Why Graphics Cards Are So Expensive in 2026: How AI Ate the Memory Supply

A graphics card with a recommended price of $2,000 is selling for $5,000. This time, miners are not to blame, and scalpers are only part of the problem: the AI boom has prompted memory manufacturers to redirect production capacity toward more profitable data center orders. As a result, both system RAM and video memory are becoming more expensive, driving up the prices of graphics cards and other components. Here is why building a gaming PC has once again become an expensive proposition in 2026 — and when the situation might change.

The GeForce RTX 5090 provides the clearest illustration of the scale of the price increases. NVIDIA's flagship launched in January 2025 with an MSRP of $1999 — an enormous figure even by high-end graphics card standards. By summer 2026, however, US retailers were asking at least $4,000 for the model. In early August, listings on the South Korean price comparison platform Danawa reached 7.3 million won, or approximately $5112. NVIDIA itself has not changed the recommended price, so the RTX 5090 still officially costs $1999.

These prices can no longer be dismissed as isolated examples from unusually expensive sellers. At the high end of the market, graphics cards are consistently selling for several times their MSRP. Resellers and automated bot purchases may be making the shortage even worse, but this time the problem begins at the manufacturing stage: key components are becoming more expensive, while memory producers are prioritizing more profitable orders from the AI industry.

Previous periods of dramatic graphics card price increases were largely connected to cryptocurrency mining. In 2017–2018 and 2020–2021, crypto miners bought gaming GPUs in huge quantities, while resellers and automated purchasing systems placed additional pressure on the market. After the cryptocurrency downturn and Ethereum's transition to the Proof-of-Stake model in 2022, demand from miners dropped sharply, and the situation gradually began to normalize.

The current crisis works differently. The problem is no longer limited to retail and instead begins much earlier in the production chain: memory chips themselves are becoming more expensive, while manufacturers are prioritizing HBM for AI accelerators and data centers. Consequently, even a decline in consumer demand would not quickly bring prices back to their previous levels. To understand the scale of the problem, it is first necessary to examine how much of a modern graphics card's cost comes from the GPU and memory.

According to industry estimates, the GPU and video memory account for around 80% of the production cost of a modern graphics card. The remaining 20% covers the printed circuit board, power delivery system, cooling solution, and other components. Until recently, the GPU was the most expensive part, but after the sharp increase in prices, memory costs have almost caught up.

The RTX 5090 comes with 32GB of video memory, spread across 16 GDDR7 chips. Before the current price surge, this configuration was estimated to cost around $80–100. It now costs as much as $320. For comparison, the RTX 5090 GPU die costs more than $300. Memory has therefore become several times more expensive and is now comparable in price to the graphics processor itself.

Higher production costs alone, however, are not enough to explain retail prices of $4,000–5,000. Reduced supply and larger markups during the shortage also play a role. More expensive components raise the baseline cost of a graphics card, while limited availability widens the gap between its recommended and actual retail prices even further.

This brings us to the central question: why has memory suddenly become so scarce? The AI boom is responsible.

Neural network accelerators require enormous amounts of memory. A single NVIDIA H200 accelerator carries 141GB of ultra-fast HBM3e memory, while newer Blackwell hardware has even more: 180GB in the B200 and 372GB in the GB200 combination. By comparison, the gaming-focused RTX 5090 has 32GB, while the mainstream RTX 5060 has just 8GB. Put simply, a single AI server contains more premium memory than an entire rack of gaming PCs.

There are effectively only 3 major memory manufacturers in the world: Samsung, SK Hynix, and Micron. All of them have redirected capacity toward HBM — expensive, ultra-fast memory for AI servers — because the margins are higher. Producing HBM requires considerably more resources than manufacturing conventional memory. Its chips are stacked on top of one another, while every gigabyte consumes several times more silicon wafer area. As a result, even a partial reallocation of production capacity toward HBM significantly reduces the volume of memory available for conventional PCs. The numbers make the situation clear: IDC analysts expect shipments of conventional system memory, or DRAM, to grow by only 16% in 2026, while NAND flash shipments will rise by 17%. Historically, growth of 20–30% has been the norm. Meanwhile, data center demand reportedly consumes as much as 70% of all high-performance memory, according to some estimates, with AI serving as the main — though not the only — driver. The accuracy of that estimate is debatable, but the overall direction is clear: less and less memory is reaching the consumer PC market.

SK Hynix and Micron sold out their entire HBM supply for 2026 before the year had even begun. In late 2025, Micron also announced that it would discontinue its consumer brand Crucial — which sells retail memory modules and SSDs — by February 2026 and redirect capacity toward enterprise AI customers. NVIDIA CEO Jensen Huang made the company's position explicit: no matter how many new memory factories are built, Nvidia will buy all of their production capacity.

With memory both expensive and scarce, graphics card manufacturers have responded in 2 ways: by raising prices and limiting GPU production.

This is how the situation looks for NVIDIA. According to supply chain rumors, the company is cutting production of RTX 50 gaming graphics cards in the first half of 2026. Estimates range from 20% to 40%, with the most memory-intensive models, such as the RTX 5070 Ti and RTX 5060 Ti, reportedly facing the largest reductions. The refreshed RTX 50 Super lineup, previously expected in early 2026, has been postponed indefinitely. The 3GB GDDR7 chips it requires cost 3 times as much as conventional 2GB chips, making the economics unworkable. The next generation, RTX 60, is now projected to arrive in 2028. This could create an unprecedented situation: according to TrendForce analysts, 2026 may become the first year in roughly 30 years in which NVIDIA releases no new gaming graphics card models. The company's financial logic is easy to understand. Gaming accounted for around 11% of NVIDIA's revenue in its latest fiscal year, and that share continues to decline. Every wafer used for an AI chip instead of a gaming GPU generates several times more money. Gamers are no longer NVIDIA's primary customers.

AMD is following the same path. Over the course of 6 months, the company twice raised the wholesale prices charged to partners for its GPU-and-memory packages — by approximately 10% each time — explicitly citing the AI-driven memory shortage. Board manufacturers have followed suit: ASUS officially raised prices on some of its products starting January 5, 2026, before CES had even begun.

The worst part is that graphics cards are not the only products affected. The shortage has hit virtually every type of hardware that uses memory.

Standard PC RAM has risen even faster than graphics cards — we previously examined the reasons behind the increase in a separate article. According to TrendForce, contract prices — the rates system builders pay when purchasing memory in bulk — for conventional DRAM surged by 90–95% in a single quarter at the beginning of 2026, setting a new record. Rising memory and flash prices are also making SSDs, gaming consoles, and laptops more expensive. Handheld PCs have not escaped the trend either: the flagship MSI Claw 8 EX AI+ launched at $1799, while the anniversary ROG Xbox Ally X20 bundle, which includes the handheld and a pair of AR glasses, is expected to cost around $2,000. The situation has reached the point of absurdity. Japanese stores have begun rationing sales, limiting customers to one graphics card with 16GB of memory or more.

According to Steam's mid-2026 figures, only around 0.45% of players use an RTX 5090. The RTX 3060, RTX 4060, RTX 3050, and even the considerably outdated GTX 1650 remain among the most common graphics cards. While the industry discusses $5,000 GPUs, the average gamer is still using a graphics card that is 5 or more years old — simply because they cannot afford a new one.

There is only one practical conclusion: in 2026, MSRP is largely meaningless. Buyers should focus on actual store prices and the performance they receive for their money, measured in cost per frame.

A few guidelines:

Specific models and current prices are covered in a separate guide to choosing a graphics card, which is regularly updated to reflect the retail market. Those planning a more substantial PC upgrade or considering a console should also read our column explaining why this may be the least bad time to buy gaming hardware: prices have already risen, but the situation could become even worse over the next few years.

A rapid improvement like the one that followed Ethereum's departure from mining in 2022 is unlikely. Back then, demand disappeared almost overnight. Demand from AI data centers, by contrast, is not going anywhere in the next few years.

The manufacturers' own forecasts are bleak. The head of SK Hynix expects the memory shortage to be worse in 2027 than in 2026, with demand continuing to exceed production capacity even after 2030. Samsung is also preparing for a more difficult 2027 and does not plan to expand production significantly until 2028. NAND flash memory may recover sooner, while conventional DRAM will take longer. Analysts are reluctant to make firm predictions about what will happen after 2028.

That said, DLSS and FSR can improve performance and allow gamers to postpone a graphics card upgrade for several more years. The next generation will probably come with more memory, but until it is officially announced, that remains only an assumption.

Prices are unlikely to return to their previous levels anytime soon. However, paying several thousand dollars extra for a flagship makes little sense either. More affordable models and discounted previous-generation graphics cards are the more reasonable options.

Because of the memory shortage. Memory manufacturers have redirected production capacity toward expensive HBM chips for AI data centers, causing video memory prices to surge while supplies have become increasingly limited. The GPU die and memory together account for as much as 80% of a modern graphics card's production cost, with memory being one of the 2 largest expenses, so graphics cards themselves are becoming more expensive as well.

Neural network accelerators consume enormous amounts of high-speed memory: a single AI accelerator carries 141–180GB, while a GB200 superchip combination has as much as 372GB, compared with 8–32GB in a gaming graphics card. Memory manufacturers have redirected factories toward producing this data center memory, leaving insufficient capacity for conventional DRAM and video memory.

The MSRP is $1999, but retailers are selling the card for $4,000–5,000 or more in 2026. In August, prices of approximately $5,112 were recorded in South Korea. NVIDIA has not officially raised the price — the difference comes from markups caused by the memory shortage. Used cards are cheaper, typically costing around $3,700–4,000.

According to a forecast from TrendForce analysts, yes. 2026 could become the first year in roughly 30 years without a single new gaming graphics card from NVIDIA. The RTX 50 Super refresh has been postponed indefinitely, while the next-generation RTX 60 series is expected around 2028.

Not anytime soon. Memory manufacturers warn that 2027 could be more difficult than 2026, while significant production expansion is not planned until 2028. A rapid price collapse like the one that followed the decline of crypto mining in 2022 is unlikely.

Focus on the actual cost per frame rather than MSRP. In the budget segment, consider the Intel Arc B580 or the 16GB AMD RX 9060 XT. In the mid-range segment, look at the RX 9070 XT or RTX 5070 Ti. Avoid cards with only 8GB of memory and consider discounted previous-generation models.

Anything that uses memory is becoming more expensive. Contract prices for PC RAM rose by 90–95% in a single quarter at the beginning of 2026, while SSDs, consoles, laptops, and handheld gaming systems are also increasing in price. The phenomenon has already been dubbed the “RAMpocalypse.”