Why GPUs Are So Expensive in 2026, and Why Memory Matters Most
Affordable graphics cards are hard to find in 2026. AI demand has pushed new prices sharply higher, in what buyers call the RAMaggedon.
NVIDIA RTX 50-series prices have risen by as much as 39 percent, while AMD GPU prices are up by as much as 10 percent, according to September 2026 reporting Engadget. The rise sits on top of a wider memory crisis. Memory chip costs rose 50 percent in 2025, then surged 80 to 90 percent in recent months Facebook Reuters. In January 2026, Counterpoint Research estimated another 40 to 50 percent rise in the first quarter alone.
The driver is factory capacity, not one product cycle. Data centers took about 50 percent of global DRAM, the main working memory in PCs and servers, in 2025, up from 32 percent five years earlier Bloomberg. GPU demand now comes from two sides at once: large companies building large language models, the AI systems behind chatbots, and people building local AI setups at home. Those buyers want specific memory sizes. Typical AI buyers seek 12 to 16GB of VRAM, the dedicated memory on a graphics card, while larger teams seek up to 48GB.
That need keeps older high-memory cards in demand. The NVIDIA GeForce RTX 3090 holds more VRAM than many newer models, so it stays expensive on the secondhand market despite age and power draw. Technicians now filter by VRAM first. Core clocks and ray tracing throughput matter less when the task is loading weights, the files that form an AI model, into local memory.
Desktop graphics card shipments reached 12.5 million, a four-year high, despite higher prices Tom's Hardware. Nvidia held 90 percent of desktop boards as gamers bought ahead of expected spikes. 2026 was on track to pass 2025 in units despite rising prices. PC Partner warned earlier in 2026 of rising GPU prices and budget card shortages, and one analyst said makers raised prices by more than memory costs Tom's Hardware.
A used GPU with ample VRAM in good condition can retain at least 40 to 60 percent of its value in 2026. Recent eBay sold listings show used NVIDIA GeForce RTX 5060 Ti and EVGA GeForce RTX 3080 cards selling for 40 to 60 percent of original price, sometimes above original MSRP. Newegg runs a GPU trade-in program, an alternative to peer-to-peer listings with their testing work and fees.
The AI boom first squeezed GPUs in late 2022, then spread to memory chips by late 2025, according to September 2026 reporting Tom's Hardware. Acer's CEO said memory makers hyped fears of shortages through 2030 to protect margins, and predicted PC prices would fall by late 2027 as cheaper Chinese capacity came online.
Vendors are also testing other memory designs. Microsoft's Surface Laptop Ultra is a 15-inch machine using NVIDIA's RTX Spark system-on-a-chip, with RTX Spark systems described as starting at around $4,000 or more. Apple's M6 chip has a 12-core GPU with neural accelerators in every core and supports up to 64GB of RAM with 307 GB/s of bandwidth, while a device identified as Neo has 8GB of RAM shared between the A18 Pro's CPU and GPU. None match a discrete 16GB or 24GB card for local model work today.
The broader context here is buyers treating this as a temporary squeeze. Sales stayed high because many bought early to beat further rises, and whether Acer's late-2027 date for lower prices holds is disputed. For buyers and IT shops, that points to inventory discipline. Audit VRAM first and performance second. If a workstation already has 16GB or more in a healthy card, delaying a refresh is reasonable. If an upgrade cannot wait, the old card can partly pay for the new one when it has the memory AI buyers want, so list it promptly and check sold listings rather than asking prices.
In my view, this pattern is familiar. I watched my own children move from a shared family PC to phones to cloud-connected laptops, and each shift brought warnings that supply would never catch up. It did. Memory remains a cyclical commodity business, even with AI changing demand, and added capacity plus softer buying tend to bring prices back into balance over time. Sellers of memory-rich hardware have the edge now. Over the longer arc, access to capable local machines should widen.
