- The cheapest available US RTX 5090 now costs over $6000 against a $1,999 launch price, with third-party listings often running past $9,000
- The GPU that already pushed past $3,500 in Feb 2026 owing to memory constraints driving prices higher has now hit fresh highs at multiple retailers
- The move is linked to AI demand, as data center operators have entered the fray to secure compute despite Nvidia’s GeForce driver license having banned ‘data center deployment’ on paper since 2017
Six weeks ago, the story was ultra-rare books being fed spine-first into high-speed scanners so that AI labs could train on text written before chatbots polluted the corpus.
Focus, however, has shifted considerably to AI labs bringing the same voracious appetite to Nvidia’s flagship consumer-grade Blackwell GPU, the RTX 5090, which has essentially disappeared from shelves unless you want to pay a 200 percent or higher premium to grab one.
Photographs circulated by Hong Kong outlet HKEPC, and picked up by multiple publications including Guru3D, VideoCardz, and TechPowerUp, appear to show wooden pallets stacked with boxed, retail GeForce RTX 5090 cards at AI server assembly sites, waiting to be unpacked and slotted into multi-GPU rigs.
The sky is the limit for Nvidia’s flagship consumer card?
This isn’t the first time Nvidia’s RTX 5090 GPUs have been deployed en masse on multi-GPU, AI-centric rigs, but this time feels different, even compared withprevious shortages.
For context, Nvidia launched the RTX 5090 at $1,999 in January 2025. Tom’s Hardware’s price tracker, updated on September 14 2026, lists the cheapest available US card at $6,389, nearly 3.2x the former.
PCGamesN reported the same week that first-party stock at Amazon, Newegg, and Best Buy had essentially vanished, leaving third-party marketplace listings above $6,000 and, in some places, over $9,000.
The pictures in circulation don’t show RTX PRO workstation GPUs, but rather GPUs from MSI, Gigabyte, and ASUS that the average gaming enthusiast or consumer would pick off a shelf.
An AI-generated image to create more FOMO?
The images have, however, been disputed by some quarters: Kevin Hofer at Swiss retailer Digitec, one of the few writers to examine the images rather than republish them, pointed at the box lettering in one frame and asked what a “50B0” or a “50R0” is supposed to be, concluding that AI was probably involved in producing the picture. Twitter has also added a “made with AI” moniker to the image in question.
TechPowerUp and TweakTown both carried the same warning in softer form. Guru3D, to its credit, noted that the report explains neither why these firms chose the RTX 5090 nor the scale of the purchases.
Coverage like this, even with disclaimers to the contrary, can create a “fear of missing out” (FOMO) where price increases become a self-fulfilling prophecy as businesses and consumers alike move purchases forward to avoid paying a further premium.
Examining the value that an RTX 5090 brings to the AI table
The Nvidia RTX 5090 is its most powerful consumer-grade card to date, packing a Blackwell-based GB202 chip and 32GB of GDDR7 memory, which checks multiple boxes for local AI use cases and gamers looking to max out games at 4K or higher.
Nvidia’s GeForce driver EULA has, since late 2017, included a clause titled “No Datacenter Deployment” that should, at least on paper, prevent data centers from stockpiling or using these.
The chip designer, currently the world’s most valuable company, has previously implemented technological restrictions such as LHR (Lite Hash Rate) on its RTX 3000 series GPUs to prevent GPU miners during the Ethereum gold-mine rush from hoarding supply.
It has also released a China-specific RTX 5090 D, which reduces AI performance while matching the RTX 5090 in gaming throughput. The RTX 5090 D, created to meet AI export restrictions, shows that Nvidia can limit the RTX 5090’s AI performance, but it remains relatively indifferent to a market that was its core revenue
The RTX 5090’s price hike, at least in part, may be due to the RTX PRO 6000’s staggering, out-of-control pricing, which Nvidia nearly doubled to $16000. The underlying cost driving the price hike is the GPU’s 96GB of VRAM, which also makes up a sizeable, if not dominant, chunk of its total bill of materials.
An RTX 5090 comes with 1/3rd the VRAM on offer, but nearly the same compute performance, and with AI-centric algorithms often assisting in breaking up tasks between GPUs and memory pools, even as the total memory pool remains similar, a triple stack of RTX 5090 GPUs should comfortably outperform the RTX PRO 6000 in certain tasks.
At the same time, Chinese OEMs have reportedly managed toslap on as much as 96GB of VRAM on an RTX 5090, which could make it an even hotter commodity in a country where both the RTX PRO 6000 and the RTX 5090 are export-restricted but do manage to make it on the regular.
The RTX 5090 is a manufacturing problem with a memory bottleneck attached for AI consumers, but with the alternatives having doubled in cost recently, AI consumers are looking to save money wherever possible, and both the stock RTX 5090 and a potentially modified one are some of the best candidates in a market that has already exhausted most of its last-generation supply of the same.
