The story on the radar this morning is that NVIDIA is discontinuing the RTX 5090, with the GB202 die reserved for the RTX PRO line. VideoCardz, Wccftech, Hardware Busters, TechPowerUp and Notebookcheck have all carried it.
Follow the links back and they arrive at the same place: one post by the leaker MEGAsizeGPU, today, saying future GB202 supply is PRO exclusive. Tom’s Hardware asked NVIDIA for comment and had not heard back when it published. There is no announcement, no end of production date, and no second independent source.
We have been here twice in three weeks, so the pattern is worth naming. Eight days ago the headlines said buying a 5090 now legally requires a federal declaration, when what had actually happened was that one retailer adopted a form of its own. Today they say a card is discontinued, when what has actually been reported is a change in where a die gets allocated.
Those two things are different
Ending new allocation of a chip to one product line is not the same as stopping sales of that product.
Dies already in the channel still become cards. Board partners build them, retailers sell them, and the practical effect of a reallocation is thinner restocks and firmer prices rather than bare shelves. Taiwan’s BenchLife frames it as capacity moving to more profitable lines, which is a business decision rather than an end of life notice.
There is a second claim riding along with it, that the top GeForce becomes an RTX 5080 with 24 GB and the 16 GB version goes away. The leaker behind the first claim has not backed that one, and it comes from a different unnamed factory message. Treat it as weaker than a claim that is already single sourced.
If you were about to buy, none of this is a reason to hurry. A rumour at this strength is not a buying signal, and we published the arithmetic on buying versus renting this exact card two weeks ago. Nothing reported today changes a line of it. Thinner supply and firmer prices push gently in the same direction that arithmetic already pointed.
The number that would actually matter
Here is why we are covering a single sourced rumour at all.
For local AI, VRAM is not a spec on a list, it is the whole constraint. Our own build writeup has said to buy VRAM rather than framerate since July, and the reason is arithmetic you can do at home. The text encoder for the video model we publish with is 15.7 GB on its own. A fifteen second clip at full canvas fills our card to around 31 GB. Thirty two gigabytes is not luxurious for this work. It is the amount that makes it possible.
So the question raised by today’s story is not whether a product name survives. It is whether 32 GB remains purchasable in a consumer card at all, or whether the ceiling for someone building a machine at home settles at 24, with 32 and above living behind professional pricing.
That would not stop anyone doing this. It would change what fits. Models get quantised, offloaded and split, and the last two months on this radar are full of people getting large things into small cards. But every one of those techniques costs either quality or time, and they are workarounds for a ceiling rather than a replacement for headroom.
What we are watching
A statement from NVIDIA, or continued silence, which after a week starts to mean something itself.
Whether a second independent source appears, or whether every version of this story keeps resolving to the same post.
And retailer stock, which a reader can check without trusting anybody. If restocks thin out over the next few weeks, the reallocation is real whatever anyone announces.
We will say so here either way, including if this turns out to be nothing.
A daily note from our news radar. The week gets the full treatment in the DIY AI Brief every Monday.