Leakers say Nvidia has stopped sending GB202 chips to GeForce and reserved them all for RTX PRO. Nvidia has said nothing. Meanwhile a $1,999 card is selling for $6,500 to $9,500 and has vanished from first-party US retail. Here is what is confirmed, what is rumour, and what you should actually buy.
Do not buy an RTX 5090 at today's prices. A GPU with a $1,999 launch MSRP is now listed at ~$6,500–$9,500 by third-party sellers after disappearing from first-party stock at Newegg, Amazon, Best Buy and B&H — with premium liquid-cooled models near $10,000. Production may already have stopped: on 10 October 2026 leakers MEGAsizeGPU and hongxing2020 reported that Nvidia has stopped allocating GB202 dies to GeForce and reserved all future supply for RTX PRO, citing a factory notice that also claims a 24GB RTX 5080 becomes the new top GeForce and the 16GB RTX 5080 is discontinued. Nvidia has confirmed none of it — no announcement, no denial, no date. Buy instead: at 1440p the AMD Radeon RX 9070 XT 16GB at ~$720–850 (MSRP $599); with ray tracing on, the RTX 5070 Ti 16GB at ~$1,000–1,250 (MSRP $749); at 4K the RTX 5080 16GB at ~$1,400–1,700 (MSRP $999); and the best raw performance per dollar anywhere is a used RTX 4090 24GB at ~$2,200–3,200. If you genuinely need 32GB of GDDR7 and CUDA, the only defensible route in the US is in-store Micro Center stock from about $4,299 — still 2.15x MSRP, in person only. Everything below is priced from real October 2026 listings, and every figure moves weekly.
Let us separate the two stories, because they are not the same story and most coverage runs them together.
On 10 October 2026, hardware leaker MEGAsizeGPU (@Zed__Wang) posted that Nvidia had stopped allocating GB202 silicon to the RTX 5090, adding in a follow-up: "All future GB202 supplies are RTX Pro exclusive." A second leaker, hongxing2020, had posted "5090 EOL, 5080 24GB ready" the same day alongside what he described as a factory notification:
Two things to hold onto. First, MEGAsizeGPU confirmed the GB202 allocation change but has not confirmed the 24GB/16GB part of the story — so the 5090 half of the leak has two independent sources and the 5080 half effectively has one. Second, Nvidia has neither confirmed nor denied any of it, has issued no end-of-life notice and has published no end-of-production date.
BenchLife's reading, echoed by Notebookcheck, is the most useful framing: this is production capacity being moved to more profitable product lines, not a formal EOL. Stopping new GB202 allocations is not the same as discontinuing a card — Nvidia's partners, distributors and retailers hold inventory of chips, boards and finished cards, and none of that disappears when new allocations stop. So the honest position is: widely reported as ending production, unconfirmed, with existing stock still trickling out.
This part needs no leaks. You can verify it in thirty seconds by searching for an RTX 5090 on Newegg, Amazon, Best Buy or B&H and setting the filter to products sold by the retailer itself.
The headline published price is from January 2025. Nothing else in this table is.
| Where you buy | ~Price (Oct 2026) | vs $1,999 MSRP | Availability note |
|---|---|---|---|
| Nvidia MSRP (Founders Edition) | $1,999 | — | The number on Nvidia's own spec page. Effectively theoretical |
| Micro Center (in store only) | ~$4,299 | +115% | The only near-sane US route. In person, per-store, moves daily |
| US 90-day median, new retail | ~$4,900 | +145% | Median of 45 tracked captures; range $4,100–$6,995 |
| Used / resale (trackers) | ~$3,800–4,900 | +90–145% | Best resale price hit $4,795 in September — 2.4x MSRP |
| Newegg marketplace, cheapest in stock | ~$6,500 | +225% | Founders Edition via a marketplace seller, 6 Oct 2026 |
| Third-party sellers, typical | ~$6,500–9,500 | +225–375% | Newegg, Amazon and Best Buy marketplace listings |
| Premium OC / liquid-cooled flagships | ~$8,700–10,000 | +335–400% | ROG Astral, AORUS, Zotac SOLID OC, ROG Matrix Platinum |
| EU street (for reference) | €5,000–6,000 | +150–200% | VAT included; Ireland tracked at €5,999 on 1 Oct |
All figures are typical US list and street ranges observed in September–October 2026, converted from retailer and tracker data, and fluctuate daily with stock, region, seller and tax treatment. Taiwan, Hong Kong and EU pricing differs substantially and is generally higher. Confirm the current price and stock on a reputable retailer before buying.
The 5090 is not expensive because it is scarce in the way a limited-run product is scarce. It is expensive because four things collided at once, and only one of them is specific to this card.
AI accelerators use HBM, not GDDR7 — but HBM and graphics memory share the same three suppliers (Samsung, SK Hynix and Micron, who together hold the overwhelming majority of DRAM production) and the same wafer capacity. When those fabs swing toward HBM, which carries several times the margin of commodity memory, the wafer starts available for GDDR fall even though no AI accelerator consumes a single GDDR7 chip.
The numbers are brutal. GDDR6 spot pricing rose from roughly $2.50 to $7.50 per gigabyte since late 2025. Conventional DRAM contract pricing climbed about 90–95% in Q1 2026, another 58–63% in Q2 and a further 13–18% in Q3. The denser 3GB GDDR7 modules cost $60–70 each against roughly $20 for the 2GB modules in today's RTX 50 cards — triple the price for 50% more capacity. AMD notified partners of a GDDR6 kit price increase effective 1 October 2026, its fourth memory-driven adjustment in ten months.
An RTX 5090 carries 32GB of GDDR7 on a 512-bit bus. It is the most memory-exposed consumer graphics card on the market, which means it absorbs more of this than anything else in the stack. That is why it moved 132% year over year in German tracking while the rest of the RTX 50 family moved 66%.
Export-unrestricted AI buyers have been purchasing retail RTX 5090s for local inference, because 32GB of very fast memory at anything under $10,000 is cheap compared with enterprise hardware. When a card becomes a tool rather than a toy, its price stops tracking its value to gamers. That is the difference between a card being "in demand" and a card being "bought by businesses with a budget line".
This is the leak's core, and its economics are not controversial even if the reporting is unverified. The same GB202 die that powers a $1,999 gaming card powers the RTX PRO 5500 with 84GB of memory at the same 21,760 CUDA cores, and the RTX PRO 6000 Blackwell with 24,064 cores and 96GB at prices that run into five figures. If wafer capacity and GDDR7 supply are both constrained, routing every die to the four-to-five-figure product is straightforwardly more profitable. Gamers are not the customer being chosen here.
Nvidia lifted GDDR6 and GDDR7 kit prices in July 2026 and has raised graphics card prices three times across the year, by 20–30% on some models. AMD has raised Radeon RX 9000 pricing by up to 20%, with the RX 9070 XT moving from roughly $830 to over $1,000 in Asian channel reporting. A separate October leak suggests TSMC may raise advanced-node manufacturing prices around 10% in Q4 2026 — another layer on top.
This is the actual answer to the question you came here with. Every card below is available in the US, Taiwan and Hong Kong, and every price is a real October 2026 range rather than a launch number.
The thing people are waiting for is the RTX 5080 24GB, and the honest answer is that waiting is unlikely to pay.
The rumour is real and reasonably well-sourced. The same factory notice cited in the GB202 leak said the highest-end GeForce would become a 24GB RTX 5080 with the 16GB model discontinued — and the leaker who confirmed the GB202 half has not confirmed this half. BenchLife reports it for Q1 2027, making CES 2027 the logical reveal, using 3GB GDDR7 modules to reach 24GB on the same 256-bit bus with the same 10,752 CUDA cores as today's RTX 5080. Reportedly expected retail is a very wide $1,499 to $2,999.
Here is the problem. Leaked partner information says the RTX 50 Super cards are already built and boxed and sitting in warehouses, held back specifically because 3GB GDDR7 modules cost $60–70 each and the margin does not work. The entire Super refresh slipped from 2026 into early 2027 for memory-pricing reasons, and the 5070 Super was shelved on the same excuse months earlier. So the card you are waiting for is delayed by exactly the force that is supposed to make waiting worthwhile — and Micron says that force gets worse into 2028.
Two more things. First, the RTX 60 series is not expected until late 2027 or 2028, so there is no generational step coming behind this. Second, a 24GB RTX 5080 at a plausible $1,700–2,200 is not an upgrade over a 4K card you can buy today — it is the same class of card with more memory, at whatever price Q1 2027 memory costs dictate.
Our position: if you need a GPU now, buy one that exists now at a price you can defend in twelve months. If you are upgrading from an RTX 4080-class card or newer and nothing you play is bottlenecked, there is no harm in sitting out the whole cycle — but that is patience, not strategy, and it should not be dressed up as waiting for a specific card that has not been announced.
Every figure in this review is one of three things: something we checked ourselves, something the manufacturer publishes, or a figure reported by a named third party. This log separates the three and dates the check, so you can see what is verified before you spend money.
| Check | Result on October 19, 2026 |
|---|---|
| NVIDIA product page (RTX 5090) | HTTP 200 — the specification page still lists the card at $1,999 MSRP. |
| Discontinuation claim | The report that GB202 allocation moved to RTX Pro is from a single supply-chain report and NVIDIA has not confirmed it. The review labels it as unconfirmed. |
| Street prices | ~$6,500-9,500 online, ~$4,299 at Micro Center in store, $3,800-4,900 used — October 2026 figures. |
| Item | Value |
|---|---|
| MSRP | $1,999 — effectively theoretical |
| Online street price | ~$6,500-9,500 |
| In-store price | ~$4,299 at Micro Center, in person only |
| Used market | ~$3,800-4,900 |
| Memory context | GDDR6 spot rose from ~$2.50 to ~$7.50 per GB; DRAM contract prices +90-95% in Q1 2026 |
What we did not test: These are manufacturer specifications and published third-party figures, not measurements from our own lab. We did not buy or bench-test this hardware, and any performance number that is not attributed to a named tester is a vendor claim.
These are the five cards that realistically take the 5090's place in a 2026 build, ordered by what most buyers should look at first.
Best for: 4K gaming on a new card, if you can live with 16GB — ~$1,400–1,700 (MSRP $999)
Best for: the best frames per dollar in the entire market — ~$720–850 (MSRP $599)
Best for: 1440p on the tightest sensible budget — ~$630–720 (MSRP $549)
Best for: 1440p with ray tracing and DLSS on — ~$1,000–1,250 (MSRP $749)
Best for: maximum performance per dollar if you accept the risks — ~$2,200–3,200
| Card | ~Street price | MSRP | Best for |
|---|---|---|---|
| GeForce RTX 5090 32GB | ~$6,500–9,500 (in-store from ~$4,299) | $1,999 | Do not buy at street price. Only for absolute top-end work if a buyer's budget absorbs it |
| GeForce RTX 5080 16GB | ~$1,400–1,700 | $999 | 4K gaming on a new card, without the flagship premium |
| GeForce RTX 5070 Ti 16GB | ~$1,000–1,250 | $749 | 1440p with ray tracing, DLSS 4 Multi Frame Generation |
| GeForce RTX 5070 12GB | ~$700–800 | $549 | 1440p on a budget, with Nvidia features. 12GB is the limit |
| Radeon RX 9070 XT 16GB | ~$720–850 | $599 | The best frames per dollar in the market. 1440p high refresh |
| Radeon RX 9070 16GB | ~$630–720 | $549 | 1440p when the XT has drifted past $850 in your region |
| Radeon RX 9060 XT 16GB | ~$380–470 | $349 | 1080p high refresh, and the cheapest sensible 16GB card |
| Used RTX 4090 24GB | ~$2,200–3,200 | $1,599 (2022) | The best raw performance per dollar, if you accept no warranty |
| Rumoured RTX 5080 24GB | Not on sale (reported ~$1,499–2,999) | — | Unannounced. Reportedly Q1 2027. Do not plan your build around it |
Typical US street ranges observed September–October 2026, in USD, excluding tax and shipping. Prices move weekly and vary substantially by region — Taiwan, Hong Kong and the EU generally sit higher. Confirm on a reputable retailer before buying. We are not affiliated with any brand listed here.
Officially, no — and that distinction matters. On 10 October 2026 the leakers MEGAsizeGPU (@Zed__Wang) and hongxing2020 independently reported that Nvidia has stopped allocating GB202 dies to GeForce and that all future GB202 supply is reserved for the RTX PRO professional line, with hongxing2020 citing a factory notice saying Nvidia will no longer sell any RTX 5090 model including the China-market D variants. Nvidia has neither confirmed nor denied any of it, has made no end-of-life announcement and has published no end-of-production date. Analysts also draw a real line between stopping new chip allocations and discontinuing the card: Nvidia and its board partners can keep selling 5090s built from the inventory already in the pipeline, and cards already sitting in distributor and retail warehouses do not vanish when new allocations stop. So the accurate framing is that the RTX 5090 has been widely reported as ending production, unconfirmed, with existing stock still trickling out — and that first-party retail availability in the US collapsed before the leak did, which is the part you can verify yourself in about thirty seconds.
Three forces stacked on the same card. First, memory. GDDR6 spot pricing rose from roughly $2.50 to $7.50 per gigabyte since late 2025, conventional DRAM contract pricing climbed about 90-95% in Q1 2026, another 58-63% in Q2 and 13-18% in Q3, and the 3GB GDDR7 modules the newest cards need cost $60-70 each against roughly $20 for the 2GB modules in existing RTX 50 cards. An RTX 5090 carries 32GB of GDDR7 on a 512-bit bus, so it is the most memory-exposed consumer card made. Second, AI demand. Export-unrestricted AI buyers have been purchasing retail 5090s for local inference because 32GB of fast memory at any price under $10,000 is cheap compared with enterprise hardware. Third, supply. If GB202 dies are genuinely being routed to RTX PRO cards that sell for four to five times the GeForce price, the consumer flagship loses its silicon while demand is still rising. German price tracking put RTX 50-series cards up 66% on average year over year in October 2026, with the RTX 5090 up 132%, and the card has now vanished from first-party stock at Newegg, Amazon, Best Buy and B&H, leaving third-party sellers at roughly $6,500-$9,500.
At $6,500-$9,500, no. That is 3.2x to 4.8x the $1,999 MSRP for a 575W consumer card with a consumer warranty, bought from a marketplace seller who may not offer returns. The comparison that settles it: at $9,500 you are two-thirds of the way to an RTX PRO 6000 Blackwell with 96GB of ECC memory and enterprise support, and a used RTX 4090 at roughly $2,200-$3,200 delivers 24GB and very strong 4K performance for a quarter to a third of the price. If you genuinely need 32GB of GDDR7 and CUDA in one card, the single defensible route in the US is in-store Micro Center stock, which has been seen from about $4,299 — still 2.15x MSRP, in-person only, per-store and moving daily, and not available in Taiwan or Hong Kong at all. Buying a 5090 at $6,500-plus in October 2026 means paying the top of a price cycle that memory-supply forecasts say will not ease before mid-to-late 2027, on a card whose production has reportedly already stopped. If you want flagship-adjacent performance now, buy an RTX 5080, an RTX 5070 Ti, an RX 9070 XT or a used RTX 4090 and upgrade the rest of the machine with the difference.
It depends almost entirely on resolution and budget, and every one of these is available in the US, Taiwan and Hong Kong. At 1080p high refresh, the RX 9060 XT 16GB at roughly $380-470 is the value pick and 16GB is the floor you should not go under. At 1440p, the AMD Radeon RX 9070 XT 16GB at roughly $720-850 is the frames-per-dollar leader, with the RX 9070 16GB at about $630-720 just beneath it; if you play with ray tracing or DLSS on, the RTX 5070 Ti 16GB at roughly $1,000-1,250 is the better card and the better all-rounder. At 4K, the RTX 5080 16GB at roughly $1,400-1,700 is the sensible new-card ceiling, and a used RTX 4090 24GB at roughly $2,200-3,200 is the best raw performance per dollar in the entire market right now. The one card we would tell you not to buy at any of today's prices is an RTX 5090 from a third-party marketplace listing.
It is a well-sourced rumour, not a product. The same 10 October 2026 factory notice that reported the RTX 5090's end also said the highest-end GeForce would become an RTX 5080 24GB and that the 16GB RTX 5080 would be discontinued — and MEGAsizeGPU, the leaker who confirmed the GB202 allocation change, has not confirmed that part. Taiwan's BenchLife reports the 24GB card is planned for Q1 2027, making CES 2027 the logical reveal, and it would use 3GB GDDR7 modules to reach 24GB on the same 256-bit bus and the same 10,752 CUDA cores as today's RTX 5080. Reportedly expected retail is $1,499 to $2,999, which is a wide band and exactly the range the current RTX 5080 already occupies. The reason it is not here yet is memory cost: leaked partner information says the RTX 50 Super cards are built and boxed and sitting in warehouses awaiting a pricing decision, because 3GB GDDR7 modules cost $60-70 each against $20 for 2GB. Our read is that waiting through Q1 2027 for an unannounced card, in a market where Micron expects the memory shortage to worsen into 2028, is not a plan — it is a hope. Buy a card that exists now if you need one now.
A used RTX 4090 is a reasonable and often smart buy; a used RTX 5090 at current prices is not. On the 4090, recent completed eBay sales put clean examples at roughly $2,200-$3,200 depending on cooler and condition, and sold-price trackers show medians clustering near $2,800-$3,200 for well-presented cards with a 90-day range of about $2,000-$3,010. What you give up is the warranty — most 4090s are out of their original term or the cover does not transfer — plus you inherit a card that may have spent two or three years at full load in an AI or mining rack, and the 12VHPWR connector history that generation is known for. Practical rules: buy from a seller with a return window and verified sold history rather than the lowest asking price, insist on a photo of the card running GPU-Z and a stress test, avoid anything sold as untested or for parts, and remember that a real bargain below the market floor is usually a card with a problem. On the 5090, prices between $6,500 and $9,500 put a used or scalped card in the same bracket as new professional hardware, with no meaningful warranty and no way to verify what the card was used for. There is no version of that trade that makes sense for a gamer.
Independent reviews of AI tools: real pricing & free alternatives.
📒 Bookkeeping ReviewsHonest accounting software reviews: Wave, QuickBooks, Xero & free options.
📄 PDF Merge & ToolsMerge, split, compress and convert PDFs — free, private.
🪪 ID Photo StudioFree passport & visa photos with AI background removal.
🧮 HK Salary Tax CalculatorEstimate Hong Kong salaries tax with official IRD rates — see what you owe before you file your tax return.
💝 HK Benefits GuideCheck which Hong Kong government benefits you qualify for.
📅 HK Calendar 香港月曆Public holidays, lunar dates & solar terms for 2020-2036 — bilingual, free, no sign-up.
★ Our own product 🏠 Home Expense BookkeepingFamily bookkeeping: income, spending, net worth and monthly bills. Data stays on your device. HK$98 one-time.