Post

Right now many services are subsidized and the companies behind them are buying up compute + memory + storage en masse, resulting in the hardware being more expensive. As subsidies decrease and prices go up, i expect demand for services and HW costs will come down. Could change the economics.

0 reposts8 likes

you can *kinda* build something reasonably-ish priced today. a couple of B60s, B65s or B70s will give you plenty of VRAM to run Qwen3.8-27B in Q6_K_XL w/ no kv quant at max context at good speeds, or you can go with 20GB 3080s to run even faster. so you're talking about ~$2K of hardware.

For sure, you can get around 20tps on a quantized qwen3.8-27B for under $2k. If you want frontier-quality intelligence (IMO glm-5.3-flash is there) it’s around 10x the parameters and 10x the cost to run it. Local is still not competitive with a 50-$200 subscription there.