RTX 5090 SE: the rumor of a castrated 5090 at $1500, let's take stock

Mid-July 2026, the Russian site GameGPU drops a bomb: Nvidia would be preparing a GeForce RTX 5090 SE , a lighter version of its flagship card based on the same GPU GB202. The thing was picked up by TechPowerUp, Clubic, Notebookcheck — in short, everyone is talking about it.

The timing is not insignificant. We are right in GPU drought : no RTX 50 Super before CES 2027, AMD postponing RDNA 5 to mid-2027, Intel Arc Celestial in uncertainty. The only gaming card released this year is the RTX 5060 Ti. Suffice it to say that any rumor of a new GPU has the effect of an oasis in the desert.

But does this RTX 5090 SE hold up? Or is it just hot air that suits everyone, including Nvidia? Let's dissect.

Click to enlarge

According to GameGPU, the RTX 5090 SE would use the the GB202 — the same silicon as the classic RTX 5090 — but in a seriously castrated version. We go from 170 activated SM (on the retail 5090) to only 110 SM , soit 14,080 CUDA cores. This is a 35% cut compared to the 5090, but it remains 30% above the RTX 5080 and its 10,752 cores.

On the memory side, the leak announces 32 GB of GDDR7 but on a bus reduced to 384 bits (against 512 bits on the 5090). The TGP would drop to 500W - yes, "drop," because the 5090 hits 575W. And the price? 1 500$ MSRP , between the RTX 5080 at $999 and the RTX 5090 at $1,999.

On paper, it looks coherent. A kind of RTX 5080 Ti that doesn't reveal its name. Except that when you dig a little, several things don't add up.

That's where it gets stuck. A 384-bit memory bus, that means 12 channels of 32 bits With the current 2GB GDDR7 chips, that gives 24GB. Not 32. To reach 32GB on a 384-bit bus, you would need to mix 3GB and 2GB chips on the same bus - eight 3GB modules and four 2GB modules, to be precise.

Notebookcheck has done the math and the result is clear: it's technically feasible, but weird . Nvidia has already done it — the GTX 550 Ti mixed 512MB and 256MB chips. But are they going to bother designing a custom PCB with mixed memory densities for a gaming GPU? That's less certain.

The alternative would be 24 GB on 384 bits (12x 2 GB) or 36 GB (12x 3 GB). But 36 GB on an "SE" model when the full fat 5090 has only 32? Nvidia would never do that, the marketing team would tear their hair out.

So either the leak is partially false about the memory config, or Nvidia has a plan that we don't see yet. A Chinese source via Sina also mentions 24 GB, not 32. It would be more technically logical.

That's the question everyone is asking. Historically, a cut-down GPU from the flagship die is a Ti. The RTX 4080 used the AD103, but a version with a reduced AD102 would have been a 4080 Ti — if it had existed. Here, we have a reduced GB202... and would Nvidia choose the name "5090 SE"?

Two hypotheses. The first one is about pure marketing The "5090" badge sells better than the "5080 Ti". That's the same reason why car manufacturers call their SUV a "coupe" — it sounds more premium, even if it's technically debatable.

The second hypothesis is cohabitation with Chinese variants Nvidia already has an RTX 5090D and an RTX 5090D V2 for China — versions limited to comply with US export restrictions. The 5090 SE would be a third variation of the GB202, but this time for the entire world. Keeping the name "5090" in the range allows for segmentation without creating a new line.

Personally, I lean towards a mix of the two. And honestly, the name doesn't matter much — what matters is the performance/price ratio.

Click to enlarge

To understand why this rumor is so exciting, you have to look at the current prices of the RTX 5090 in Europe. Spoiler: it hurts.

The official MSRP of the Founders Edition is 2,349 euros at Infomax — but good luck finding one in stock. The custom AIB models, that's a different story: The ASUS ROG Astral OC sells for around 4,578 euros on Amazon , the ProArt ranges from 4,100 to 4,500 euros instead of its 3,499 euros MSRP, and anniversary editions like the ROG Matrix Platinum exceed 6,500 euros at LDLC. We are dealing with large-scale scalping.

In this context, a $1,500 MSRP card (let's say 1,600-1,700 euros with the conversion and taxes) that offers performance between the 5080 and the 5090? That would interest people. A lot of people.

Except that if availability is the same as for the 5090... you would probably pay 2,500-3,000 euros in practice. And there, the interest collapses.

No. Not in the current state of affairs.

First because this card probably does not exist yet . The leak comes from a single source (GameGPU), the memory specs are dubious, and Nvidia has not confirmed anything. Club386 even published an article titled "No, Nvidia is not launching an RTX 5090 SE." That's the level of reliability.

Then, even if it exists, it would not come out before the THE EARLIEST 2027 — at the same time as the Super refresh. That's six months of waiting minimum for a ghost product.

If you're looking for a high-end GPU right now, the RTX 5080 at $999 remains the rational choice. Want something better? The RTX 5090 - if you can get your hands on one without bleeding. Want something in between the two? For now, it simply doesn't exist.

This leak is interesting for one reason: it shows that Nvidia knows that the gap between the 5080 and the 5090 is too large . 999$ against 1 999$, with nothing in between? It's a huge gap in the lineup. Whether through a "5090 SE," a "5080 Ti," or a beefy "5080 Super," Nvidia will have to fill this void. The question is when, not if.

No. As of July 17, 2026, Nvidia has made no official announcement. Everything is based on a leak from GameGPU picked up by TechPowerUp. Several sites like Club386 consider the rumor to be unreliable.

Concretely, none on the technical level. Both refer to a reduced GB202 GPU positioned between the 5080 and the 5090. The difference is purely marketing — Nvidia seems to prefer keeping the "5090" badge for prestige reasons.

The rumored MSRP is $1,500. With the conversion and European taxes, we would be around 1,600-1,700 euros. In practice, given the state of the GPU market in 2026, one should expect real prices well above — probably 2,000-2,500 euros from retailers.

If it exists, not before early 2027. The most likely window would be CES 2027 in January, potentially at the same time as the RTX 50 Super refresh.

If you need a GPU now, take the RTX 5080. The 5090 SE is a ghost product based on a single leak. Waiting six months for an unconfirmed product is like playing the lottery. The 5080 is available, performs well, and at $999 it's the best performance/price ratio in the Blackwell range.