这位工程师给算了一笔账：微调7B模型只要几十英镑，而且托管推理按量付费，建议老板们别再犹豫了。
作者研究后发现，在Together AI或Fireworks上对7-16B模型做LoRA微调，计算成本仅需£10-50，70B QLoRA也不超过£200。推理方面，托管端点按基模型费率收费约£0.20-0.90/百万token，自托管spot H100约£1-2.50/小时，但流量波动时容易超支。数据准备和GDPR合规仍需投入时间。整体门槛已大幅降低，适合企业用自有数据试点。
This, literally, got me thinking about actually do…
This, literally, got me thinking about actually doing it.
I went in and did a quick bit of research on how practical it really is. Fine-tuning itself is now stupidly cheap on platforms like Together AI or Fireworks. LoRA on a 7-16B model often £10-50 in compute for a decent business dataset. Even 70B QLoRA stays under a couple hundred quid.
The gotcha is inference and ops. Managed endpoints let you serve the fine-tune at base model rates (roughly £0.20-0.90 per million tokens for Llama-class), pay-per-use, no idle GPUs. Self-host on spot H100s (around £1-2.50/hr) wins at high steady volume but eats cash if usage is spiky or you ignore monitoring.
Data prep and evaluation still take real time, and UK/telecom data means thinking GDPR and residency hard. But the barrier has dropped hard. It's less can we and more does leadership see custom AI as proper infrastructure.
For businesses tired of rising token bills and access risks, this is looking like the sustainable play. Worth a small pilot on your own data? 🤔