# AI capex — X 热门讨论 (2026-09-25 22:13 UTC)
## @business (Bloomberg) · 09-25 15:05 · ♥30 ↻10 💬9 Spending on AI infrastructure by the five largest US hyperscalers is set to increase by more than half next year to $1.2 trillion, according to strategists at Goldman Sachs https://t.co/4DyXuckJhf https://x.com/business/status/2103501039362080772
## @econcallum (Callum Williams) · 09-25 20:54 · ♥33 ↻7 💬4 In July we argued that the AI ecosystem required $2-3trn annual revenues in perpetuity to deliver a payback on AI capex. Yesterday Goldman published a similar report with extremely similar results https://t.co/vj4RiBpayd https://x.com/econcallum/status/2103589021738029156
## @loraclexyz (🔮(𝕏ᵀ𝕏) - loracle.hl) · 09-25 21:34 · ♥37 ↻0 💬3 People are going to realize it takes a frontier human to extract productivity out of a frontier model, and there is simply not enough frontier human to compensate for the capex relying on revenue from training frontier models (yes if you remove training compute demand the whole thing is unprofitable already, and that’s before token price depreciation). Put it simply, demand for intelligence isn’t infinite and is in the shape of a bell curve, and cheaper models are slowly eating their way into the bell curve, leaving frontier training cost rely on a smaller and smaller tail of intelligence demand. In the end you are left with a billion dollar training run which only use is to solve Navier Stokes, because all the other tasks that are useful for the current economy can be run on cheaper models > 引用 @zerohedge: "Token demand growth will need to outpace declining token prices to support continued growth in investment spending. Frontier models are currently a key source of demand for hyperscaler compute. However, the rise of competitive open-source models has contributed to a decline in average token prices. Measures of frontier token demand slowed in July..." - Goldman https://x.com/loraclexyz/status/2103599024947523645