GPU prices just jumped up to 39% in two months — here's what's actually worth buying for local AI now

If you were planning to buy a GPU for local AI this month, the advice from two months ago is now wrong. Newegg's median RTX 5060 Ti 16GB price jumped from $569.99 in June to $804.99 in August — a 39% hike — while the RTX 5080 barely moved, up just 3% to around $1,500. That's not a uniform price increase. It's a reshuffle, and it changes which card is actually the smart buy.

What actually changed in August

This is the same DRAM and GDDR7 shortage that's been rattling the whole GPU market, but it's landing unevenly. Mainstream, high-volume cards — the 5060 Ti and 5070 — got repriced hardest. The low-volume flagship 5080 barely moved. That's worth sitting with for a second: the card that was the 'safe mid-range pick' in June is now worse value than the tier above it.

Every tier, priced honestly, right now

The budget tier is the standout. Arc B580 and the RX 9060 XT weren't caught up in the same repricing wave as the RTX 50-series, so they're now the cheapest VRAM per dollar by a wide margin — but capacity is still capacity. If you're unsure whether 12GB or 16GB actually covers your use case, our breakdown of what fits in 12GB is worth reading before you buy on price alone.

The catch with the budget cards

Arc B580 has no CUDA. Most local-AI tooling runs fine on it through Vulkan or SYCL backends, but anything CUDA-only is a dead end. The RX 9060 XT leans on ROCm, which has closed a lot of ground this year but still trails CUDA in day-one support for new model releases — our own switch from Nvidia to AMD covers exactly where that friction shows up in practice.

What about the RTX 5090 and workstation-class cards?

For 32GB-plus needs, the RTX 5090 or a step into workstation territory is the honest answer — and it's a different budget class entirely. Most people chasing a specific 20–30B model don't need it. If you're running something like gpt-oss-20b, a 16GB card already covers full context comfortably; save the 5090 money for a model that actually needs 32GB.

Prices are moving fast enough now that this breakdown needs a recheck in a month, not a year. If the pattern from June to August repeats, the mainstream tier keeps getting worse relative to the high end — which is a strange sentence to write, but here we are.