Nvidia has reportedly stopped making the only consumer graphics card with 32 gigabytes (GB) of video memory [1]. That silicon is going to professional boards instead. Nine days earlier, on 1 October, the discount cloud firm Nebius raised its pay-as-you-go rental rates for the second time in three months. Its largest increase, 41%, landed on memory rather than on compute [3]. Neither move is about raw speed. In both, what is scarce and what sets the price is memory space.

The big picture:

The choice to build or rent has always rested on two numbers: the price of a card, and the price of an hour. In the past 30 days, one input moved both.

On 10 October, Tom’s Hardware reported that Nvidia has stopped supplying the GB202 chip to makers of consumer cards. That is the die inside the 5090. The outlet cited a leaked factory notice, plus a second hardware leaker who said future supply of the die is reserved for data-center and professional cards [1]. Nvidia has not confirmed it. It has given no end date, and it did not answer the outlet’s question [1].

The part I keep circling is what Nvidia did list. In mid-September, with no press release and no blog post, product pages appeared for its 5500 Blackwell Workstation Edition. The card has the same 21,760 cores as the 5090, joined to 84 GB of GDDR7 memory that carries error-correcting code (ECC) [2]. It splits into one 84 GB slice or two of 42 GB. There is no price, no ship date, no speed figure, and the spec table is marked preliminary [2].

Nvidia points the card at rack-mounted gear. Its own copy says graphics chips (GPUs) are centralized “into managed racks” and shared across users [2]. That is an IT purchase, not a desk purchase. The money is blunt: the same die on a professional board sells for “substantially more”, and the full 96 GB professional card “can easily sell for $15,000” [1].

By the numbers

  • 84 GB vs 32 GB — Same core count: the professional card Nvidia listed in September carries about 2.6 times the video memory of the consumer card it shares a core count with. Nvidia has published no price for it [1][2].
  • 41% — The memory line: Nebius lifted its rate on memory from 1 October, against 17% to 21% on graphics chips and 25% on some processor-only instances. Its 96 GB workstation tier did not move at all [3][4].
  • $4,999 for 64 GB — The desk box inversion: Nvidia’s new half-size Spark machine ships on 23 October “starting at $4,999”. The 128 GB model that “launched at $3999 about a year ago” now lists at $6,950 [5].
  • $3.47 — The H100 hour: that was the median on-demand rental across 41 providers in the week of 5 October, up 4.0% in four weeks and 16% in a year, even as the premium the big cloud firms charge narrowed to 75% [6].

What I’d watch:

One line in the rental data matters more to me than the price rises. The premium the big three clouds charge over the discount providers narrowed to 75% because the discount side moved up [6]. The floor rose. The ceiling did not.

Three things I am watching from here.

  • Where the 84 GB card gets priced. A card with the 5090’s core count and 2.6 times its memory, priced near the old consumer tag, would undercut the professional line beneath it. Priced at professional money, it says the premium is on purpose [1][2].
  • Whether the second rise holds. Two hikes in three months test pricing power, not costs. If usage holds up after them, a memory index becomes how this industry prices everything [3][4].
  • The 32 GB resale market. Those cards listed at $1,999 and were already fetching three times that before the leak. What used ones bring in November tells you what buyers think the replacement is worth [1][7].

The catch

My read is that this is a memory story wearing a compute story’s clothes. Three limits still apply.

First, the halt is a leak, not a statement. Nvidia has not confirmed it, has disputed it in private to at least one outlet, and cards already built can keep selling for a while [1]. The half I can stand behind is the pricing above it.

Second, renting has not closed. A one-year reservation still runs about 32% below on-demand, and a three-year about 49%. Spot capacity sits near 59% of the on-demand rate [6]. A team that will tie up capital, or put up with jobs dying mid-run, can still move its number down.

Third, the rises are not universal. Of 22 rental models tracked this year, 14 went up and 8 fell, and the 96 GB workstation tier did not move [4][6]. What repriced is memory-heavy gear, which is the same thing Nvidia is re-cutting its silicon to carry more of. TrendForce’s numbers from August, before this window, already had memory at 68% of major cloud providers’ capital spending for 2027, from 47% this year [8].

At a glance

  • The Big Shift: Nvidia reportedly stopped making its only 32 GB consumer card, to reserve that silicon for professional boards. It also listed an 84 GB professional card with the same core count and no price. In the same stretch, Nebius raised rental rates twice in three months, hardest on memory, at 41%.
  • Why It Matters: Both halves of the artificial intelligence (AI) compute decision, buy or rent, now move with memory capacity rather than with compute. A 2027 plan built on price per unit of compute, or on today’s hourly rate, will come in low.
  • What I’d Watch: three signals that will show whether memory has become this industry’s price index.
  • The 84 GB card’s price: a price near the old consumer tag undercuts the line beneath it; professional money says the premium is deliberate.
  • Whether the second rise holds: usage after two hikes in three months is the test of pricing power.
  • The 32 GB resale market: a card that just stopped being restocked tells you what buyers think its replacement is worth.
  • The Catch: the halt is an unconfirmed leak that Nvidia has disputed, and reservations and spot capacity still price below on-demand. The rent side has not closed. It has just stopped being the cheap answer by default.

Related reading

Sources

[1] Tom’s Hardware, “Nvidia reportedly halts GeForce RTX 5090 production in favor of AI data center and professional GPUs” (2026-10-10) — https://www.tomshardware.com/pc-components/gpus/nvidia-reportedly-halts-geforce-rtx-5090-production-in-favor-of-ai-data-center-and-professional-gpus-impending-supply-drought-expected-to-drive-up-prices-rtx-5080-24gb-rumored-as-new-gaming-flagship [2] NVIDIA, 5500 Blackwell Workstation Edition product page (fetched 2026-10-13) — https://www.nvidia.com/en-us/products/workstations/professional-desktop-gpus/rtx-pro-5500/ [3] Reuters, “Nebius hikes AI cloud prices again as demand for computing power soars” (2026-09-17) — https://www.reuters.com/technology/nebius-hikes-ai-cloud-prices-again-demand-computing-power-soars-2026-09-17 [4] Nebius, AI Cloud pricing, GPU-hour rates effective 1 October 2026 (fetched 2026-10-13) — https://nebius.com/prices [5] ServeTheHome, “NVIDIA DGX Spark 64GB Launched and Big 128GB GB10 Price Increases” (2026-10-02) — https://www.servethehome.com/nvidia-dgx-spark-64gb-launched-and-big-128gb-gb10-price-increases and the NVIDIA store listing at $6,950 — https://marketplace.nvidia.com/en-us/enterprise/personal-ai-supercomputers/dgx-spark [6] getdeploying.com, GPU Price Trends, week of 2026-10-05 (fetched 2026-10-13) — https://getdeploying.com/gpu-price-trends [7] Tom’s Hardware, “Nvidia’s RTX 5090 vanishes from online retail in the US” (2026-09-14) — https://www.tomshardware.com/pc-components/gpus/nvidias-rtx-5090-vanishes-from-online-retail-in-the-us-third-party-sellers-now-demand-as-much-as-usd9-500-for-nvidias-fastest-gpu [8] TrendForce, DRAM and NAND to account for 68% of major cloud providers’ capex in 2027 (2026-08-25, dated before this window) — https://www.trendforce.com/presscenter/news/20260825-13198.html