Best UPS for a Home AI Server: Sizing for GPU Loads, Not Just Runtime
Most UPS buying guides size for home routers, internet gateways, and small-office servers—machines drawing 50–200W and needing a few minutes of runtime to shut down gracefully. A home AI server is not that machine. A single modern GPU rig pulls 500–800W sustained, and a dual-GPU build can draw 1000–1200W. More important: GPUs have active-PFC power supplies that demand a true sine wave inverter and enough capacity to absorb the current inrush when the cards spin up. A UPS undersized for GPU loads is not merely inconvenient—it trips under training runs and can corrupt your work in progress. Per LocalRig’s safety framework, undersized power delivery is a Persona-4 betrayal and belongs in the same category as a bad PSU.
This guide sizes UPS by load tier, explains why pure sine wave and VA rating matter, and gives you the math to buy once and stop worrying. It assumes you’ve already chosen a PSU; if not, read PSU for multi-GPU AI rigs first.
Core principle: VA, not watts—and pure sine wave is non-negotiable
When manufacturers quote a UPS capacity, they usually print both watts and VA. Ignore the watts. Buy based on VA, and only pick a UPS with a pure sine wave inverter.
VA (volt-amperes) is the UPS’s total output capacity. It accounts for both real power (watts, which do work) and reactive power (which your load draws but does not convert to heat). An active-PFC power supply has reactive impedance: it draws more apparent current (VA) than it converts to real power (W). If you size the UPS by watts alone, you’ll undersell the inverter and watch it trip under load.
Watts is what the UPS can deliver steadily to a purely resistive load. For a load like a heater, watts and VA are the same. For anything with active-PFC (modern 80+ Bronze and above PSUs, most energy-conscious hardware), they are not.
Pure sine wave vs stepped-wave: A cheap UPS with a stepped or modified-sine inverter saves money by approximating a true AC waveform in steps. It works fine for resistive loads—heaters, lights, fans. It does not work reliably with active-PFC. Active-PFC circuits sample the voltage waveform and sync their input current to it; a stepped approximation confuses the PFC controller, causing it to behave erratically or shut down. Many community reports describe the frustration: a UPS rated for 1500W trips repeatedly under a 1000W GPU rig drawing 800W at that moment. The reason is almost always a stepped-wave inverter combined with undersizing. Pure sine wave UPS units cost more, but they are the only option here.
Size for peak transient, not nominal draw. When a GPU powers on, its current inrush can spike 20–40% above steady-state for a few milliseconds. A UPS sized to nominal power will see this spike and interpret it as an overload. The safety margin is in the VA rating: buy a UPS that is at least 1.5× the sustained wattage of your rig.
Wattage anchors: single-GPU and dual-GPU reality
The homelab community sites these figures repeatedly, and they are estimates based on TDP + PSU overhead, not independently verified by LocalRig:
- Single RTX 5090 rig: ~575W nominal. With transient headroom, plan for 800–900W at the PSU input. Size the UPS for 900W absolute, which means a pure sine wave UPS rated 1200–1500 VA.
- Dual-GPU rig (2× RTX 4090, 2× RTX 5090, or mixed): 1000–1200W nominal. With headroom, the peak draw can touch 1500W. Size the UPS for 1500W, which means a pure sine wave UPS rated 2000–2500 VA.
- Mini-PC with embedded GPU (Intel Arc, AMD Ryzen AI): ~300W nominal. Still needs pure sine wave, and oversizing does not hurt. A 1000 VA UPS is overkill but reliably safe.
These are rule-of-thumb estimates. For your exact rig, use the PSU calculator to model the actual components.
The UPS sizing formula
To convert watts to VA:
VA = watts ÷ power factor
For active-PFC PSUs, the power factor is typically 0.7–0.85, depending on the PSU and load. Use 0.75 as a planning estimate.
Example: A single-GPU rig with a 750W PSU.
- Nominal draw: ~600W.
- Peak draw (transient headroom): 600W × 1.5 = 900W.
- VA needed: 900W ÷ 0.75 = 1200 VA.
A 1200 VA pure sine wave UPS is the minimum; a 1500 VA unit adds extra margin and costs little more.
Master comparison table: UPS by load tier
| Load Tier | Workload | Sustained W | Peak W | Recommended VA | Example UPS | Pure Sine? |
|---|---|---|---|---|---|---|
| Mini-PC | Intel Arc 380M, AMD Ryzen AI | 250–300 | 350–400 | 600–1000 | Cyber Power CP1000PFCLCD | Yes |
| Single GPU | RTX 5090 / RTX 4090 single rig | 600–800 | 900–1000 | 1200–1500 | APC Back-UPS Pro 1500VA; Cyber Power CP1500PFCLCD | Yes |
| Dual GPU | 2× RTX 4090, 2× RTX 5090, or mixed | 1000–1200 | 1500–1600 | 2000–2500 | APC Smart-UPS 3000VA; Cyber Power CP3000PFCLCD | Yes |
All recommendations are pure sine wave; anything else risks nuisance trips and PSU damage.
The picks, by buyer constraint
Single-GPU rig (600–800W) on a modest budget: Cyber Power CP1500PFCLCD
The Cyber Power CP1500PFCLCD is the community standard for single-GPU builds. It’s rated 1500 VA / 900 W continuous (the split reflects the active-PFC reality), has a pure sine wave inverter, and ships with USB/serial monitoring. At roughly $200–$250 new, it’s a sensible price for a pure sine wave UPS in this capacity class.
Runtime: Cyber Power publishes ~10 minutes at half load (450W) and ~3–4 minutes at full load (900W). Expect 2–3 minutes at your actual GPU peak draw. Use it to gracefully shut down, not to continue working through the outage.
Check Cyber Power CP1500PFCLCD pricing on Amazon →
Single-GPU rig with monitoring & longer support: APC Back-UPS Pro 1500VA
The APC Back-UPS Pro 1500VA is a step up in build quality and software support. It includes APC’s PowerChute UPS monitoring software (Windows, Linux, macOS) so you can safely shut down remotely if needed. Pure sine wave, USB/network connectivity, and a track record of reliability in homelab setups.
Runtime at full load is similar (~3–4 minutes), but the software integration and longer warranty (3–5 years depending on variant) make it worth the premium if you want to integrate UPS status into your shutdown automation. Prices typically run $250–$350 new.
Check APC Back-UPS Pro 1500VA pricing on Amazon →
Dual-GPU rig (1000–1200W): Cyber Power CP3000PFCLCD or APC Smart-UPS 3000VA
For a dual-GPU rig, you need 2000–2500 VA minimum. Two solid picks:
Cyber Power CP3000PFCLCD (3000 VA / 1500 W): The budget-friendly option. Pure sine wave, USB/serial, and rated for ~5–6 minutes at half load, ~1–2 minutes at full load. Typically $350–$450 new. Good value and enough headroom to absorb GPU transients.
Check Cyber Power CP3000PFCLCD pricing on Amazon →
APC Smart-UPS 3000VA (3000 VA / 1500 W): The enterprise-grade choice. Better power conditioning, longer warranty, and the SmartSlot modular expansion option if you want to add monitoring or network connectivity later. Costs $500–$700 new, but it’s built for 24/7 uptime and the MTBF is industry-leading. If you plan to run the rig heavily or expand it, this is the safer choice.
Check APC Smart-UPS 3000VA pricing on Amazon →
Multi-GPU or high-wattage setup (>1500W): APC Smart-UPS 5000VA or larger
For builds exceeding 1500W peak draw—three or more GPUs, or a 1000W PSU with auxiliary equipment—move to 5000 VA or higher. The APC Smart-UPS 5000VA is the industry standard. It’s large, expensive ($1000+), and probably overkill for a home setup, but it handles the load safely and will outlast the hardware. Consider it only if you’re running a small cluster or a production-grade homelab. Otherwise, the dual-GPU tier covers you.
Check APC Smart-UPS 5000VA pricing on Amazon →
Runtime: the honest math
Every UPS manufacturer publishes a runtime curve showing how long the unit supplies power at various load levels. The curves assume a purely resistive load (like a space heater). GPU rigs do not behave that way.
At half load, expect near the published curve. A UPS rated 10 minutes at half load will give you roughly 10 minutes if you’re drawing 450W from a 1500 VA unit.
At full load, expect 50–60% of the published curve, sometimes less. The reason: active-PFC and transient dynamics eat into the time. A UPS publishing 3 minutes at full load will realistically give you 1.5–2 minutes with your GPU rig drawing peak current.
The honest use case: A UPS is not a battery backup to keep your training run alive. It is a graceful shutdown tool. When you lose power, the UPS buys you 1–3 minutes to save your checkpoint, shut down cleanly, and avoid filesystem corruption. That’s what you’re paying for, and that’s what you’ll get. Do not size a UPS expecting to work through a long outage; use it to stop safely.
If you need long runtimes, add external battery packs (some Cyber Power and APC models support modular expansion). Alternatively, accept the trade-off and invest in a generator for extended power loss scenarios.
Why pure sine wave is non-negotiable
Active-PFC power supplies are standard on all modern 80+ Bronze and above PSUs. They draw current in phase with the AC voltage, which is energy-efficient and reduces waste heat. The catch: they are sensitive to the waveform shape. When they see a stepped or modified-sine wave inverter, they can misfire in several ways:
- Nuisance shutdown. The active-PFC controller interprets the stepped edges as voltage transients and shuts down as if a fault occurred.
- Harmonic distortion. The PSU’s input filter reflects the stepped waveform back into the UPS, causing the inverter to oscillate or overheat.
- Reduced efficiency and higher losses. The PSU works harder trying to correct for the approximation, generating more heat and drawing more current, which can trip a UPS that would otherwise be fine.
Cheap UPS units (often found at box stores, labeled “600VA” or “850VA”) use stepped or modified-sine inverters to save cost. They work great for routers, monitors, and fans. They fail repeatedly on GPU rigs. The cost difference between a stepped-wave and a pure sine wave UPS in the 1500 VA range is roughly $50–$80. That is a bargain for the safety and reliability you gain.
Integration: shutdown automation
A UPS is more useful when it integrates with your shutdown logic. Most modern UPS units ship with USB or network connectivity and support software that monitors battery level and initiates a graceful shutdown if the outage lasts beyond a threshold.
For Linux-based setups (your AI server is likely Linux), apcupsd (for APC units) and nut (Network UPS Tools, broader support) are standard. Both can monitor the UPS, estimate remaining runtime based on load, and trigger a shutdown script when battery hits a threshold (e.g., 20% remaining). This automation is the difference between a corrupted model checkpoint and a clean state file.
For Windows or macOS, APC’s PowerChute and Cyber Power’s software provide similar functionality. Set it up once and do not ignore the alerts.
Who this is NOT for
This guide is for home AI enthusiasts and small-team setups running local inference or fine-tuning on consumer hardware.
It is the wrong guide if:
- You are running a data center or production cluster. Hire a UPS and power infrastructure consultant. That budget and scale require three-phase power, generator coordination, and redundancy this guide does not cover.
- You are trying to squeeze an extra 30 minutes of runtime by undersizing the UPS and oversizing the battery. It does not work. The UPS inverter has a VA rating; exceeding it will trip the unit or damage it. Buy the right VA for the load, then expand runtime with modular batteries if needed.
- You are running on a budget so tight that you cannot afford a pure sine wave UPS. Do not buy a cheap stepped-wave unit hoping it will work. It will fail under GPU load. Either stretch the budget to a true sine wave UPS or delay the GPU rig until you can afford the full system safely. A bad UPS is worse than no UPS; at least without a UPS you know you are vulnerable.
- You expect the UPS to keep your training run alive through a long outage. It will not. Use it as a graceful shutdown tool, not an extended-runtime battery. If long runtimes are critical, add external battery packs or a generator.
Bottom line
A home AI server needs a UPS sized for peak GPU transient current, with a pure sine wave inverter, and rated in VA (not watts). Size at 1.5× the PSU’s wattage converted to VA, and plan for 2–3 minutes of real runtime at full GPU load. The Cyber Power CP1500PFCLCD is the community standard for single-GPU builds; the Cyber Power CP3000PFCLCD or APC Smart-UPS 3000VA handles dual-GPU. Do not cheap out on the UPS and then blame the GPU or PSU when nuisance shutdowns ruin your work. Buy the right tool for the load, integrate monitoring, and rest easy knowing a power loss will shut you down safely, not corrupt your data.
For the full context on power delivery in a home AI rig—PSU sizing, power conditioning, and the rest—see PSU for multi-GPU AI rigs, the PSU calculator, and the server rack guide if you’re building a larger setup.