I wouldn’t get the cart before the horse here. A lot of this is still up in the air. Tokens are still subsidized, and that bill comes due eventually. However…
- Massive compute build-out (and even electrical infrastructure) is underway and will be for another half-decade at least. Higher supply of compute means compute is cheaper.
- Open-weight models are really not that far behind proprietary frontier models. A few months behind, at most.
- Researchers are constantly finding strategies to lower per-token costs, even if the supply of compute weren’t increasing (and it is)
- Running LLMS locally for security and regulatory reasons is already a thing and will be increasingly important. Those local LLMs will require in-house or contracted building and maintenance.
So it’s still way, way too early to make any of these pronouncements about whether it’s all over and we should just pack up and become plumbers or if it will be too expensive for anyone except giga-trillionaires to write any code etc.
