Graphics card prices have quietly doubled since last year, and this isn't a post-launch scalper spike that settles down after a few months. A structural memory shortage has already reshaped NVIDIA's and AMD's entire roadmaps, and it won't fix itself before 2028.
The bottleneck is memory, not silicon
TSMC fabs GPU dies without trouble. NVIDIA and AMD have plenty of silicon. The shortage sits in DRAM: that memory is getting pulled away from consumer graphics cards and funneled into AI data centers instead.
Samsung, SK Hynix, and Micron make nearly all the world's DRAM, and they've been reallocating production capacity toward High Bandwidth Memory (HBM), the stacked memory used in AI accelerators like NVIDIA's own data center GPUs. HBM carries far higher margins than the GDDR6/GDDR7 memory used in gaming cards, so when a fab picks which product gets priority, gaming loses. IDC now estimates AI data centers will absorb roughly 70% of global memory production in 2026, up from just 20-30% in 2022. The memory industry's output has permanently shifted toward AI infrastructure.
Modern graphics cards are memory-hungry by design, so the knock-on effect hits hard. On some high-end SKUs, VRAM now accounts for more than 80% of the card's total bill of materials. A stack of 16GB of GDDR7 that cost board partners roughly $65-80 in mid-2025 now runs north of $200. NVIDIA and AMD pass that increase straight to buyers, and in some cases still eat margin on top of it.
What cards actually cost right now
Street pricing as of mid-September 2026, measured against launch MSRP, shows the scale of the shift.
| Card | Launch MSRP (late 2025) | Street price (Sept 2026) | Change |
|---|---|---|---|
| RTX 5090 | $1,999 | $4,300-$5,000+ at official retail (when in stock); $6,000-$9,500+ on third-party marketplace listings | +115% to +375% |
| RTX 5070 Ti | $749 | ~$1,149 | +53% |
| RTX 5060 | ~$300 | Discontinued, pulled from the market entirely | N/A |
| 32GB DDR5-6000 kit (system RAM, not VRAM) | $80-120 | $400-470 | +230% to +400% |
The RTX 5090 is also hard to find at official retail. Amazon, Newegg, and Best Buy sit largely out of first-party stock in the US, and Micro Center remains one of the few chains still selling it in-store, around $4,299 when it shows up. NVIDIA has reportedly cut production across the entire RTX 50 lineup by 15-20% through at least Q3 2026, and pulled the RTX 5060 outright rather than sell it at a loss on thin margins.
The shortage runs DRAM-wide, not GDDR-specific, so plain system RAM has been dragged into the same spiral. A 32GB DDR5-6000 kit that cost under $120 a year ago now runs $400-470. Anyone pricing out a full build in 2026 pays the memory tax twice: once on the GPU, once on the motherboard's RAM slots.
Why there's no quick fix coming
NVIDIA and AMD have both reportedly pushed their next-generation architectures back a full year. NVIDIA's Rubin-based RTX 60 series and AMD's RDNA 5 lineup were expected in 2027; both are now tracking for 2028 instead. Industry reporting agrees this isn't a GPU design or TSMC capacity problem. Neither company can ship a next-gen card without enough GDDR to put on the board, and that memory pipeline is booked out by AI infrastructure for the foreseeable future.
Memory suppliers aren't promising relief soon, either. Samsung has flagged that 2027 could run tighter than 2026 before new fabrication capacity comes online, with a real correction more realistically landing in 2028. Anyone holding out for prices to normalize by next spring is banking on a timeline the supply chain itself isn't describing.
Buy now, or wait
If your current GPU is failing, bottlenecking you badly, or you have a real need this year (new build, VRAM-starved card choking on modern titles), buy now. Waiting for a price drop that the entire memory industry places past 2028 is a losing bet, and prices could still climb before they stabilize. Buying today at an inflated price beats buying in six months at an even more inflated one.
If your current card still does the job, an RTX 4000-series or RX 7000-series card especially, skip the panic-buy. You're sitting out a spike, not missing a window. Watch Micro Center's in-store stock and official retailer restocks rather than third-party marketplace listings, skip the RTX 5090 unless you have a real 8K or AI workload to justify it, and treat any GPU purchase in this window as buying what you need, not what you want. The market won't reward patience the way it usually does, but it will punish impulse upgrades at these prices.
