9 October 2026
Interconnects predicts AI inference costs will fall sharply within years
First reported
Interconnects ran this on .
- The author expects agents to optimize measurable metrics like tokens per second and cost per answer.
- He predicts the cost of model intelligence will fall near-exponentially as inference approaches hardware limits.
- Longer-term gains are expected from designing chips and models together, which he calls accelerator-model co-design.
Our read
Only Interconnects, out of 30 newsletters tracked, covered this story. It follows our August 28 story, 'AI model prices fall while building costs rise,' on the same cost thread.
Earlier on this story
Written by Newskeryx from which newsletters covered this story and from our own archive.
How it was covered
InterconnectsNathan Lambert
Training and inference efficiency metrics, cost to serve models, accelerator-model co-design, and automating pretraining research