Best PSU for an AI Workstation (2026) — Wattage Sizing, ATX 3.1 / 12V-2×6, and the Sustained-Load Rules Gaming Guides Get Wrong

Every component guide on this site so far has been about throughput — how much RAM, which NVMe, which GPU. This one is about the part that decides whether the other three stay powered on. It is also the part where almost every article you will find is wrong for your use case, because it was written for gamers.

The Short Answer: PSU Wattage by AI Build

An AI workstation should be sized so its power supply runs at 40–80% of rated capacity under sustained load, not at its peak rating. A gaming PC spikes briefly and idles between frames; an inference or fine-tuning box holds its GPUs near 100% for hours at a time. In practice that means a single RTX 5090 (575W TGP) wants a 1000W minimum and 1200W recommended ATX 3.1 unit with a native 12V-2×6 cable, a dual-5090 build needs 1600W or more, and a 4× RTX 3090 rig exceeds what a standard US 15-amp/120V circuit can safely deliver continuously (~1,440W) before the PSU's own conversion losses are counted — making the wall circuit, not the power supply, the real ceiling on a home AI rig.

"Sustained system draw" assumes one mainstream high-core CPU at ~150–200W PL2 plus ~100W of platform overhead (board, RAM, NVMe, fans, pumps). GPU figures are board TGP as published by NVIDIA and reflected in our product specs. Recommended wattage is derived by dividing sustained draw by a 0.65–0.75 target load factor — the arithmetic is in the next-but-one section, so you can re-run it for your exact parts.

Why AI Load ≠ Gaming Load (The Rule Everyone Gets Wrong)

Search "best PSU for RTX 5090" today and every result is a gaming article. That is not a snark — it is a sizing problem. Gaming and AI stress a power supply in opposite ways, and the recommendations that fall out of each are genuinely different.

Gaming is a transient problem. Frame-to-frame, a GPU swings between near-idle and full boost in single-digit milliseconds. The failure mode is a microsecond excursion — the card momentarily pulls two or three times its nominal draw and trips the PSU's over-current protection, and your machine reboots mid-match. ATX 3.x exists largely to codify this: the spec requires units to ride out defined excursions above rated power without shutting down. Gaming guides therefore optimize for headroom against spikes.

AI is a thermal and duty-cycle problem. A fine-tuning run or a batched inference server holds the GPU at 95–100% utilization continuously, for hours or days. There is no idle between frames to shed heat into. Seasonic — a PSU manufacturer, writing about its own products in AI Workstations vs Gaming PC: What's the Difference in Power Supplies Load — draws exactly this distinction and lands on the same conclusion we do: AI systems sit at a high, flat load for extended periods, which pushes them toward higher efficiency tiers and more conservative sizing than a gaming build of the same nominal wattage.

Three consequences follow, and they are the spine of this guide:

• The 40–80% band matters more than the peak rating. The 80 PLUS certification program tests at 20%, 50%, and 100% load for a reason — efficiency peaks around the middle of the curve and falls off at both ends. A PSU parked at 90% load runs hotter, spins its fan harder, ages its capacitors faster, and converts a larger fraction of your electricity bill into case heat. A PSU parked at 65% does none of that.
• Rated output is a 25°C number. Continuous-output ratings are specified at an ambient temperature most multi-GPU cases never see. As intake air rises — and in a four-card chassis it rises a lot — usable headroom falls. Check the manufacturer's derating curve; a unit rated for full output to 50°C is worth paying for in a dense build, and a unit that quietly derates above 30°C is not.
• Duty cycle changes the efficiency math from a rounding error into money. Two percentage points of efficiency on a box that runs four hours a week is noise. On a box that runs 168 hours a week it is a line item, and we do that arithmetic below.

None of this shows up in a gaming recommendation, because none of it matters when your GPU is at 100% for ninety seconds at a time.

How to Actually Size It: The Math

Do not multiply your GPU's TDP by "1.5 or so." Do this instead:

Sustained draw = (GPU TGP × count) + CPU PL2 + ~100W platform
PSU rating     = Sustained draw ÷ 0.65   (target ~65% load)

The divisor is the whole trick. Sizing by peak-times-a-guess gives you a number with no defined relationship to the load band; dividing by a target load factor puts you in the band on purpose. Use 0.65 for a 24/7 box, 0.70–0.75 if the machine is a workstation that also idles, and never above 0.85.

Worked example 1 — single-GPU upgrader

An RTX 5090 ($1,999 – $2,199, 575W TGP) with a Ryzen 9 or Core Ultra 9 class CPU (~200W PL2) and ~100W of platform overhead:

575 + 200 + 100 = 875W sustained
875 ÷ 0.70 = 1250W  →  buy 1200W (nearest common rating)
875 ÷ 1000 = 87.5% load on NVIDIA's stated minimum

NVIDIA's official 1000W system-power specification is a minimum, and it is correct as a minimum — the machine will boot and run games all day. But 87.5% sustained is the wrong place to live for a machine that fine-tunes overnight. 1200W puts you at 73%. That is the entire argument for the extra $60.

Worked example 2 — multi-GPU / home server

Four used RTX 3090s ($699 – $999 each) — still the best VRAM-per-dollar stack in 2026, as we argued in used RTX 3090 vs RTX 5060 Ti — on a workstation platform:

(350 × 4) + 150 + 100 = 1650W sustained
1650 ÷ 0.70 = 2357W  →  no single consumer PSU exists at this rating

Two things break here at once. There is no mainstream 2400W ATX unit, and 1650W of DC output means roughly 1790W drawn from the wall at 92% efficiency — well past a 15A circuit. This is the build that forces either power limits, a dual-PSU split, or a different chassis strategy; all three are covered below.

Ampere cards deserve one extra warning. The 3090's 350W nominal understates what the PSU sees: independent testing has measured transient spikes to roughly 464W on 3090-class cards against that 350W rating. Sustained draw sets your load band; transients still set your OCP margin. A unit with a generous single +12V rail and ATX 3.x excursion compliance handles both. A tired 1000W unit from 2018 handles neither.

Worked example 3 — "is my existing PSU fine?"

Frequently, yes — and we would rather tell you that than sell you a PSU. An RTX 5060 Ti 16GB ($429 – $479) is a 150W card. Even with a hungry CPU:

150 + 150 + 100 = 400W sustained
400 ÷ 650 = 62% load on a decent existing 650W unit

That is dead center of the band. If you are running a 16GB-class card and your PSU is a reputable 650W+ Gold unit under about six years old, buy VRAM instead. The same is true for the RTX 4060 Ti 16GB ($399 – $449). See AI on a budget for where that money goes further.

80 Plus Tier: When Gold Is Fine and When You Need Platinum or Titanium

The certification tiers are efficiency at defined load points. On 115V, at the 50% test point, the relevant floors are 90% for Gold, 92% for Platinum, and 94% for Titanium — Titanium additionally has to hold ~90% at a 10% load point, which is why it behaves better on a machine that idles between jobs.

Put dollars on it. Take a workstation pulling 1200W of DC output around the clock:

Rate of $0.17/kWh is a US national-average placeholder; substitute yours. Full regional math is in what local AI actually costs to run in electricity.

The Gold-to-Platinum step saves about $43/year on a 24/7 box — roughly the price gap amortized over three years, which is a wash on money alone. The tiebreaker is the last column: 29W of heat that never enters the case, in a chassis that is already fighting four GPUs. That is why we recommend Platinum for always-on rigs and Gold for everything else. Titanium is a luxury unless power is expensive where you live or the rig lives in a closet with no air conditioning.

One caveat that gaming guides skip entirely: the certification is measured at 25°C ambient. Every number above degrades in a hot case. If your build is dense, weight the manufacturer's temperature derating curve more heavily than the badge on the box.

12V-2×6 and the Connector Problem (2026 Status)

This section is the one with real news value, and the one manufacturer-adjacent content will not write plainly. High-current 16-pin connectors on 4090- and 5090-class cards were still melting through mid-2026 — and not only in the ways the internet's folk wisdom predicts.

The documented cases matter because of what they rule out:

• TechPowerUp reported an RTX 5090 melting its 12V-2×6 connector despite a −100W power limit. Reducing the load did not prevent it.
• Guru3D covered a melt under a 500W cap — again, below stock.
• Tom's Hardware reported a power wire catching fire on a first-party cable. "Use the cable that came in the box" is necessary but not sufficient.
• VideoCardz documented a ZOTAC RTX 5090 melt on a below-recommended 850W PSU — the one case that is straightforwardly about undersizing, and the reason the table above starts at 1000W.

The accepted root-cause explanation comes from independent testing by der8auer (Roman Hartung), whose measurements of per-wire current on a live 5090 found the load distributed wildly unevenly across the six 12V conductors — tens of amps concentrated on one wire while others carried almost nothing, against a per-wire design margin far below that. The six 12V pins are wired in parallel with nothing forcing them to share equally; any difference in contact resistance, seating depth, or crimp quality steers current toward the lowest-resistance path, and that path heats, which changes its resistance further.

What compounds it is a design trend: the RTX 3090 Ti monitored its 12V input across three shunt resistors and could detect imbalance across pin pairs; the 4090 reduced that, and the 5090 effectively measures its input as a single node. A card that cannot see per-pin current cannot throttle or fault on an imbalance. That is why a power limit does not save you — you can halve the total current and still push 20A+ down one 16AWG wire.

This is not a reason to avoid a 5090. It is a reason to be exacting about five things:

1. Native 12V-2×6, no adapters. Buy an ATX 3.1 PSU and use its own 12V-2×6 cable to the card. Every 4×8-pin adapter, extension, and third-party "custom" cable adds two more contact interfaces where resistance can diverge.
2. Never daisy-chain or split. One cable, one card, straight from the PSU. A GPU pulling 575W should not share a cable run with anything.
3. Seat it fully, then check. The 12V-2×6 revision (PCIe CEM 5.1) shortens the four sense pins and lengthens the power and ground pins specifically so a partially-inserted plug negotiates down to low power instead of arcing at full current. That safety only works if the connector is a native one and you hear the latch click.
4. Keep the bend radius sane. Do not fold the cable within ~35mm of the plug. Side-loading the housing is one of the ways contact pressure goes uneven in the first place.
5. Thermal-check it once, then occasionally. After your first long run, feel the connector housing or point an IR thermometer at it. Warm is normal. Hot compared to the cable ~10cm away means uneven current, and that is your early warning.

Multi-GPU builders should read this alongside our multi-GPU local LLM setup guide, where cable routing and slot spacing interact with all of the above.

Recommended PSUs by Build Tier

Three picks mapped to the three builds people actually have. All are units already carried in our accessory bundles — you will find them on the build pages for the GPUs they pair with.

Single mid-tier GPU (5060 Ti / 4060 Ti / 5080 / 4080 SUPER) — Corsair RM850x ATX 3.1

850W, 80+ Gold, fully modular, with a native 12V-2×6 cable — roughly $130–$170. This is the default for a one-card workstation up to the 360W class. At an RTX 5080 build's ~660W sustained draw it sits at 78% load, which is the top of the acceptable band; at an RTX 5060 Ti build's ~400W it sits at a very comfortable 47%.

What it is not enough for: an RTX 5090 (~875W sustained is over 850W of rated output — do not), a 4090 running 24/7, or any two-card configuration. If you are building around a 5080 and think you might add a second card later, skip this tier now and buy once.

Single 5090 or dual mid-tier GPU — Corsair HX1500i (2025) ATX 3.1

1500W, Cybenetics Platinum, ATX 3.1, native 12V-2×6, roughly $420–$500. This is the workhorse recommendation of the post. A single 5090 build at ~875W sits at 58% load; a 2× RTX 3090 build at ~950W sits at 63%; a 2× 5090 build at ~1450W sits at 97% and is not supported without power limits. Its digital monitoring (per-rail current and temperature telemetry over USB) is genuinely useful on an always-on box, because it lets you log real sustained draw rather than trusting this article's arithmetic.

What it is not enough for: two stock-power 5090s, or four 3090s. Also note that at 1500W of DC output this unit can, on its own, exceed what a US 15A circuit will deliver continuously — see the next section before you assume the wall keeps up.

4-GPU / 2000W+ tier — no single-unit recommendation

We will be honest rather than pad the list: there is no consumer ATX unit we would recommend for a stock-power 4× 3090 or dual-5090 build, because the honest answer at that tier is not a bigger PSU. It is one of three things:

• Power-limit the cards. Four 3090s capped at 250W each (nvidia-smi -pl 250) draws 1000W of GPU instead of 1400W, costs roughly 10–15% of inference throughput, and brings the whole build to ~1250W DC / ~1360W at the wall — under the 1440W continuous limit. This is the single highest-leverage command in multi-GPU local AI.
• Split across two PSUs on two circuits. Covered in the FAQ; it solves the electrical problem, not just the PSU problem.
• Buy a chassis engineered for it. A Supermicro SYS-421GE-TNRT ($8,000 – $15,000 barebones) ships with redundant hot-swap power supplies, 240V-capable input, and airflow designed around passive datacenter cards. If your build has crossed 2000W, you have left the ATX design envelope, and the OEM path — also covered in best prebuilt AI workstations and local AI servers for business — is cheaper than engineering it yourself twice.

Passive datacenter cards deserve a footnote here: an A100 80GB PCIe (300W) or H100 PCIe (350W) draws less than a 5090, but has no fan of its own. It expects a server chassis pushing high static pressure front-to-back. Dropping one into a quiet desktop case is a thermal failure, not a power one — a distinction the same mistake gets made in both directions.

The Part Nobody Plans For: Your Wall Circuit

Here is the fact almost nothing ranking for "best PSU for AI workstation" will tell you: your power supply's rating and your outlet's capacity are different claims, and the outlet usually loses.

A standard US 15A/120V branch circuit is rated 1800W. The NEC continuous-load rule — a load running three hours or more must not exceed 80% of the branch rating (articles 210.19 and 210.20) — brings the usable figure to 1440W. An AI workstation running a six-hour fine-tune is the definition of a continuous load. The relevant capacities:

Three practical notes. First, the number that counts is AC draw at the wall, not PSU DC output — divide your sustained draw by the efficiency figure, so 1650W of DC at 92% is ~1793W from the outlet. Second, that circuit is almost certainly shared with room lighting, other outlets, and possibly a space heater someone plugs in each winter; the 1440W is a ceiling for everything on it, not for your rig alone. Third, if a long inference run trips a breaker, the loss is not the breaker — it is the checkpoint.

The fix hierarchy, cheapest first: power-limit the GPUs → move the rig to a different circuit than the rest of the room → have an electrician pull a dedicated 20A line → split across two circuits with two PSUs → 240V. Non-US readers on 230V mains can generally skip to the next section; your outlet has roughly twice the headroom and this constraint rarely binds.

UPS, Surge, and the Always-On Rig

Sustained-load boxes are precisely the ones that lose a 40-hour fine-tune to a two-second brownout. Four rules:

• Pure sine wave, not simulated. This is not optional. Every modern PSU uses active power factor correction, and active-PFC circuits can misbehave — buzzing, refusing to transfer, or shutting down — on the stepped-square approximation that cheap "simulated sine" units output. Pure sine costs more and is the entire point.
• VA is not watts. Consumer UPS units are rated in volt-amps with a power factor around 0.6, so a 1500VA unit typically delivers 900–1000W. Size against the watt rating, not the marketing number. The CyberPower CP1500PFCLCD (1500VA, ~$200–$260) is the standard pick for a single-GPU workstation; the APC Smart-UPS 2200VA rackmount (~$1,000–$1,400) is the pick for a multi-GPU or rack build, and the smaller CyberPower CP850PFCLCD (~$130–$180) is right-sized for a Mac Mini M4 Pro or Mac Studio.
• Size for graceful shutdown, not ride-through. A 900W-capable UPS carrying an 800W rig gives you a few minutes, not an hour. That is enough — configure NUT or the vendor daemon to checkpoint and halt at 50% battery. Trying to ride out a real outage on a 1200W AI rig requires a battery bank, not a UPS.
• Surge protection is separate and cheap. A 12-outlet 4000J strip (~$45–$65) for peripherals and monitors costs less than one afternoon of debugging a corrupted model file. The UPS covers the rig; the strip covers everything around it.

Our home AI server build guide covers the always-on side in more depth, including the monitoring daemon setup.

Cooling, Sag, and the Rest of the Power Path

Sustained load promotes three "optional" parts into load-bearing ones.

The CPU cooler stops being decorative. Data prep, tokenization, and CPU-offloaded layers in llama.cpp can peg a CPU for hours alongside the GPU. A 360mm AIO (~$110–$150) handles 280W+ of CPU TDP without throttling; a mid-range tower cooler sized for gaming bursts will not.

Case airflow becomes a PSU spec. Everything in the derating discussion above depends on intake temperature. A five-pack of good 140mm static-pressure fans (~$40–$55) is the cheapest way to keep the PSU's ambient — and therefore its actual usable output — near its rated conditions. If noise is the constraint, our quiet AI PC guide works through the trade-off; the short version is that a larger PSU running at 55% load is quieter than a smaller one at 85%, because its fan curve barely engages.

Gravity is a 24/7 problem. A 2–3kg triple-slot card hanging off a PCIe slot for two years, thermally cycling the whole time, works its connector and its slot. A $15–$25 magnetic support bracket per card is trivially worth it, and it also keeps the 12V-2×6 plug from being side-loaded by a sagging card — which loops directly back to the connector section.

Don't Want a 1,500W Space Heater? The Low-Power Path

If you have read this far and concluded that a dedicated 20A circuit is more electrical work than your AI hobby justifies — that is a legitimate conclusion, not a failure of nerve.

A Mac Studio M4 Max ($1,999 – $5,999) draws roughly 160–200W under sustained inference load. It runs Llama 4 Maverick 70B, Qwen 3 72B, and DeepSeek R1 70B at usable speeds out of unified memory, plugs into a normal outlet, has no PSU decision, no 12V-2×6 connector, no derating curve, and no circuit math. A Mac Mini M4 Pro ($1,399 – $1,599) does the same at a smaller scale, and clusters if you want more.

The honest trade-off: Apple Silicon's memory bandwidth is well below a 5090's 1,792 GB/s, so tokens per second on models that fit in 32GB of VRAM will be lower — sometimes much lower. Training and fine-tuning throughput is not close, and the CUDA ecosystem is still where the tooling lives. What you get in exchange is capacity: 128GB of unified memory holds models a 32GB card cannot touch at any power level, in a box you can leave running under a desk. We compared the two directly in Mac Studio M4 Max vs RTX 5090, and if you want the raw power-per-token framing, RTX 5090 vs RTX 3090 and RTX 4090 vs RTX 3090 put the generational efficiency gains in context.

The Verdict

• Single 5090: 1200W ATX 3.1 Platinum, native 12V-2×6. Not the 1000W minimum.
• Single 5080 / 4080 SUPER or below: 850W ATX 3.1 Gold. Your existing good 650–750W unit is probably fine for a 5060 Ti.
• 2× 5090: 1600W+ with cards power-limited to ~450W each, or two PSUs on two circuits. Check the wall first.
• 4× 3090: power-limit to 250W each and split across two circuits — or buy a chassis that was engineered for it.
• 24/7 anything: Platinum over Gold for the heat, not the money. Pure-sine UPS sized for shutdown.
• None of the above: the Apple Silicon path has no PSU chapter at all.

A $200 power supply next to a $2,000 GPU is the cheapest insurance in the build, and — per where GPU prices sit in 2026 — the only component here whose price is not being set by a shortage. Buy the right one once.

Next in this component series: how much system RAM local AI actually needs and the NVMe guide. For the full build, start at how to build an AI workstation step by step and what it costs, or browse the AI GPU buying guide hub.