Purchases and Costs
Last updated
RTX 5090 Discontinuation Rumors: Should You Buy One for Local AI?
Separate the supply story from the local-AI capability you actually need.
The RTX 5090 discontinuation rumor spread with an October 10 claim that future GB202 supply would go exclusively to RTX PRO products. The NVIDIA material checked at publication did not announce an RTX 5090 production-end or support-end date. Owners can keep using their existing local-model setup; prospective buyers should compare the memory they need with actual asking prices.
The October 10 claim concerns GB202 allocation
The RTX 5090 is an NVIDIA Blackwell graphics card with 32 GB of VRAM1. It is used for local LLMs and image generation that load model data into GPU2 memory. GB202, the chip mentioned in the news, is the GPU used by this card.
On October 10, MEGAsizeGPU claimed that future GB202 supply would be reserved for RTX PRO products. hongxing2020 also posted about ending RTX 5090 production and a 24 GB RTX 5080; VideoCardz reported those posts.
There is no official notice yet confirming a production plan or end date. The RTX 5090 reports remain a discontinuation rumor, not a confirmed decision.
A retailer's out-of-stock notice describes inventory you can buy now. Future factory output and the driver-support schedule are separate manufacturer decisions. If you are shopping, check the retailer's restock estimate and NVIDIA's product announcements separately.
On October 16, NVIDIA said the RTX 5090 is not discontinued
At the RTX Spark launch event in Taipei on October 16, NVIDIA senior vice president Jeff Fisher told 4Gamers that supply is constrained but the RTX 5090 is not discontinued and will keep shipping. Around the same date the card disappeared from the GPU filters on NVIDIA's US Marketplace, and Micro Center changed its listings to "Not Carried".
The denial means the chip line continues; it does not mean more cards are shipping. Fisher gave no shipment volume or GB202 allocation figure, and in Germany the cheapest available card was €6,699 on October 15, nearly double the €3,400 January reference price. Decide whether to buy or keep your current card from shelf prices and stock.
If you own a 5090, keep working with your current setup
If your model works with the current driver and runner, you can keep using that setup. Changes to new-card supply leave your installed model files and existing compute features available. Record the driver, runner version, and model filename so you can restore a known working setup after an update.
Keep the seller details, receipt, and board manufacturer's warranty terms for a card still under warranty. Reduced supply could lengthen replacement availability, so ask the board manufacturer about repair and replacement procedures. Check its warranty alongside NVIDIA's product news.
If heat and fan noise become a problem during long local-AI sessions, try a power limit. Compare power draw, answer speed, and temperature using the same model and input, then decide whether the performance trade-off works for you. See the GPU power-limit guide for device-specific adjustments including the 5090.
Before buying, decide whether your workload needs 32 GB
If 32 GB of VRAM is why you want a 5090, choose the model and context length3 first. A short chat with quantized Qwen3.8 27B, a long-document session, and keeping an image model loaded at the same time use different amounts of memory. Defining the workload lets you decide whether a higher asking price is worth paying.
Official specifications list 32 GB for the RTX 5090, 16 GB for the RTX 5080, and 24 GB for the RTX 3090. The rumored 24 GB RTX 5080 does not yet have confirmed pricing, availability, and local-AI performance suitable for a purchase comparison. Compare products you can order using their current specifications and prices.
| Graphics card | VRAM | Use in a comparison |
|---|---|---|
| RTX 5090 | 32GB | More room for long context or larger models on one card |
| RTX 5080 | 16GB | Compare actual speed for workloads that fit 16 GB |
| RTX 3090 | 24GB | A baseline when keeping an existing card |
NVIDIA published specifications for the RTX 5090, 5080, and 3090. The unannounced 24 GB RTX 5080 is excluded from the table.
If 24 GB finishes your workload, keeping your current device is an option. If you repeatedly exceed it and stop working, compare a single 32 GB card, multi-GPU partitioning, and larger-memory devices together with cost and runner support.
Include the whole PC in your budget
Even a card-only upgrade requires checking case length and thickness, power connectors, and the power supply. NVIDIA lists 575 W total graphics power and a 1,000 W required system power supply for the RTX 5090 reference specifications. Match your purchase to its board manufacturer's dimensions and power requirements.
Our checked RTX 5090 price range is ₩8,989,320~₩9,289,050, dated 2026-10-12. The retailer's live checkout price can differ. Add any power-supply or case replacement when comparing alternatives; estimate ongoing costs with the electricity tool.
If you want a large model, try Strata too
Strata runs a large MoE4 model such as Qwen3.8-Flash-Next using the GPU, system RAM5, and CPU6 together. It caches frequently used expert weights on the GPU and uses RAM and CPU for the remainder. Published setups include a 12 GB GPU with sufficient system RAM, so trying a large model may be possible on your existing PC.
Match your current RAM and GPU against the Strata hardware guide. After installation, send your actual prompts7 and long documents and record the first-answer and completion times. If it finishes your work, you can use local AI while waiting for clearer 5090 supply information.
If you already own a 5090 and want to try Strata, continue with the RTX 5090 Strata guide. It covers model choice for 32 GB of VRAM and your RAM capacity, long-context settings, and published run conditions.
Terminology notes
VRAM — Memory used by a graphics card’s GPU to store model weights and intermediate values.
Back to the textGPU — A processor designed to handle many calculations in parallel. It performs model computations during AI inference.
Back to the textContext window — The token span of input and generated content a model can handle in one request. The supported limit and memory use depend on the model and runtime settings.
Back to the textMoE — A model architecture that selects some of several expert subnetworks for each input. Total parameters can differ from the number activated for one token.
Back to the textSystem RAM — System memory that temporarily holds data while programs run.
Back to the textCPU — The central processor that runs the operating system and general-purpose instructions; in AI workloads it also handles tasks such as preprocessing and data movement.
Back to the textPrompt — The input text sent to a model. For generative models it describes the desired result and strongly shapes the output.
Back to the text
Frequently asked questions
- What should you check first in this article?
- The RTX 5090 discontinuation rumor spread with an October 10 claim that future GB202 supply would go exclusively to RTX PRO products. The NVIDIA material checked at publication did not announce an RTX 5090 production-end or support-end date. Owners can keep using their existing local-model setup; prospective buyers should compare the memory they need with actual asking prices.
- How much memory does it need?
- The RTX 5090 is an NVIDIA Blackwell graphics card with 32 GB of VRAM. It is used for local LLMs and image generation that load model data into GPU memory. GB202, the chip mentioned in the news, is the GPU used by this card.