GPU Power Connector Melting: What Actually Happened, from the RTX 4090 to the 5090

Your new card is in. The chunky 16-pin cable went in a little stiff, so you pushed until you heard the click, closed the side panel, and ran a first benchmark. Everything looked fine.

Then you remembered those photos of charred connectors. You reached around the back of the case and touched the cable — warm. How warm is too warm? You looked it up, and the answer you found was probably this: that was 2022, back when the 4090 launched, NVIDIA said people weren't plugging it in all the way, the connector got revised, it's over.

It isn't over. The two most recent public cases happened on 3 September 2026 and 7 September 2026. In the four years between, the cards that melted spanned two generations, two GPU brands, first-party and third-party cables, and everything from a 575 W flagship down to a 360 W mid-range card.

This post walks you through what is actually on the record, in order — official statements, a government recall filing, numbers independent testers measured with instruments in their hands. Where the record is solid and where it simply doesn't exist, this post says so. Why the current piles onto one wire, and what you can actually do about it, are later posts in this series.

The timeline starts three weeks earlier than you think

Most coverage starts in mid-October 2022, after the 4090 went on sale. But the earliest known record is dated 20 September 2022 — three weeks before launch — and it comes from NVIDIA's own internal test document, later obtained by GamersNexus.

That document covers NVIDIA validating whether the new connector could hold the 55 amps of continuous current the spec calls for — which is what your card actually pulls at full load. The test conditions included severe bending and roughly 40 mating cycles. The result: pins 3 and 4 measured high resistance, and one of them carried 36.4 amps — at 12 volts that is 436 watts through a single pin — producing a 180°C hotspot.

How hot is 180°C for this part? The connector's rated operating ceiling is 105°C. That pin was sitting at nearly twice what it was designed to survive.

Which means "current piles onto one pin" was not a surprise discovered after the fact. The manufacturer measured it before the card shipped. So what did the public statement say? That took another two months.

Two months later, NVIDIA's public line was: about 50 cases, and you didn't plug it in properly

On 18 November 2022, NVIDIA finally spoke. Two points: roughly 50 cases worldwide, and you hadn't seated it fully — the wording was "our findings to date suggest that a common issue is that connectors are not fully plugged into the graphics card." Any cable, any card, you could RMA it, and the company said it was looking into ways to "ensure that the connector is secure before powering on."

The figure that circulated afterwards was 0.04% — 50 cases divided by roughly 125,000 4090s sold. But it's worth knowing where both halves of that fraction come from: the 50 is from NVIDIA's statement, and the 125,000 is also a number NVIDIA gave the press. Neither side of it was independently audited.

Two weeks after that, PCI-SIG — the body that writes this connector's specification — issued a statement of its own: the specs "provide necessary technical information for interoperability and do not attempt to address proper design, manufacturing methods, materials, safety testing, safety tolerances or workmanship," and members are "responsible for the design, manufacturing, and testing, including safety testing, of their products."

In plain terms: the spec guarantees the plug fits, not that it survives. That part is each vendor's problem. So how many actually failed? This is where the numbers get hard to find.

Exactly one number in this story is backed by a government filing, and it isn't 0.04%

On 8 February 2024, the US Consumer Product Safety Commission published recall 24-112, covering CableMod's angled GPU adapters. This is the only statistic in the entire story with a documented numerator and denominator:

These were adapters, not the cable that shipped with your card, so the rate doesn't transfer directly to what's in your case. But it is a number somebody signed their name to, and NVIDIA's 0.04% is not.

There's one more indirect signal from the same period: the US repair shop NorthridgeFix said in April 2024 that roughly 200 melted graphics cards were coming through the door every month, up from about 100 the previous autumn. One shop, 200 a month, doubled in a year.

And all of that is still the 4090. Three years on, with a new generation and a revised connector, things were supposed to get better.

By the 5090 generation the story changed, because this time the people measuring had instruments

In February 2025, shortly after the RTX 5090 launched, the cases came back. What was different this time: the people doing the measuring had instruments, and the cables were properly seated.

Overclocker der8auer tested one case and found more than 22 amps on a single wire — each wire should normally carry around 6 — with the connector at the power supply end reaching 150°C. That cable was a third-party MODDIY unit, but der8auer considered its build quality good, better than stock even.

Germany's Igor's Lab measured the same card: two of the twelve wires in your cable were each carrying over 20 amps — against a design value of 8; 150°C at the PSU end, around 90°C at the GPU end. You didn't break it by plugging it in wrong, is essentially what he concluded: not an isolated incident, and certainly not simple user error.

Here's the way to remember that number: 22 amps on a wire designed for 8. That isn't slightly over, it's nearly triple — and the cable was seated.

The cases that followed got progressively harder to pin on you:

The Hardware Unboxed case was found the same way you started this article: the cable had been seated properly in the PSU beforehand, and somebody happened to touch it and noticed it was unusually hot. And "one wire at 22 A, the others at zero" states the current-sharing problem about as plainly as it can be stated — you plug in one cable, but the six wires inside it go their own way.

Which raises the question: why does one wire carrying extra current melt anything? That starts with some very simple arithmetic.

The root cause: this connector has about 10% margin left at full load

A little arithmetic here, but it's easy enough to follow along. 600 watts divided by 12 volts is 50 amps, spread across 6 × 12 V wires — about 8.3 amps each. The spec requires each pin to be rated for 9.2 amps.

So at full load each pin is using 90% of its rating, with 10% left over. What about those older 8-pin connectors in your case? 60% used, 40% left.

And the most damning line comes from the specification itself. PCIe CEM 5.1 states, in writing: "Due to variations in contact resistance, an individual pin may see more than 9.2 A of current depending on cable contact resistance nonuniformity, but the total current for the assembly shall not exceed 55 A RMS in each direction."

Put those two together: the permitted total already consumes 90% of the margin, and the spec itself concedes that an individual pin may exceed its rating. This needs no measurement to establish, because it is written into the standard.

So what did that 2023 connector revision actually fix?

The 12V-2x6 revision fixed "not fully seated" and did nothing about current sharing

After the spec update in late 2023 the connector was renamed 12V-2x6, and that's the version on the RTX 50 series you'd buy today. What it actually changed:

• Power terminals lengthened by 0.25 mm, so they make contact earlier
• The sense pins shortened by 1.5 mm, so they make contact later

Together those act as a mechanical interlock: if the plug isn't fully home, the sense pins don't connect and the card won't draw full power. If you have ever pushed a plug in twice because of those photos, this change was made for you, and it does solve the partly-seated case.

But it does nothing about current sharing. Not one wire changed on the cable side; the cables are interchangeable between the two versions. Put another way, the revision fixed "did you push it all the way in" — and every case from 2025 onward involved a cable that was pushed all the way in. The RTX 5080, Club386 and Hardware Unboxed entries in the table above are all 12V-2x6.

So the problem was never fixed. And in the most recent case, "did you seat it" stopped being the question at all.

The latest one: you changed nothing, and a software update pushed the card to 613 W

The case from 7 September 2026 is a new shape again. An MSI RTX 5090 (rated 575 W from the factory) was running DLSS 5's new rendering feature when monitoring software logged board power at 613.5 W — nearly 40 W above its rated figure, and not as a momentary spike: it stayed above 610 W. Inspection afterwards found at least five scorched pins.

This one needs a clear label: it is a user report plus press coverage, with no vendor or independent lab confirmation. It's here because it represents something that hadn't appeared before — not a badly seated plug, not aged hardware, but a software update leaving the card running sustained above its own rated power. You changed no hardware and adjusted no settings.

Which probably leaves you with a few questions still unanswered.

Frequently asked questions

Is this old 2022 news that's already been fixed?

No. The most recent public cases were in September 2026. The connector was revised once, but the revision added a mechanical interlock against partial seating; it did not change how current divides between the wires.

Your card isn't a 5090 — does that mean you're fine?

No. The 360 W RTX 5080 and AMD's RX 9070 XT both have cases. The failure is one contact carrying too much, not total wattage being exceeded — current concentrating on a single pin can do it at modest power levels.

You're using the cable that came with your PSU — is that safe?

No guarantee. In the cases listed above, the RTX 5080 one was on a PSU's own bundled cable, and Club386's and Hardware Unboxed's were on dedicated first-party cables. Conversely, a third-party cable der8auer rated as high quality melted too.

How worried should you be? How many cards have actually failed?

There's no reliable total. The only figure backed by a government filing is CableMod's 1.07%; NVIDIA's 2022 figure of 0.04% has a vendor-supplied number on both sides of the fraction. For the RTX 5090 generation there is no official or independent count at all — the most conspicuous gap in the record.

Does lowering the GPU power limit prevent melting?

No. That RTX 5090 from February 2026, already undervolted and capped at 500 W, melted anyway. Lowering the limit brings down average current per wire, but it does nothing about uneven current sharing, nothing about a bad contact, and nothing about microsecond power excursions. That deserves its own post, and it's coming later in this series.

Where this stands, honestly

Everything in this post is somebody else's work: official statements, a government recall filing, numbers independent testers measured with instruments. I don't have a lab, I have never measured a cable myself, and I have no way to verify any of the figures above. What I can do is mark the sources clearly, and say plainly where the record is missing — the total number of 5090-generation cases, for instance, simply isn't out there.

This post also deliberately doesn't tell you "what to do." The reason: in the public record, the one root cause independent testers consistently point at is uneven current sharing, and that lives at the graphics card's board design level, not anywhere you can reach from software. Every setting claimed to "prevent melting," lowering the power limit included, already has a counter-example.

There's also a case where you can close this tab right now: if your card is powered by traditional 8-pin connectors, none of this applies to you. You don't need to buy anything because of this article, and you don't need to install anything either.

Later posts in this series cover: why current concentrates on one wire, how those old dual 8-pin connectors in your case actually differ from today's 12-pin, why the spec permits a card to spike to three times its rated draw, and what lowering a power limit can and cannot do.

If you do want to cap your card's power draw yourself, the command is nvidia-smi -pl 500 from an administrator prompt (check your card's legal range with nvidia-smi -q -d POWER). Two things to say up front: this does not prevent connector melting — that 5090 capped at 500 W is the counter-example — and the setting is wiped by the driver on every reboot, so you'd need to schedule a startup task yourself.

One last thing, and it has nothing to do with melting, so let's be clear about that. PowerDoze handles a different question: whether your card is running inside the envelope it's rated for. The DLSS 5 case above is exactly that situation — you changed nothing and a software update pushed the card to 613 W. If you want a power limit put back without retyping that command after every reboot, PowerDoze does that part — the limit lives inside a power mode, and an all-day, every-day schedule rule pointed at that mode re-applies it at every startup. It cannot prevent connector melting, and this post has explained why.

Download PowerDoze free →