This guide is built from published specifications, manufacturer testing and third-party measurement rather than a specific build we assembled. Component prices move frequently, so verify before purchasing.
Most people size a power supply for AI by adding up TDP figures and adding a bit. That method is why systems crash at 3am mid-training-run and the owner spends a week blaming software.
An RTX 5090 is rated at 575W. Its transient spikes reach roughly 957W, about 79.8 amps on the 12V rail, for periods measured in milliseconds. No calculator that models average draw will tell you that, and your protection circuitry does not care that the spike was brief.
The silent killer, though, is not wattage at all.
In this guide:
- Why Calculators Get This Wrong
- The Transient Spike Numbers
- AI Load Is Not Gaming Load
- ATX 3.1 Is Not Optional Anymore
- The Silent Killer: Connector Seating
- Sizing by Build
- Why Bigger Is Not Always Better
- Who Should Buy What
- Frequently Asked Questions About a Power Supply for AI
- Verdict
- Sources and Further Reading
Why Calculators Get This Wrong
The standard power supply approach is to sum component TDPs, add 20%, and buy that. It worked when an RTX 3070 peaked near its 220W rating. It does not work now.
Most power supply calculators model average draw rather than transient behaviour. GamersNexus measured RTX 5090 transient spikes hitting 645W over sub-millisecond intervals, roughly 12% higher than one popular calculator predicted for the same card. Other testing puts peak excursions considerably higher still.
Two further power supply omissions compound it.
The platform is missing. Motherboard VRM losses, SSDs, RAM, pumps and fans add 50 to 150W that rarely appears in a quick calculation. On a build already near the edge, that is the difference between stable and not.
Simultaneous boost is missing. A CPU all-core load landing at the same moment as a GPU spike produces a combined peak neither component’s spec sheet describes. An RTX 5090 alongside an i9-14900K or Ryzen 9 9950X3D can pull over 900W in that condition.
The symptom is distinctive: a system that runs for hours and then reboots without warning. People blame drivers, thermals and software, and the cause is a power supply tripping into overcurrent protection on a spike lasting a fraction of a millisecond.
The Transient Spike Numbers
What the cards actually demand from a power supply, as distinct from what they are rated at:
| GPU | Rated TDP | Measured transient peak |
|---|---|---|
| RTX 5090 | 575W | ~957W (approaching 2x) |
| RTX 5080 | 360W | ~500W |
| RTX 4090 | 450W | ~1.7x rated |
The generational change matters. An RTX 4090 spiked to roughly 1.7 times its rating; RTX 5090 measurements approach a full 2x multiplier. That additional 30% is precisely what pushes a marginal supply past its protection threshold.
For anyone building around the cards this site covers, the RTX 5090 review is the relevant case. Mid-range cards are far less demanding: the RTX 5070 Ti and RX 9070 XT at 220W sit comfortably inside a quality 750W unit and create none of these problems.
AI Load Is Not Gaming Load
This is the part generic power supply guides miss, and it changes the recommendation.
Gaming load is bursty. Frames render, the GPU idles briefly, load fluctuates. A power supply spends most of its time well below peak.
AI inference and training are sustained. A fine-tuning run holds the GPU near maximum draw for hours. An always-on inference server does it indefinitely. The supply is not visiting high load, it is living there.
Two consequences.
Headroom should be larger. Published guidance for AI and long workloads recommends 30 to 40% headroom, landing at 1200W for a build where total draw reaches 900 to 1000W. That is more conservative than a gaming recommendation for identical hardware, and deliberately so.
Efficiency and thermals matter more. A unit running at 85% load for eight hours runs hotter, louder and less efficiently than one at 60%. Over a machine used daily, that shows up in noise, component lifespan and the electricity bill.
The same reasoning appears in our dual-GPU build guide, where power supply sizing is one of the failure modes people discover after buying rather than before.
ATX 3.1 Is Not Optional Anymore
For current-generation hardware, an ATX 3.1 power supply has moved from nice-to-have to requirement.
ATX 3.1 exists specifically to handle GPU power excursions. A certified unit can absorb transient loads well above its continuous rating for short bursts, with published figures around 200% of rated power for brief periods. A 1000W ATX 3.1 unit can theoretically ride out a 2000W spike lasting 100 microseconds. An older ATX 2.x unit of the same wattage may simply shut down.
Intel revised the standard after the RTX 4090 era exposed weaknesses in the original 12VHPWR connector. The two main changes are a redesigned 12V-2×6 connector with shorter sense pins, and tighter specification around fail-safe seating behaviour. Together they produce a connector that fails safely rather than dangerously.
The practical rule: buy ATX 3.1 for any build with a current-generation GPU, meaning RTX 4060 Ti or newer and anything in the RTX 50 series. For older hardware such as a used RTX 3090 or an RTX 3060, ATX 2.x remains fine.
One caveat that undercuts the reassurance. ATX 3.1’s transient tolerance assumes you are not already near the limit. If average draw sits at 850W on a 1000W unit, spikes push into territory where even good supplies struggle with voltage regulation. The standard buys margin; it does not replace it.
The Silent Killer: Connector Seating
Here is the power supply finding that contradicts most advice, and it is the reason the wattage conversation can mislead.
Bad seating, not insufficient wattage, causes most connector failures.
The 12V-2×6 connector needs to be fully inserted. A partially seated connector concentrates current through fewer contacts, which produces heat, which produces the melted-connector photographs that circulate after every GPU launch. That failure has nothing to do with whether your supply is 850W or 1200W.
Two specific practices follow.
Use a single native cable per GPU. Daisy-chaining PCIe-to-16-pin adapters is a documented failure pattern. If your supply has a native 12V-2×6 cable, use it. If it does not, that is an argument for a different supply rather than for an adapter.
Push until it clicks, then check. The redesigned connector has shorter sense pins precisely so an incompletely seated cable fails to power on rather than running hot. That safety mechanism only works if you let it, rather than forcing a partial connection and assuming it is fine because the system boots.
This is worth more attention than the last 200W of headroom, because a correctly sized supply with a badly seated connector is the more dangerous configuration.
Power Supply Sizing by Build
Power supply recommendations for local AI machines, assuming sustained load rather than gaming bursts.
| Build | Recommended power supply |
|---|---|
| Single mid-range GPU (5070 Ti, 9070 XT, ~220W) | 750W ATX 3.1, quality unit |
| Single RTX 5080 (360W) | 850W ATX 3.1 |
| Single RTX 5090 (575W) | 1000W floor, 1200W for AI workloads |
| RTX 5090 plus high-end CPU | 1200W minimum |
| Workstation, sustained training | 1500W 80 Plus Titanium, ATX 3.1 |
| Dual mid-range GPUs | 1200W |
| Dual RTX 5090 | 1600W and careful planning |
NVIDIA and board partners list 1000W as the floor for a 5090. Published guidance is consistent that real-world stability lives above that number, and that 850W with a 5090 should be avoided.
For AI specifically, the recommendation moves to 1200W on a single-5090 build, because the workload holds high draw for hours rather than seconds.
Why Bigger Is Not Always Better
The opposite power supply error is real and less discussed.
Power supply efficiency curves peak around 40 to 70% load. A 1000W unit powering a 350W machine spends its life below 20% load, where efficiency drops and, on some units, fan behaviour becomes less predictable. You paid more for a supply that runs less efficiently at your actual workload.
Aim for 40 to 70% load at your heaviest sustained draw. That is where quality units run quietest and coolest, and it gives ample transient headroom without overshooting.
For a mid-range AI build drawing around 400W sustained, that points at 750W rather than 1000W. For an RTX 5090 build at 900W sustained, it points at 1200W to 1500W. The rule scales in both directions, which is why “just buy the biggest one” is not the answer despite everything above.
Efficiency rating matters too, though less than wattage choice. 80 Plus Gold is the sensible default; Platinum and Titanium pay back only on machines running many hours a day, which admittedly describes a lot of AI builds.
Who Should Buy What
| If you are | Buy | Why |
|---|---|---|
| Running one mid-range GPU | 750W ATX 3.1 Gold | Lands in the efficiency sweet spot |
| Running an RTX 5090 for AI | 1200W ATX 3.1 | 1000W is the floor, not the target, under sustained load |
| Training for hours daily | 80 Plus Platinum or Titanium | Efficiency pays back on long runtimes |
| Building dual-GPU | 1200W minimum, plan connectors | Count native cables before buying |
| On older hardware (3090, 3060) | ATX 2.x is fine | Transient behaviour was less extreme |
| Reusing an existing power supply | Check ATX version first | An ATX 2.x unit with a 5090 is the classic crash |
| Tempted to overshoot | Size to 40–70% load | Efficiency drops below 20% load |
Frequently Asked Questions About a Power Supply for AI
How many watts of power supply do I need for an RTX 5090?
NVIDIA and board partners list 1000W as the floor. For AI workloads holding sustained load, 1200W is the better target because it keeps you in the efficient band and leaves room for simultaneous CPU spikes.
Do I really need ATX 3.1?
For any current-generation GPU, yes. It handles transient excursions up to roughly 200% of rated power for short bursts, and includes the redesigned 12V-2×6 connector. For older cards like a used RTX 3090, ATX 2.x is fine.
Why does my system reboot randomly under load?
The classic cause is a transient spike tripping overcurrent protection. It looks like a software or thermal problem because average draw appears fine in monitoring tools, which sample far too slowly to catch a sub-millisecond excursion.
Is 850W enough for an RTX 5090?
Published guidance strongly advises against it. The card alone spikes toward 957W, and adding a high-end CPU pushes combined peaks past 900W sustained.
Can I use an adapter for the 12V-2×6 connector?
Avoid it. Use a single native cable per GPU. Daisy-chained PCIe-to-16-pin adapters are a documented failure pattern, and most connector failures trace to seating rather than wattage.
Does a bigger power supply make my system faster?
No. Beyond adequate headroom it changes nothing except efficiency, which gets worse below about 20% load. Size to your actual draw rather than to the largest number you can afford.
Diagnosing a Power Supply Problem
If a machine already misbehaves, a short sequence that isolates the power supply from everything else.
Check whether crashes correlate with load transitions, not with sustained load. A system that runs a benchmark happily for an hour but dies when a job starts or a model loads is showing transient behaviour, not a thermal problem.
Reseat the GPU power connector. Fully unplug and reconnect, listening for the click. This costs nothing and addresses the most common cause.
Check the ATX version on the label. An ATX 2.x unit paired with a current-generation card explains a great many mysterious reboots.
Power-limit the GPU temporarily. Capping the card at 70 or 80% through your driver reduces transient peaks substantially. If crashes stop, you have confirmed the diagnosis, and you can decide between a new supply and living with the cap.
Then look at software. Drivers and frameworks get blamed first and deserve to be checked last, because a power fault mimics almost anything.
Verdict
Sizing a power supply for AI is a different problem from sizing one for gaming, and the difference is sustained load. A machine holding 900W for six hours of fine-tuning needs more headroom than the same hardware running a game, which is why the recommendation for an RTX 5090 AI build sits at 1200W rather than the 1000W floor NVIDIA quotes. Add ATX 3.1 for any current-generation card, because transient excursions approaching twice rated TDP are what actually trip protection circuits.
The honest caveat is that the wattage conversation gets more attention than it deserves relative to the thing that actually causes hardware damage. Bad connector seating, not insufficient wattage, is behind most 12V-2×6 failures. A properly sized 1200W supply with a partially inserted connector is more dangerous than an adequate 1000W unit correctly connected. Buy enough headroom, then spend the extra thirty seconds making sure the cable clicks home.
Concrete next step: before buying anything, work out your sustained draw rather than your peak, then pick a supply where that figure lands between 40 and 70% of rated output. That single calculation gets you adequate transient headroom and good efficiency at the same time, which the “add 20% to TDP” shortcut does not.
Sources and Further Reading
- Seasonic on RTX 5090 PSU calculation — manufacturer analysis of transient behaviour and the 1200W AI recommendation
- RunAIHome PSU sizing for AI workstations — ATX 3.1 guidance specific to AI builds and when older standards suffice
- Digital Gamer Hub on Blackwell transients, the 2x multiplier measurement and 12V-2×6 connector redesign
- PC Game Check PSU guide, the connector seating finding and adapter warning
- OrdinaryTech power supply calculator analysis, GamersNexus transient measurements against calculator predictions
For related coverage on this site, see the dual-GPU build guide for multi-card power planning, the AI workstation build for complete configurations, the RTX 5090 review for the card that drives these requirements, and the home AI server build for always-on machines where efficiency matters most.
