Dark square chip package standing on its edge against a plain background, illustrating GPU die allocation

The Die That Went to the Workstation: What GB202 Allocation Says About the 32 GB Tier

What actually happened

On October 10, 2026, the leaker MEGAsizeGPU posted a short confirmation and a follow-up: no more GB202 for the RTX 5090, and all future GB202 supply goes to RTX Pro products. A second leaker, hongxing2020, posted “5090 EOL, 5080 24GB ready” alongside what appeared to be a factory notice.

Nvidia’s answer, on October 17, was narrower than either claim. Jeff Fisher said the card has not been discontinued and will continue to ship when possible, and that Nvidia is doing everything it can to get the product on the market. He also confirmed the operational detail that matters: some GB202 production is being prioritized for RTX Pro GPUs.

Those two statements are not in conflict. “Not discontinued” is a product-status claim. “No longer able to order” is a supply claim, and it is the one that changes what you can actually buy.

The clearest evidence on the supply side came from an anonymous Nvidia board partner, reported by der8auer. The partner said it has not been told the 5090 is end-of-life, but it can no longer purchase the GB202 chip or the 32 GB of GDDR7 needed to build a card. It also said it can no longer place backorders. Historically, backorders were always eventually filled once stock came back. A backorder you cannot place is not a shortage. It is a reallocation.

Market behavior matches the reallocation, not the shortage. Nvidia removed all RTX 5090 SKUs from its own store. Micro Center switched most of its listings from “out of stock” to “not carried.” Tom’s Hardware found the card priced at $8,500 on Amazon on October 17. GPUPoet’s October 2026 market report, compiled from August and September listings, found 59 of 81 tracked GPUs got more month over month expensive and only 17 got cheaper, with the 5090 at a best resale of $4,795, up 23 percent from August.

Why one die goes to two different customers

GB202 is a single silicon part. The RTX 5090 uses 21,760 CUDA cores of it. The RTX PRO 6000 Blackwell uses 24,064. Both run a 512-bit GDDR7 interface at 28 Gbps effective. TechPowerUp lists the same 1.79 TB/s of memory bandwidth for both cards; Nvidia’s own product page for the server edition of the PRO 6000 states 1,597 GB/s, so the two vendor sources do not agree on the workstation number. The difference between them is not compute. It is the memory bill: 32 GB on one card, 96 GB on the other.

That is where the allocation decision gets made. The scarce input is not the die. It is the stack of GDDR7 sitting next to it. A workstation card carries three times the memory content on the same interface and sells at a list price in the $16,000 range, while a July 2026 street tracking put the same card at $8,499. The gaming card was set at $1,999 and was expected to move another $300 purely because of GDDR7 cost.

When a supplier can put the same wafer into a product that carries three times the memory content at eight times the price, the gaming SKU is not a design decision. It is an allocation decision. And allocation decisions reverse when demand cools. Product status does not.

What the reallocation produces at retail

Card Launch MSRP Verified price point Source and date
RTX 5090 32GB $1,999 $8,500 Amazon listing Tom’s Hardware, Oct 17, 2026
RTX 5090 32GB $1,999 $4,795 best resale, +23% MoM GPUPoet, Oct 2026 report
RTX 5090 32GB $1,999 $6,549 to $9,300 third-party, one listing above $10,000 TweakTown, Newegg third-party
RTX 5080 16GB $999 $1,899.99, ASUS TUF, its own recorded high Camel price history, Oct 15, 2026
RTX 5080 16GB $999 $1,899.99 third-party, GIGABYTE WINDFORCE SFF, its own recorded high Camel price history, Oct 7, 2026
RTX PRO 6000 96GB not published $8,499 street / $16,000 list qscompute July 2026 tracking, TechPowerUp

Every row above is a price someone could actually pay on a real page. The pattern is consistent: the premium widens as the memory content grows. The 5060, which uses less memory and draws no interest from AI buyers, stayed close to its original price. The 5090, which carries the most, went to 2.4 times its MSRP.

What this means if you run models locally

For a local build, the 32 GB tier is the tier that decides whether a model lives on the card or spills off it. A 32B model at Q4_K_M is roughly 19 to 20 GB of weights. That fits in 32 GB with room for the key-value cache at a normal context length. It does not fit in 16 GB.

The decode ceiling on that card is arithmetic, not marketing. At the 1,792 GB/s TechPowerUp lists for the 5090, reading a 20 GB weight file once per token gives an upper bound of about 90 tokens per second. Real engines land lower than that because of request overhead, cache management, and the fact that a decode pass never reads memory at the theoretical peak. On my own Ryzen 9 5900X I measured 17.9 GB/s of copy bandwidth against a 42.7 GB/s theoretical channel, which is 42 percent of the channel counted as data moved. The same gap exists on every GPU memory system. The ceiling is a ceiling, not a delivery.

So the practical question is not whether the 5090 is discontinued. It is whether a 32 GB card is worth paying 2.4 times MSRP for when the alternative is a 16 GB card at 1.9 times MSRP plus offload. Offload runs at system memory speed. A consumer dual-channel DDR5 board is on the order of 40 to 50 GB/s theoretical, and the measured copy pass on my own board was under 20 GB/s. A layer that spills off the card runs at that speed, not at 1,792 GB/s. That is the whole cost of the 16 GB tier, and it is why the price gap between 16 GB and 32 GB cards has widened faster than the price gap between any other pair.

The 24 GB card that was supposed to answer this

The RTX 5080 24 GB was the mid-cycle fix. Reports put it in Q1 2027, and board partners were still in the dark about it as of October 12, 2026. The design work on the refresh was reportedly finished. It has not shipped. Nvidia never announced it as a public product, so there is nothing to call cancelled. The practical effect is that the 16 GB cap stays in place for at least another two quarters.

The RTX 60 series is reported to slip to 2028, and no outlet that published the estimate has walked it back. Nvidia has not confirmed or disputed it. The one part of the story that has moved from rumor to consensus is the cause: memory and packaging supply, not design readiness.

Three paths, with prices

Pay the auction price

If you need 32 GB of GDDR7 today, you are buying from third-party stock, not from a manufacturer. The MSI GAMING TRIO OC is the clearest example of what that looks like on a product page: the listing is live, the Amazon price is out of stock, and the last recorded third-party price was $1,399.99 in June 2026. Listings that were in stock in October ran from $6,549 to above $10,000. You are not buying a card. You are buying the last allocation of one.

Buy the 16 GB tier and accept the offload

The 16 GB cards are still orderable, which is the actual difference between them and the 5090. The ASUS TUF RTX 5080 was at $1,899.99 on October 15, 2026 against a $1,699.99 list price. The GIGABYTE WINDFORCE SFF has never been tracked below its $999.99 list price, and the last recorded third-party price was $1,899.99 on October 7, 2026. That listing is out of stock. Both are real purchases. Neither fits a 32B model. The decision is whether your workload is a 14B model that lives entirely on the card, or a 32B model that does not.

Buy capacity instead of bandwidth

The path that does not compete for GDDR7 is a unified-memory machine. A Ryzen AI Max+ 395 box with 128 GB was at $3,599 on October 8, 2026. It runs a 70B model that a 32 GB card cannot load at all, at roughly 256 GB/s. That is a trade of speed for capacity, and it is the same trade the memory market is making on the other side of the counter. If the model you actually run is bigger than 32 GB, the bandwidth ceiling on local decode matters less than the fact that you can load the model. The ceiling on local decode is a hard number on every one of these machines, and paying 2.4 times MSRP does not move it.

The memory price is the reason, and it is still climbing

TrendForce projected conventional DRAM contract prices up 10 to 15 percent quarter over quarter in Q4 2026 and NAND up 15 to 20 percent. Consumer DRAM demand came partly from SSD customers, and suppliers stayed committed to production cuts. The retail end of that is measurable. The DDR5 median tracked at $18.75 per GB on October 27, 2026, up 4.3 percent in 30 days. A 32 GB CL30 kit that listed at $101.99 in 2023 and bottomed at $92.99 in April 2025 was at $599.99 on October 14, 2026. That is the same DRAM price squeeze that took wafer capacity from consumer modules, and it is the reason a 16 GB card costs what it costs.

The inference cards that skip HBM entirely are the other response to the same constraint. Intel’s GPU D-Series puts 480 GB of LPDDR5X on an air-cooled PCIe card, which is a different answer to the same question: how do you serve a model without paying the stacked-memory price. That is worth reading next to this story, because both are attempts to route around the same bill.

What to watch

Three things decide whether this reverses.

First, whether board partners can place backorders again. That is the operational signal, not the statement about product status. A company that says a product is not discontinued and cannot sell you the parts to build it has described a supply position, not a roadmap.

Second, whether GDDR7 allocation shifts. Micron said more than 75 percent of its expected fiscal 2027 output is already committed under long-term agreements. SK hynix’s NAND capacity is reportedly booked through the end of 2026. New capacity from the major makers is not expected to meaningfully increase supply until mid-to-late 2027, and ASML has said AI demand exceeds chip-tool supply through 2028.

Third, whether consumer demand absorbs the prices it is being quoted. PC brands are reducing SSD capacities in mainstream models to lower bill-of-materials costs. That is what a market looks like when it stops absorbing a price, and it is the only mechanism that would pull GDDR7 allocation back toward gaming cards.

For a local build, the practical conclusion is narrow. The 32 GB tier is priced at 2.4 times its launch price because the memory on it is worth more somewhere else. That is not a card problem. It is an allocation problem, and allocation moves when the price stops being absorbed. Until it does, the buying decision is about which constraint you accept: bandwidth, capacity, or price. The hardware ladder covers the tiers, and the bandwidth per dollar arithmetic is what decides which tier is worth buying at any given price. Right now the arithmetic does not favor the top of the stack.

1 Comment

Leave a Reply

Your email address will not be published. Required fields are marked *