The single most under-spec’d component in home AI builds is the power supply. Builders who would never short-change a $1,500 GPU happily pair it with a $60 generic 750W PSU — then wonder why the system crashes during long inference runs or the GPU melts a power connector. Right-sizing the PSU is the difference between a stable, multi-year AI workstation and a build that randomly reboots under load.

This piece runs the actual wattage math for AI workstation builds in 2026, gives a clear PSU recommendation by GPU tier, and explains which 80 PLUS rating is worth paying for versus which is marketing. If you’re spec’ing a build with an RTX 4090, RTX 5090, or any modern AI card, the wattage answer is here.

PSU specifications and standards verified against the PSU Wikipedia overview on May 5, 2026.

The first principle: 40% headroom

The standard PSU sizing rule is “size your PSU 40% above the calculated peak system power draw.” Why 40%:

1. Peak draw is higher than typical. A GPU rated at 450W TDP can spike to 600W+ for milliseconds during intense inference or training workloads. A PSU sized for the rated TDP will fail those spikes.

2. PSUs run efficiently around 50% load. An 80 PLUS Gold PSU is most efficient at 40-60% of rated capacity. Running at 80%+ load constantly causes heat-related degradation over years.

3. Future-proofing. Adding a second drive, a fan controller, or upgrading to a more power-hungry GPU mid-build shouldn’t force a PSU swap.

The 40% rule is conservative but correct for AI workstations specifically — these systems run sustained load for hours during training or batch inference, unlike gaming rigs that hit peak only briefly.

Component-by-component power draw

For a typical AI workstation in 2026, peak component-level power draw:

Component	Typical peak draw	Notes
CPU (Ryzen 9 9900X / 9950X)	170-230W	Higher under all-core load
Motherboard	30-50W	Includes SSD slots, USB power
RAM (64GB DDR5)	20-30W	Negligible compared to GPU
NVMe SSD (Gen4 2TB)	10W	Trivial
HDD (if any)	10W	Trivial
Case fans (4-6)	15-25W	Trivial
Liquid cooling pump	5-15W	Only if AIO/custom loop
GPU (varies by card)	see table below	Dominates the budget

The GPU is 60-80% of total system power for any AI workstation. Sizing the PSU is essentially “size around the GPU.”

GPU power draw by tier

Verified TGP (Total Graphics Power) for AI-relevant cards:

GPU	TGP (rated)	Peak transient	NVIDIA-recommended PSU
RTX 3060 12GB	170W	~200W	550W
RTX 4060 Ti 16GB	165W	~200W	550W
RTX 5060 Ti 16GB	180W	~220W	600W
RTX 5070 12GB	220W	~265W	650W
RTX 5070 Ti 16GB	300W	~360W	750W
RTX 5080 16GB	360W	~430W	850W
RTX 3090 24GB (used)	350W	~430W	750W
RTX 4090 24GB	450W	~550W	850W
RTX 5090 32GB	575W	~700W	1000W

The transient peaks (the rightmost numbers) are why the headline TGP isn’t enough. A 5090 rated at 575W can briefly draw 700W+ during dense inference workloads. A PSU sized for 575W will trip protection circuits during those spikes, causing system reboots.

For details on the RTX 5090 vs RTX 4090, used RTX 3090, and RTX 5060 Ti vs 4060 Ti at each tier, see our companion buying guides.

Total system wattage by GPU tier

Combining CPU + GPU + everything else, then applying the 40% headroom rule:

GPU	Calculated peak system	+40% headroom	Recommended PSU
RTX 5060 Ti 16GB	~450W	~630W	650-750W
RTX 5070 Ti 16GB	~600W	~840W	850W
RTX 5080 16GB	~660W	~924W	1000W
Used RTX 3090 24GB	~660W	~924W	1000W
RTX 4090 24GB	~750W	~1050W	1000-1200W
RTX 5090 32GB	~900W	~1260W	1200W minimum
Dual RTX 3090 (multi-GPU)	~1050W	~1470W	1500W
Dual RTX 4090 (multi-GPU)	~1250W	~1750W	1600-2000W (HEDT) or 2× 1200W

The headline numbers: a 5090 build needs a 1200W PSU. A 4090 build needs 1000W minimum, 1200W comfortable. A 3090 or 5070 Ti / 5080 build needs 1000W. Anything below RTX 5070 Ti class works on 850W.

These recommendations assume a Ryzen 9 9900X / 9950X CPU. For a Ryzen 7 (lower TDP) or Intel Core Ultra (different power profile), shift one tier lower.

80 PLUS rating: which to pay for

The 80 PLUS efficiency ratings measure how efficiently a PSU converts wall AC to DC for your components. The ratings:

• 80 PLUS Bronze: 82-85% efficient at typical loads
• 80 PLUS Silver: 85-88% efficient (rare; mostly skipped in 2026)
• 80 PLUS Gold: 87-90% efficient
• 80 PLUS Platinum: 90-92% efficient
• 80 PLUS Titanium: 94-95.4% efficient at 50% load

For most AI workstation builds, Gold is the right tier. The math:

• A 1000W Bronze running at 600W average draws ~720W from the wall (83% efficient)
• A 1000W Gold running at 600W average draws ~690W from the wall (87% efficient)
• A 1000W Platinum running at 600W average draws ~670W from the wall (90% efficient)
• A 1000W Titanium running at 600W average draws ~660W from the wall (91% efficient)

The Gold-vs-Platinum gap is ~3%. At $0.15/kWh and 4 hours/day of 600W AI inference, that’s roughly $4-$5/year in electricity savings. The price gap between Gold and Platinum is typically $50-$100. Platinum doesn’t pay back at home AI usage levels.

Titanium is for 24/7 always-on home AI servers where the 4-7% efficiency over Gold accumulates to meaningful savings over 5+ years. For occasional-use workstations, Titanium is overkill.

Don’t go below Gold for a serious AI workstation. Bronze PSUs typically have weaker capacitor quality, less stable voltage regulation under load, and shorter warranties. The $30-$50 you save buying Bronze isn’t worth the long-term reliability difference.

ATX 3.0 vs ATX 3.1 vs ATX 12VO

The current standard is ATX 3.1 as of mid-2025. Key features for AI builds:

• Native 12V-2x6 connector: the backward-compatible successor to the 12VHPWR connector that had melting issues on early RTX 4090 cards. ATX 3.1 PSUs ship with a fresh 12V-2x6 cable that’s safer to mate.
• Better transient response: ATX 3.0/3.1 PSUs are designed to handle the high transient peaks of modern GPUs (200% rated power for short durations).

Practical recommendation: buy an ATX 3.1 PSU for any build with a current-gen GPU (RTX 4060 Ti or newer, RTX 5000-series). For older cards (RTX 3090 used, RTX 3060), ATX 2.x PSUs work fine.

ATX 12VO is a different design philosophy (motherboard handles voltage rails) — niche, not relevant for most home AI builds. Skip it unless you’re specifically optimizing for OEM-style efficiency.

The 12V-2x6 connector and the melting story

If you’re upgrading to an RTX 4090 or 5090 in 2026, you’ll encounter the 12V-2x6 connector. The history:

• RTX 4090 launched with the 12VHPWR connector in 2022
• A subset of cards experienced melting at the connector — typically attributed to incomplete insertion, sharp cable bends near the connector, or aftermarket adapter issues
• 12V-2x6 was introduced as the backward-compatible successor with redesigned safety features (sensing pins shifted, longer power pins for proper engagement)
• Modern RTX 5090 cards and ATX 3.1 PSUs both use 12V-2x6 by default

Practical safety advice when installing a high-end GPU:

1. Plug the connector all the way in until the latch clicks audibly
2. Do not bend the cable within 35mm of the connector
3. Avoid 8-pin to 12V-2x6 adapters; use a native cable that came with the PSU
4. If the PSU is older ATX 2.x, check the manufacturer’s compatibility list before using their 12V-2x6 adapter

Most current-generation PSUs from reputable brands (Corsair, Seasonic, EVGA, Cooler Master, be quiet!, Super Flower) ship with a native 12V-2x6 cable. Use it.

Modular vs semi-modular vs non-modular

Fully modular (recommended for AI builds): all cables detachable. You only install the cables you need, reducing case clutter and improving airflow.

Semi-modular: 24-pin and 8-pin CPU cables fixed; PCIe and peripheral cables modular. Cheaper than fully modular, almost as good in practice.

Non-modular: all cables fixed. Cheapest, but cable management is a pain in cases with limited space.

For a $1,500+ AI workstation, the $20-$40 premium for a fully modular PSU is worth it. Cable management directly affects airflow, which affects GPU thermals during sustained AI workloads.

What about server-tier PSUs and Platinum HEDT options?

For multi-GPU AI servers (dual or triple RTX 4090/5090):

• Single 1500W ATX PSU: works for dual 4090s; tight for dual 5090s
• Single 2000W server-grade PSU: works for dual 5090s (rare in consumer builds; loud)
• Dual 1200W PSU setup: more complex but provides redundancy; requires a dual-PSU adapter and careful planning

For most home AI builds, single-GPU configurations are sufficient and a single 1000-1200W PSU covers them. Dual-GPU and beyond is enterprise-adjacent territory and the PSU advice shifts to “consult a builder” rather than “follow this guide.”

Specific brand recommendations

I won’t name specific SKUs — pricing fluctuates and exact models change quarterly. Stick to these reputable PSU brands for AI workstations:

• Corsair (RMx, HX, AX series)
• Seasonic (Focus, Prime series)
• EVGA (SuperNova series — though EVGA exited GPU market, their PSUs are still excellent)
• Cooler Master (V Gold V2, MWE series)
• be quiet! (Dark Power, Straight Power series)
• Super Flower (Leadex series — OEM for many premium PSUs)
• Thermaltake (Toughpower GF, GFX series)

Avoid generic / bargain-brand PSUs (EVGA W series, Apevia, Diablotek, etc.) for AI workstations. The premium PSU brands cost $30-$80 more but the difference in capacitor quality, voltage stability, and warranty (typically 7-10 years for premium brands vs 3-5 years for budget) directly affects how long your $1,500+ GPU stays alive under sustained load.

For current pricing, search Newegg, Amazon, or B&H for the wattage tier and brand from the list above. Verify ATX 3.1 compliance and 12V-2x6 connector inclusion before buying.

Total cost-of-PSU by build tier

For sanity-check, typical PSU pricing in May 2026:

Wattage / Rating	Typical price	When to buy
750W Gold	$90-$130	RTX 5060 Ti / 5070 builds
850W Gold	$120-$170	RTX 5070 Ti / 3090 / 5080 builds
1000W Gold	$160-$220	RTX 4090 / 5080 builds
1200W Gold	$200-$280	RTX 5090 / dual-GPU builds
1000W Platinum	$230-$310	When you specifically want efficiency premium
1500W Gold	$300-$420	Multi-GPU / serious AI server

For a typical $2,000-$2,500 AI workstation build, plan $150-$220 for the PSU. That’s 7-11% of your total budget — a fair allocation for the component that determines stability and longevity of the build.

Honest verdict by build profile

Profile	PSU recommendation	Reasoning
Entry build (RTX 5060 Ti / 4060 Ti)	750W 80 PLUS Gold ATX 3.1	$90-$130, comfortable headroom
Mid build (RTX 5070 Ti / 3090 / 5080)	850W 80 PLUS Gold ATX 3.1	$120-$170, sweet spot for single-GPU
Flagship (RTX 4090 / 5090)	1000-1200W 80 PLUS Gold ATX 3.1	$160-$280; 1200W for 5090 specifically
Multi-GPU (dual 4090)	1500W 80 PLUS Gold or Platinum	$300-$420; PSU is the bottleneck
24/7 home AI server	1000W 80 PLUS Platinum or Titanium	Efficiency premium pays back over years
Used legacy build (RTX 3060/3090)	750-850W 80 PLUS Bronze or Gold	Spend savings on better GPU

The single most important rule: don’t pair a $1,000+ GPU with a $60 generic PSU. Spend $150+ on a PSU from the reputable brands listed above with adequate wattage headroom. A weak PSU under sustained AI load is the most common failure mode in home AI workstations — and the failure mode that takes other components with it.

For the broader picture of AI workstation building, see our GPU buying guide for local AI, system RAM sizing guide, and the used RTX 3090 evaluation.

If you’re a developer planning to use this workstation primarily for AI coding workflows like Cline + local LLM, the same PSU sizing applies — coding workloads draw the same GPU power as inference workloads of equivalent compute intensity.

Sources

• Power supply unit specifications and standards — Wikipedia
• GPU buying guide for local AI — TGP reference for major cards
• RTX 5090 vs RTX 4090 — Power requirements detailed comparison
• Used RTX 3090 power and thermal characteristics
• System RAM sizing guide — companion build component reference
• 80 PLUS efficiency rating standard
• ATX 3.1 specification overview — Tom’s Hardware coverage

Last updated May 5, 2026. PSU prices fluctuate weekly; verify current pricing on retailer pages before purchasing. Wattage recommendations assume single-GPU consumer builds with reasonable component selections; HEDT and server builds need bespoke calculations.