What PSU Do You Need for a Local AI Rig? (August 2026)

Bottom Line

• Single GPU up to a 3090/4090 (350-450W): 850W ATX 3.1. Real headroom, no drama.
• RTX 5090: 1000W ATX 3.1 with a native 12V-2x6 cable, minimum. The card’s board power is 575W and independent testing caught spikes of 901W in under a millisecond. ATX 3.1 units are rated to absorb excursions up to 200% of GPU power — that rating is the entire reason 1000W works.
• Dual GPUs (2x 3090, or any pair): 1200W. Two cards is not twice the average draw problem, it is twice the spike problem.
• Prices as of August 2026: PSUs are one of the only components the DRAM shortage has not touched. Do not economize here to fund an overpriced GPU — this is the cheap part of the build right now.

Why AI Duty Cycle Is Different From Gaming

A gaming session is bursts: heavy scenes, menu screens, loading pauses. Every mainstream PSU calculator assumes that shape.

A local inference box does two other things instead:

1. Sustained load for hours. A long agent run or an overnight batch keeps the GPU near its power limit continuously. Heat soak in the PSU is a duty-cycle problem, not a peak-watts problem. This is where a unit that “meets spec” on paper gets loud, hot, and old before its time.
2. Millisecond transients. Modern GPUs swing power draw far above board power for sub-millisecond windows. Igor’s Lab and TechPowerUp both measured the RTX 5090 spiking to 901W — roughly 57% above its 575W rating. Your wattmeter never sees it. Your PSU’s over-current protection does, and on an older unit the result is a hard shutdown mid-run.

The ATX 3.x standard exists because of point 2: a compliant unit must ride through GPU power excursions of up to 200% of rated GPU power. An ATX 3.1 850W survives spikes that trip a 2019-era 850W. Same number on the box, different unit.

The Ladder: One PSU Per Build Tier

The one non-obvious row is the first. An 850W unit on a 180W card looks wasteful. It is not: a PSU running at 20-40% load is at peak efficiency and near-silent, PSU prices are flat while everything else in the build inflated, and the next GPU you buy drops straight in. Buying the PSU for your next card is the cheapest upgrade insurance in the whole machine.

The Undervolt Escape Hatch

If you already own a good 850W unit and want a 5090-class card, you have one honest alternative to a new PSU: power-limit the card.

Cutting a 5090 to a 450-480W limit costs single-digit percent in inference speed — token generation is memory-bandwidth bound, and the power limit mostly bites compute-heavy prompt processing. It also shrinks the transient spikes, which scale with the operating point. Many AI builders run limited cards on purpose, PSU aside: the machine is quieter, cooler, and cheaper per month.

Do the arithmetic before trusting it, though. A limited 5090 plus a 200W CPU on 850W leaves thin margins, and thin margins on a machine that runs unattended overnight is exactly where you do not want them. We treat the power limit as the fix for noise and electricity, and the 1000W unit as the fix for stability.

At the US average of about $0.18/kWh (August 2026), the efficiency tier also stops being a rounding error: a rig averaging 400W around the clock puts roughly 290 kWh/month through the PSU — about $52 of electricity. Two or three efficiency points between a budget unit and a Platinum one is a few dollars every month, forever. Our electricity break-even numbers cover the full picture.

What Not To Buy

• A used mining-era PSU. It spent years at exactly the sustained duty cycle that ages a unit. You cannot inspect capacitor wear from a listing photo.
• An old 1000W with a 12VHPWR adapter for a 5090. The wattage is fine; the pre-ATX-3.0 transient behavior and the extra connector junction are not. The native-cable unit costs less than the card’s sales tax.
• A 1600W “just in case” for a single-GPU build. A PSU loafing at 15% load is below its efficiency sweet spot. Buy the tier that matches the build; the ladder above exists so you only ever buy one.

See Also

• Dual RTX 3090 vs one RTX 5090 — the two builds that define the 1200W tier
• Best budget local AI PC under $1,000 — where the 850W unit slots in
• 64GB local AI rig parts list — the full tower this ladder belongs to
• Local LLM electricity cost break-even — what 24/7 wattage actually costs
• Cloud GPU vs local GPU cost — whether to build at all