor grid and no solar; or capital constraint favours lower upfront cost (a 20 kW diesel genset at €8,000 is still a lower entry point than a 100 kWh BESS).

Transfer switches: MTS, ATS, and CTTS

A transfer switch is the device that moves the load between utility grid and generator. Choosing the right type is not complicated for AI labs, but the distinctions are worth knowing.

Manual transfer switch (MTS). A physical rotary switch that an operator throws by hand. Cheap, simple, code-compliant for some applications. Not appropriate for an AI lab — by the time a human arrives and throws the switch, the UPS is drained and the servers have crashed. MTS is for hobbyists and construction-site power.

Automatic transfer switch (ATS). The standard for any serious installation. The ATS continuously monitors the utility feed voltage and frequency. On loss of utility — or voltage/frequency excursion beyond preset thresholds — the ATS sends a start signal to the generator, waits for stable output (typically 10–30 seconds depending on the genset's warm-up time), then transfers the load. When utility returns, the ATS transfers back (after a configurable delay to confirm stability, typically 3–5 minutes) and signals the generator to cool-down stop.

Standard ATS + UPS topology: utility or generator feeds ATS; UPS bridges the open-transition gap and genset start window; servers see nothing.

ATS switchover is an open-transition ("break-before-make") event: the load is de-energised for a brief moment — typically 50–150 ms — while the contactor opens the utility connection and closes the generator connection. This is precisely why the UPS exists upstream of the ATS. The UPS sees the brief drop, stays on battery, the servers never notice. Without the UPS, that 100 ms break would cause a hard-reboot on every server.

Closed-transition transfer switch (CTTS). A CTTS synchronises the generator output to the utility waveform before transferring, then briefly parallels both sources (typically under 100 ms) before disconnecting utility. The result is a truly zero-break transfer — the load sees no interruption at all. This is what hospitals and large datacenters use when they cannot tolerate even a brief disturbance.

For AI labs with a properly sized UPS, CTTS is overkill. The UPS already provides zero-break protection at the server side; the 100 ms ATS dead-time never reaches the load. CTTS adds meaningful cost (30–60% premium over ATS) and complexity, and requires utility approval to operate in parallel with the grid, even briefly. Unless a specific load on your site cannot tolerate the 100 ms ATS gap even with UPS upstream, stick with standard ATS.

UPS–generator coordination: the bridge problem

The 10–30 seconds between utility loss and generator on-load is the critical window. If the UPS runs out of battery before the generator stabilises and the ATS transfers, every server hard-reboots. Sizing this correctly is not optional.

The chain of events on utility loss:

T=0 s    Utility drops below threshold
T=0 s    UPS detects, switches to battery — servers see nothing
T=2 s    ATS starts generator (after 2–5 s intentional delay to avoid nuisance starts)
T=12 s   Generator reaches stable voltage and frequency
T=14 s   ATS transfers load — brief 100 ms open-transition event
T=14 s   UPS sees utility restored (generator = new utility), returns to float
T=15 s   UPS battery at 97–99% — barely used

The UPS bridge target: 2 minutes minimum at full rated load, not 30 seconds. Why the buffer? Because generators do not always start on the first crank. A cold diesel in a Czech winter below −10 °C may need two or three start attempts, adding 30–60 seconds. The ATS typically allows two or three start attempts before alarming. The UPS must hold through all of them.

One coordination detail that causes trouble in practice: UPS charger walk-in. When the ATS transfers the load to the generator, the generator suddenly sees not just the IT load but also the UPS charger trying to rapidly recharge batteries that just discharged. A UPS charger on a modest generator can represent 5–15% of the generator's rated output — it needs to be factored into sizing. Most modern double-conversion UPS systems have adjustable charger walk-in settings. Configure this; it is not the default on most units.

Fuel storage: runtime math and fire code

A diesel generator running at 80% load consumes approximately 0.27–0.30 L per kWh of electrical output.

Most industrial gensets ship with an integral "base tank" or "belly tank" of 60–250 L. A 15 kW genset with a 100 L base tank gives you roughly 27 hours of runtime at 80% load — adequate for most grid events without a separate day tank.

Czech and EU fire code for diesel storage. The general EU regime (implemented in CZ via relevant CSN standards) applies: up to approximately 100–200 L of diesel stored outside a building requires no formal permit in most CZ municipalities, but placement, bunding, and separation distances from buildings apply. Over 200 L (and for any fixed installation above 50 L inside a building) requires a bunded containment structure capable of holding 110% of tank volume. Over 1,000 L typically triggers a fire category assessment and potentially a permit from the local fire authority (Hasičský záchranný sbor in CZ context). Large bulk tanks above 5,000 L are classified as hazardous substance storage.

For the medium AI lab case (15–50 kW genset with a 100–500 L day tank): outdoor placement with a double-walled or bunded tank, minimum 2 m clearance from the building facade, weatherproof canopy for the genset. No exotic permitting required, but verify with your local fire authority before install.

Diesel fuel stored in a day tank degrades in approximately 12 months. Fuel polishing (circulating through a filter) every 6–12 months extends life. The standard practice: annual fuel polish plus top-up with fresh diesel, monthly exercise run to burn through some inventory and keep the injectors live.

Concrete scenarios

Scenario 1: Small lab, ~5 kW IT, one K-AI 4-GPU server plus a robot charging dock.
Generator verdict: skip it. A 5 kVA double-conversion UPS (Eaton 9PX 5000 or APC Smart-UPS SRT 5000) gives 10–15 minutes for graceful shutdown. The probability of an outage long enough to justify a €10,000+ generator install is very low in an urban CZ location. UPS total: €2,000–3,500.

Scenario 2: Medium lab, ~10 kW IT (1× 8-GPU K-AI node + cooling + network).
Generator: 15 kW diesel with acoustic canopy, ATS, and 100 L day tank. ATS: standard open-transition, single- or three-phase matched to site power (P01). UPS: 10 kVA double-conversion lithium (Eaton 9PX 11 kVA or Vertiv Liebert GXT5 10 kVA) for 5-minute bridge runtime. Approximate installed cost including electrical work, ATS, fuel tank, and commissioning: €12,000–20,000. Annual maintenance budget: €1,000–1,500.

Scenario 3: Larger lab, ~30 kW IT (4× 8-GPU servers, cooling, storage, networking).
Generator: 50 kW diesel, three-phase ATS, 500 L day tank for roughly 40-hour autonomy at full load. UPS: 30 kVA modular double-conversion (Schneider Galaxy VS or Vertiv Liebert EXM) feeding the whole rack, 5-minute bridge runtime. The UPS is centralized (per P05's recommendation). Approximate installed cost: €30,000–50,000 all-in. Annual maintenance: €2,000–3,000. This is a serious installation that warrants a licensed electrical designer.

Generator manufacturers and EU sourcing reality

Cummins — strongest service network in Eastern Europe, CZ distributor presence is solid. Full Stage V range from 10 kW to megawatt class. Long lead times on bespoke configurations (8–16 weeks).

Kohler-SDMO — French-owned, widely distributed in EU. SDMO brand is the European industrial range. Available through Czech distributors.

MTU Onsite Energy — premium tier, Rolls-Royce Power Systems brand, strong presence in German-speaking markets and Czech industrial sector. Higher price point, appropriate for the larger (80 kW+) installs.

Broadcrown / FG Wilson — well-distributed in CZ through industrial equipment channels. Solid mid-market range.

For a lab-scale install (15–50 kW), Cummins or Kohler-SDMO through a local CZ distributor is the practical recommendation. Both provide local service contracts, which matters for the annual maintenance requirement.

Maintenance: the ongoing cost that surprises people

A diesel generator that sits in a canopy and never runs will not start when you need it. The maintenance burden is real and non-negotiable.

Monthly: exercise run under load (minimum 30% rated load for 30 minutes — "no-load" idling does not clear carbon or maintain the wet seals); visual inspection of coolant level, oil level, starter battery condition, fuel level.

Quarterly: fuel water separator drain; air filter condition check; battery load test (starter battery).

Annually: full load bank test at 100% rated load for 2 hours; oil and filter change; fuel polish or full drain and refresh; coolant condition test; governor calibration check.

Every 5 years / 1,500–2,000 hours: top-end inspection, valve adjustment, injector service, full tune-up.

This is part of the Kentino Wiki, a reference series on AI compute, robotics, and the systems that connect them. Comments and corrections welcome at info@kentino.com.