About a year ago, Andreessen Horowitz (the largest venture capital firm in the world), published a blog post showing that the cost of LLM inference is dropping 10x per year.

Andrej Karpathy, a member of the founding team at OpenAI and subsequently Tesla’s Sr. Director of AI, says that in 2019 it cost $43k to train GPT-2. Today he can train a model that performs similarly for $73. That’s 600x cheaper in 7 years. Not the same metric, but maybe more important.