Open Model Fine-Tuning Costs Plummet, Bottleneck Shifts to Inference and Ops
koltregaskes · x · 2026-07-03
The author details the current costs of fine-tuning open-source models: performing LoRA on 7-16B models via platforms like Together AI and Fireworks typically requires only 10-50 GBP in compute power, and 70B QLoRA generally stays within a few hundred GBP. The real challenge lies in inference and operations. Hosted endpoints offer pay-as-you-go billing for fine-tuned models at base model rates (around 0.2-0.9 GBP per million tokens for Llama-like models), eliminating idle GPU costs. However, for large-scale, stable traffic, self-hosting spot H100s (approx. 1-2.5 GBP per hour) remains more cost-effective.
More from Infra
- Encrypted GPT-2 speedrun challenges FHE inference: best is 3.031 s/token, goal sub-1s — tensorqt · 2026-10-27
- Open-source Latch compiles PyTorch into training megakernels, 2.75x faster on H100 — sguada · 2026-10-27
- Stanford launches GPU-centric supercomputer Marlowe for data science and AI research — StanfordHAI · 2026-10-27
- NVIDIA Claims HGX Rubin NVL8 Delivers 10x More Tokens per Megawatt Than B200 — thursdai_pod · 2026-10-27
- Together lets multiple LoRA post-training experiments share the same GPU capacity — togethercompute · 2026-10-27
- Commoditization dilemma: either human labor or GPUs become the commodity — _AustinCalvert_ · 2026-10-27