On October 10, 2026, reports surfaced that NVIDIA will end the supply of GPUs for the GeForce RTX 5090. The source was a post by MEGAsizeGPU, a leaker known for hardware information, who claimed that GB202 chips will be redirected to the professional RTX PRO series. Several outlets have since picked up the claim. However, NVIDIA has not officially announced that the RTX 5090 is being discontinued, and even if the rumor is true, no specific end date has been given. Meanwhile, the specifications NVIDIA has announced for the RTX PRO 5500, including 84GB of memory and a GPU partitioning feature, differ from those of a top-end consumer card. The rumor highlights how differently the same Blackwell generation of GPUs is expected to serve consumers and businesses.
What we know so far about the RTX 5090 supply-end rumor
MEGAsizeGPU's claim is that GB202 chips supplied from now on will be dedicated to the RTX PRO series and will no longer be used to build new RTX 5090 cards.
This is not a claim that the specifications of existing RTX 5090 cards will change. It is information about how GPU chips will be allocated to graphics cards manufactured in the future. If true, the chips needed to build new RTX 5090 cards would no longer be supplied.
On the same day, another leaker, hongxing2020, also mentioned the end of the RTX 5090 and a 24GB RTX 5080.
A notice that this leaker reportedly showed says that handling of the RTX 5090, including the D model for the Chinese market, will end, and that a 24GB RTX 5080 will become the top model in the GeForce lineup. However, the origin of the notice cannot be verified, and there is no evidence that NVIDIA has made any official decision.
At this point, all of the information is circulating through reposted post text and images.
The claim that GB202 supply is being redirected and the claim that a 24GB RTX 5080 will become the new flagship should be considered separately. Even if the former proves true, that does not establish the specifications or release plans of the latter.
It is also premature to assume that the RTX 5090 will vanish from store shelves immediately because of the phrase "end of production."
The end of new GPU chip supply, the sell-through of finished-product inventory, and the end of driver support for cards already purchased are three separate matters. Nothing in the current reports indicates when stock will run out or when support will end.
Redirecting GB202 supply is not the same as ending the chip itself
NVIDIA's Blackwell GPU architecture whitepaper presents GB202 as the GPU used in the GeForce RTX 5090.
GB202 is the name of a GPU chip design, and the RTX 5090 is a product built on that chip. Even if supply for the RTX 5090 ends, it does not necessarily mean GB202 itself will stop being produced.
Furthermore, even with the same GPU chip, the number of enabled compute units differs from product to product.
According to the whitepaper, the full GB202 configuration has 192 SMs (Streaming Multiprocessors), which handle compute processing, and 24,576 CUDA cores. The RTX 5090, by contrast, has 170 SMs and 21,760 CUDA cores enabled.
In other words, not every product built on GB202 uses the same number of the chip's compute units.
This helps explain what the rumor about directing GB202 to the RTX PRO series could mean.
The question is not whether production of Blackwell-generation GPUs as a whole is ending, but which products GB202 will be preferentially allocated to.
However, supply volumes for each product and the criteria used to select chips have not been made public. It is therefore impossible to calculate how many RTX PRO cards could be made from chips that had been used for GeForce.
Caution is also warranted regarding the view that NVIDIA prioritized more profitable professional products.
Manufacturing costs, profit margins, and any official policy on this supply allocation have not been confirmed. The specifications of professional products reveal differences in intended use, but they alone cannot explain why NVIDIA would have decided to end the RTX 5090.
84GB of memory and GPU partitioning: what RTX PRO prioritizes
The RTX PRO 5500 Blackwell Workstation Edition that NVIDIA has announced is a professional GPU with 84GB of GDDR7 memory.
NVIDIA envisions it not only for AI workloads and complex 3D data processing that require large amounts of memory, but also for deployments in which GPUs are consolidated in racks and shared by multiple users.
The design priorities differ from those of the GeForce series, which is used in personal PCs for gaming and creative work. Comparing raw compute performance alone does not fully capture the difference between the two.
Comparing NVIDIA's official specifications as of October 10, 2026 gives the following. Note that the RTX PRO 5500 is an upcoming product, and the published specifications may change.
| Item | GeForce RTX 5090 | RTX PRO 5500 Workstation Edition |
|---|---|---|
| GPU memory | 32GB GDDR7 | 84GB GDDR7 (with error correction) |
| Published power spec | TGP 575W | Maximum power consumption 600W |
| Memory allocation unit | 32GB for the whole card | Up to 2 partitions via GPU partitioning, up to 42GB each |
| Product status | Existing product on sale | Upcoming, provisional specs |
Sources: NVIDIA GeForce RTX 5090 official specifications and the official RTX PRO 5500 specifications and FAQ on GPU partitioning. RTX 5090 figures are based on the Founders Edition or reference design and may differ for manufacturers' custom cards.
What distinguishes the RTX PRO 5500 is not just that a single user can access 84GB of memory.
According to NVIDIA, a GPU partitioning feature called MIG (Multi-Instance GPU) can divide one GPU into up to two independent partitions, each with its own dedicated memory, cache, and compute cores.
This allows multiple users to share a single GPU while each uses the compute resources assigned to them.
Comparing nominal memory capacity alone, the maximum 42GB per partition when the RTX PRO 5500 is split in two is 31.25% more than the RTX 5090's total memory of 32GB.
The calculation is as follows:
(42 − 32) ÷ 32 × 100 = 31.25%
In other words, even when the RTX PRO 5500 is split into two, each partition can be assigned more memory than the entire RTX 5090 has.
This is only a comparison of nominal memory capacity, however. It says nothing about which card is superior in compute performance, processing speed, or the free memory actually available to applications.
Still, the difference is important for understanding the uses the RTX PRO series is designed for.
NVIDIA lists large AI models and complex 3D data processing, both of which require large amounts of memory, as primary uses of the RTX PRO 5500. Moreover, when GPUs are consolidated in racks and shared among many users within a company, what matters more than the performance of a single personally owned graphics card is how efficiently compute resources can be allocated across the organization.
On the other hand, more memory does not make every workload faster.
According to NVIDIA's published figures, the RTX 5090 has a memory bandwidth of 1,792GB/s, while the RTX PRO 5500 has 1,398GB/s.
The RTX PRO 5500 can hold more data in memory, but the RTX 5090 is faster at transferring data to and from memory.
The RTX PRO 5500 is therefore not higher-performing than the RTX 5090 in every task; it is better seen as a product that excels where large memory capacity and GPU sharing matter.
Of course, these specification differences do not prove that GB202 supply is being redirected.
Still, if NVIDIA does reduce supply to GeForce and prioritize professional products, its product strategies for consumer and enterprise GPUs could diverge significantly.
Could the rumored 24GB RTX 5080 succeed the RTX 5090?
NVIDIA has not officially announced the 24GB RTX 5080 that hongxing2020 mentioned, either. Its specifications and release date are unknown.
Furthermore, according to NVIDIA's Blackwell GPU architecture whitepaper, the existing RTX 5080 uses GB203, a different GPU chip from the RTX 5090's GB202.
Even if a 24GB RTX 5080 appears, therefore, simply having more memory does not mean it would inherit the RTX 5090's performance.
More memory could allow users to run larger data sets and AI models that previously didn't fit.
However, improvements in compute unit configuration, such as CUDA cores, and in memory bandwidth are not guaranteed. Game frame rates and AI processing speeds will need to be checked against actual product specifications and benchmark results.
Also, becoming "the top GeForce model" is not the same as having "performance equal to or greater than the RTX 5090, the previous flagship."
Even if the RTX 5090 is discontinued and a 24GB RTX 5080 becomes the top of the GeForce lineup, that alone does not make it a full successor in performance.
The key questions going forward are whether NVIDIA officially confirms the end of RTX 5090 supply and whether the rumored 24GB RTX 5080 actually arrives.
Once both are clarified, users will be able to weigh alternatives to the RTX 5090 not just by memory capacity but also by compute performance, memory bandwidth, and price.
Meanwhile, in professional GPUs, large memory capacity and compute-sharing features are growing in importance. Whether or not the supply-end rumor proves true, it is clear that even on the same GPU architecture, the value expected of products differs between consumers and businesses.