$3-$10 in GPU time can now fine-tune a 7B-8B open-weight model with LoRA or QLoRA, according to the report, versus vendor quotes that can reach $50,000 for custom model work.
Five techniques drive those savings: LoRA trains under 1% of parameters, QLoRA quantizes base weights to 4-bit, Unsloth reduces training time, DPO avoids RLHF's extra reward-model pipeline, and GPU right-sizing cuts rental waste.
An RTX 4090 at about $0.31 an hour is presented as sufficient for many 7B-8B adapter runs, while A100 or H100 hardware is framed as necessary mainly for larger 13B-70B-plus jobs or multi-GPU workloads.
The article argues compute is no longer the main cost center: complete projects still often land at $5,000-$15,000 once dataset creation, repeated experiments and evaluation are included.
It also warns fine-tuning is often unnecessary, citing a 2026 guide that says roughly 80% of requests framed as fine-tuning are better solved with retrieval.