“The business of selling inference is highly profitable, while selling raw compute is less so, with fine-tuning APIs like Tinker's positioned in between.”