Artificial Intelligence: How Much Will It Cost You?
summary
Generative models require huge up-front capital: accelerator hardware costs dominate (H100-class cards cost tens to hundreds of thousands each) and training a Llama-scale model runs on the order of $1M–$15M, while GPT‑4–scale training plausibly hits $10M–$50M. Serving adds per-query cost (roughly $0.005–$0.07 per conversational turn) plus the need to amortize training, so a $4M model needs on the order of 100M conversational turns before serving equals half the paid training. Rent cloud accelerators unless demand and budget justify owning.
AI-written summary · about
chapters (8)
- 0:00AI costs introduction
- 3:30Corporate money basics
- 10:12Capex vs Opex in AI
- 18:20Generative AI explained
- 22:55Why AI accelerators cost
- 27:05LLM training and serving costs
- 33:25Practical AI cost estimates
- 35:54Train your own LLM?
