the RTX 5070 just hit $900. I ran the local-AI math and almost couldn't recommend it

The RTX 5070's median US retail price hit $899.99 this month, according to Newegg listings tracked by Tom's Hardware — up from $659.99 in June. That's a 36% jump in roughly two months, on a card Nvidia launched at $549. I've been telling people the 5070 is the sane entry point into local AI for months. At $900, I had to actually redo the math instead of repeating myself.

Why the price moved, and it's mostly not Nvidia doing it on purpose

This isn't a scalper story, and it isn't really an Nvidia story either. TechPowerUp's tracking points to rising GDDR7 costs as the driver, with board partners passing along higher wafer, packaging and power-delivery costs on top. The RTX 5060 Ti 16GB actually got hit worse — up 39% to $804.99, about 88% above its original launch price — because it packs more memory chips per card, and memory is exactly what's short right now. The RTX 5090 only rose 9% this round, to roughly $4,700, but that's because it was already 135% over its $1,999 MSRP before this cycle even started. Guru3d's independent retailer check found the same pattern across a separate set of listings.

Line them up and the RTX 5070 is no longer the obvious pick it was in June. It still has the fastest memory bandwidth of this group by a wide margin, and Nvidia's CUDA stack is still the path of least resistance for anything you copy off Hugging Face. But you're now paying a 64% premium over MSRP for that convenience, while the 5060 Ti — more VRAM, slower bandwidth — is paying an even bigger one. The two budget outsiders, Arc B580 and RX 9060 XT, are both sitting closer to 20% over their launch prices. That gap is basically the whole story.

What 12GB and 672 GB/s actually get you in 2026

None of that changed with the price. The 5070 still does about 70 to 80 tokens per second on Llama 3.1 8B at Q4_K_M, and 50-80 tok/s on denser 9B models like Qwen3.5 — genuinely fast, genuinely usable for daily coding-assistant or chatbot work. What changes is the price you're paying per token of that speed. Push past 12GB — a 14B model, a longer context window, an image model bolted on top — and you're offloading to system RAM, and the 5070 becomes a 7 tokens-per-second card instead of an 80 tokens-per-second one. 12GB was already the tight end of comfortable before the price jumped; now you're paying luxury pricing for a card with a real ceiling.

What I'd actually buy at today's prices

That math is blunt on purpose, and bandwidth isn't in it — the 5070's 672 GB/s beats the Arc B580's 456 GB/s by a real margin, so this isn't "just buy the cheapest GB." But if what you actually want is headroom to load a 14B or 20B model without falling off a cliff, the RTX 5060 Ti 16GB is the more defensible Nvidia buy right now even at $805, because you're buying VRAM, not raw throughput. And if you can live without Nvidia's CUDA tooling, both the Arc B580 and the RX 9060 XT are pricing far better than either Nvidia card, in a market where everything just got more expensive than it was in June.

I don't think $900 is a permanent price, but I also don't think anyone honestly knows when it stops — the DRAM and GDDR7 shortage driving this is a supply problem, not a marketing one, and those don't resolve on a predictable schedule. If your local-AI build can wait a month, wait. If it can't, buy based on what you're actually running, not what card sounded right back when it cost $650.