The lack of GPUs never truly went away. Its emphasis just changed. The antagonist was clear in 2021: bitcoin miners pursuing digital riches while storing graphics hardware in fluorescent-lit warehouses. That tale seems almost archaic by 2025. The new strain comes from expansive AI data centers that are working around the clock to train models that the majority of consumers will never see, not from underground mining rigs.
On paper, manufacturers such as AMD and NVIDIA have expanded manufacturing. However, the glass display cabinets felt oddly empty when I recently walked into a computer hardware store in Manhattan. Even experienced players were taken aback by the cost of a few mid-range cards that were sitting under bright lights. The expensive models? The clerk shrugged, “Out of stock,” with the well-honed indifference of someone who has said it ten times that morning. It’s quieter now, the shortage. less disorganized. more organized.
| Category | Details |
|---|---|
| Key Manufacturer | NVIDIA |
| Competing Manufacturer | AMD |
| Flagship AI GPU | NVIDIA H100 |
| Consumer Models Mentioned | RTX 5060, RTX 5070 Ti |
| Market Condition | 30%–70% markups above MSRP |
| Official Reference | https://www.nvidia.com |
Due to their greater margins, data center GPUs like NVIDIA’s H100 are gaining popularity. It’s hard to argue with the reasoning. Why would silicon wafers be diverted to a $399 gaming card while hyperscalers are willing to pay tens of thousands per unit? Nevertheless, that decision is subtly changing the consumer market by reducing supply while claiming in public that production is robust.
Another layer is added by memory limitations. High-bandwidth memory, or HBM, is still scarce, which causes bottlenecks for both high-end gaming GPUs with 16GB or more VRAM and AI accelerators. Whether these limitations are entirely technical or somewhat deliberate is still up for debate, but the result is the same: selected shortages that rarely garner media attention yet persist for months. The shadow market thrives in that space.
When a new RTX 5060 was introduced in early 2025, automated bots apparently snatched up over 70% of the stock within hours. It was almost dramatic to watch the listings evaporate in real time, as if the tickets had vanished before the event had even started. The identical $399 card was selling for $675 or more by the end of the week on resale sites, representing a markup of about 70%.
It’s not always clear that the brokers in this market are scalpers. Some act as “fences,” rerouting supply from areas with lower demand into markets with higher prices by transferring bulk inventory across borders. Others work as “fixers,” surreptitiously locating premium, customized cards for customers who are prepared to pay more. Far from official stores, deals are frequently conducted via private emails or encrypted conversations.
This gray market seems to have reached a certain level of maturity. Compared to the epidemic era, it is more ordered and less chaotic. worldwide. Weeks later, cards that were delivered to Southeast Asia can reappear in Europe with new labels and prices. Even after fees and shipping, the margins are still appealing.
Retailers, meanwhile, are picking up tips from the strategy. Checkout pages often contain hidden costs. Bundles are becoming more popular; a GPU connected to an undesired motherboard or power supply raises the effective price without technically breaking MSRP regulations. Even when stock is accessible, some third-party merchants on popular platforms post their products at 15% to 30% more than suggested costs, counting on customer tiredness to seal the deal.
Customers seem to have adjusted their perception of what is “normal” after years of price inflation. A mid-range card that costs $600 no longer shocks as much as it used to. One of the most effective strategies used by the shadow market can be that normalization trap. There is less motivation for intermediaries to compete at lower prices if customers accept higher prices as the norm.
It’s difficult to ignore how this seems different than shortages in the past. Not all of the shelves are empty. Supply crises are not announced in press announcements. Rather, there is a constant tension as availability fluctuates. There is not enough stock to bring back true price competition, but there is enough to prevent panic headlines.
In 2025, Japan provided a brief counterexample, claiming that customer opposition caused prices in some segments to drop by about 20%. However, high-performance cards—especially ones with 16GB or more VRAM—remain hard to come by worldwide. Gaming just cannot keep up with the rate at which AI demand is consuming supply.
As this plays out, there is an unsettling comparison to home markets in hot cities: purchasers extending their budgets, speculative buying, and restricted inventory because waiting seems riskier than overpaying. According to investors, AI’s demand for computing power will only increase, maintaining upstream pressure on output well into 2026.
It’s unclear if that assumption turns out to be accurate. There are plans to increase semiconductor capacity, with new factories opening in Taiwan and Arizona. However, it takes years for fabrication operations to ramp up, and the demand for AI is currently rising. In that lag, the shadow market flourishes.
The impact is noticeable for both small AI businesses attempting to prototype models without enterprise funds and gamers constructing setups in spare bedrooms. Spreadsheets for budgets are long. Timelines are prone to slipping. Calculation takes the place of enthusiasm, which somewhat wanes.
Unlike before the pandemic, GPUs are now commonplace. They are accessible, although frequently not at reasonable costs. Perhaps the most noticeable aspect of this new shortage is the stillness surrounding it. No spectacular crashes. No anger on the internet. Only steadfast markups, brokers working quietly in the background, and consumers grudgingly adjusting. As it happens, the shortage persisted. It changed.
