Hyperscale clouds will get their ROIC on AI capex, but there is still a debate to be had.

While $亞馬遜 (AMZN.US)$ $微軟 (MSFT.US)$ may overbuild in the short term if AI demand weakens, they are not renting GPUs at a loss. They've been very clear that capex is fungible for other purposes if demand disappoints vs. forecasts, perhaps w/ some time lag (would be more concerned abt others like $甲骨文 (ORCL.US)$ / Coreweave recovering ROI in a downside case).

The debate regarding AI capex ROIC should be centered around:

1) The ROIC of GPU buyers ex-clouds... Mostly: will frontier model companies make $? OpenAI has negative gross margins if you count training costs as COGS. $谷歌-C (GOOG.US)$ $Meta Platforms (META.US)$ P&L's dont yet reflect the lagged depreciation expense of their recent GPU ramp, & cant cut opex forever to drive EBIT. They will begin to feel pressure of rising Depreciation next year & need to overcome w/ revenue acceleration or else cons. margins have risk;

2) Are the end-renters of GPUs (i.e., cloud customers) getting adequate ROI? $Snowflake (SNOW.US)$ said this week they're feeling downward pricing pressure on AI products while the price for chips isn't coming down, pressuring margins. They now mostly have enough supply and wont buy any more GPUs until they see revenue to support it (implying they dont see revenue to support it today).