Why compute prices can't collapse: you can't buy 10 flops, only trillion-fold bundles
davidmanheim · x · 2026-09-23
Continuing the LLM cost-curve debate: a cited argument notes you can't buy 10 flops of compute separately — the same price buys a trillion times more — and the inverse (everything a trillion times cheaper) is impossible due to fixed-cost bottlenecks. Manheim adds that distilled models do cut costs, but the $0.01/mtok serving floor means the cheapest tiers go unserved, distorting the decline curve.
Related event: Toby Ord and Others Challenge Viral "AI Prices Fell 100,000x" Chart(5 posts)→
More from Infra
- Dreamscale serves NVIDIA's DreamZero robot model 8% faster with half the silicon — chris_j_paxton · 2026-09-23
- Agents Make Data Stacks Unforgiving: Stale Data Can Trigger Wrong Orders — bigdata · 2026-09-23
- Silent LLM Cost Errors: Inconsistent Token Accounting Undercounts Usage by Up to 46x — qaiser_mehdi · 2026-09-23
- Redis ships Radar, Search on Flex and agent memory tools as 97% back context engineering — antirez · 2026-09-23
- 24GB AMD GPU Struggles With Local Video Models: 2 Minutes for 2-Second 480p — OneMoreName1 · 2026-09-23
- ByteDance accounted for nearly 75% of Nscale's 2025 sales, used Norway site to access Nvidia chips — nathanbenaich · 2026-09-23