Memory went up, SSDs went up, hard drives went up, and graphics cards went up. The 2026 build market has produced something odd: a year after a new generation launched, the cards did not fall in price, they rose across the board. That has almost never happened in the past decade. The normal pattern is that prices loosen six to twelve months after launch, as the next generation approaches.
What the increase looks like on four models
| Model | Launch MSRP | 2026 median street price | Change |
|---|---|---|---|
| RTX 5060 Ti 16GB | US$569 | US$804.99 | +41% |
| RTX 5070 | US$659 | US$899 | +36% |
| RTX 5090 | US$1,999 MSRP | US$4,699-4,900 | 135-145% above MSRP |
| AMD RX 9070 | No MSRP published | +13% over launch price | Relatively stable |
Those are medians of overseas e-commerce transactions, on the Newegg basis. Domestic channel prices moved the same way. The RTX 5070 went from RMB 4,200-4,500 at the start of the year to RMB 5,500-6,000, the RTX 5060 Ti 16GB from RMB 3,300 to about RMB 4,200, and the 5090 has sat above RMB 20,000 with very little stock to buy at any price.
Why: memory cost ate the graphics card
The key figure is counterintuitive: GDDR7 now accounts for more than 80% of the manufacturing cost of a high-end graphics card. A decade ago the GPU core and its wafer cost dominated, alongside PCB, power delivery and cooling, and memory was 15% to 25% of the bill. Two things changed in this generation.
- Memory capacity doubled. Most of the RTX 50 series starts at 16GB, and the 5080 and 5090 carry 16GB to 32GB of GDDR7 on 256-bit to 512-bit buses. That is twice as many memory packages as the previous generation at the same tier.
- GDDR7 unit prices soared. It is a new product with a limited capacity ramp, and its lines compete for wafers with GDDR6 and DDR5. When DRAM rises across the board, GDDR7 rises harder, because a new part has no inventory buffer to absorb the shock.
Multiply the two and the arithmetic gets ugly. The 32GB of GDDR7 on an RTX 5090 may now cost as much as the GPU core itself. If memory goes up 50%, the card has to go up more than 30%, and the manufacturer has no other lever.
The second push is AI demand buying consumer flagships. Professional cards such as the H200 and B200 are expensive and in short supply, so small AI teams, university labs and individual developers have turned to the 5090, whose 32GB is well suited to local model inference. That demand, not gaming demand, is what lifted the street price of the high-end card.
Why AMD has held steadier
The RX 9070 series is up only about 13%, well below NVIDIA. Three reasons:
- More conservative memory. The RX 9070 uses 16GB of GDDR6 rather than GDDR7. GDDR6 is a mature part, and while it rose too, it rose less than the new memory.
- No AI premium. AMD cards are at a disadvantage in the CUDA ecosystem, so AI developers mostly do not buy them. The RX 9070's price is set by gaming supply and demand alone, with no extra demand pushing it up.
- Deliberate pricing. AMD chose to take share on price-performance this generation and to compress its margin rather than track rising costs upward.
The signal for builders is clear. For pure gaming in 2026, AMD offers noticeably better value than NVIDIA. If the workflow depends on CUDA — whether it is AI training and inference, Blender GPU rendering,, some Adobe acceleration or professional simulation, NVIDIA's premium is still a necessary cost.
Buying advice for three kinds of user
Gamers
At 1080p with high frame rates, take an RTX 5060 8GB or an RX 9060 XT 16GB. The AMD card's larger memory suits 2026 titles, many of which now exceed 8GB at 1080p maximum settings. At 2K, the RX 9070 at about RMB 4,500-5,000 is the value pick, close to the previous generation's 4080 for roughly 30% less. At 4K, choose an RTX 5080 at about RMB 8,500-9,500 or an RX 9070 XT, with DLSS 4 multi-frame generation as NVIDIA's argument. Do not buy a used older card to save money right now: RTX 30 series prices rose with everything else, and a used 3080 at RMB 3,000 or more is worse value than adding the difference for an RX 9070 with a warranty.
Design, video and rendering studios
Memory matters more than core performance. 4K editing, large-scene rendering and AI image generation all eat VRAM, so 16GB is the floor and 24GB is comfortable. CUDA dependency decides the brand: Blender, OctaneRender and the Adobe suite mean NVIDIA, while DaVinci Resolve, which is well optimized for AMD, gives you a real choice. And one large card beats two small ones. A single 24GB card is more useful in most studios than two cards with 12GB each, because much of the software does not use the second GPU.
AI development and local inference
For a 7B to 14B quantized model, 16GB is enough: an RTX 5060 Ti 16GB or an RX 9070 at RMB 4,200-5,000. For 32B models or LoRA fine-tuning, 24GB is the practical floor, which means a used RTX 4090 or a 5090. If this is real business rather than a hobby, rent cloud capacity or a card instead: a 5090 now costs RMB 20,000 or more, and the same money buys hundreds of hours of A100 or H20 time. A local card is worth owning for data that cannot leave the building and for latency, not for cost per token. Enterprises should also look at domestic accelerators from Ascend, Cambricon and Moore Threads, whose inference value is now competitive, with ecosystem maturity the main reservation.
How to split a 2026 build budget
With storage and graphics rising at the same time, the old allocation rules need rewriting:
| Use | Typical 2024 split | Suggested 2026 split |
|---|---|---|
| Office PC, RMB 3,000-4,000 | GPU 0, integrated; memory 20%; SSD 15% | Integrated graphics; 16GB is enough, since memory is expensive and 32GB is not worth it; 512GB SSD |
| Gaming PC, RMB 7,000-9,000 | GPU 40%; memory 10% | GPU 45-50%; 32GB memory, down to 8% of the budget but up in absolute terms |
| Design workstation, RMB 15,000-20,000 | GPU 35%; memory 15% | GPU 35%; memory 12%, since 64GB now costs RMB 7,000; put the savings into the monitor and storage |
In one line: do not chase the top spec on memory or graphics in 2026. Put the budget into the parts that change daily experience: monitor, SSD capacity, cooling and PSU quality. and power supply quality. Those items have risen far less than memory and cards, so the value is better.
ACCPC wholesales graphics cards in Guangzhou across the full NVIDIA and AMD ranges, including ASUS, MSI, Gigabyte, Galax and Colorful, and builds custom machines. We ask about your actual workflow before quoting instead of copying a configuration off the internet. Bulk orders of five units or more get channel pricing, with a three-year on-site warranty. Call 020-39029800.