How the Math Works
The author's math goes like this: frontier models' API prices are falling, but usage is rising faster, and agentic workflows readily burn massive amounts of tokens, so the bill keeps climbing. On the other side, open source models' capabilities are rapidly closing in, running on self-bought or rented hardware at extremely low marginal cost, and paired with engineers in lower-labor-cost regions for supervision and backstopping, the blended unit price may come out ahead. The conclusion: "subscribing to a frontier lab" isn't necessarily an enterprise's long-term default.
The Two Points the Objectors Hit
The pushback aims mainly at two spots. One is hidden costs: the ops of self-hosting and the rework caused by the open source model's capability gap are hard to put into the spreadsheet. The other is the quality ceiling—the hardest tasks are still only doable by the top models, and the money saved may be paid back at a critical moment. Even so, the trend the post represents is real—the cost structure of inference is loosening, frontier labs' pricing power isn't ironclad, and enterprises are starting to seriously do the homework on alternatives.
via: Hacker News