Matt Dratch(@DratchCap):Great post by one of my new fav investors @FredaDuan. Have a few thoughts to add on the return side, including why ROE not ROI is the metric to watch for large % of spenders: 1. A lot of 2027 capex is for GW that turn on in 2028. So ‘27 revenue vs ‘27 capex overstates the gap. 2. ~30% of the spend is shells and power that last 25-30 years. 15 is v conservative. 3. **Not every buyer pays cash**. The hyperscalers mostly do, but neoclouds, labs and SPVs are borrowing 75-85% against 5yr contracts (much of which is hyperscaler risk). Freda’s $2.5-3T is more than double the ~$1.2-1.5T the hyperscalers spend, and a lot of the rest is levered. Once it's financed like infrastructure, ROE is actually the metric that matters. Easiest to think about it per GW. On her numbers a GW costs ~$55B. Pay all cash and it needs ~$17B/yr of rent to make 15%. Use 80% debt, even at 10%, and it only needs ~$14B to make 15% ROE (13bn at 7.5%). So the bar is $13-17B depending on who's buying. Her own math has a GW bringing in ~$20B today tho (prices appear to be going up…) On the whole 50GW that's ~$700-850B a year once it's live, against her ~$1T.

2026.10.18 14:30

Great post by one of my new fav investors @FredaDuan. Have a few thoughts to add on the return side, including why ROE not ROI is the metric to watch for large % of spenders: 1. A lot of 2027 capex is for GW that turn on in 2028. So ‘27 revenue vs ‘27 capex overstates the gap. 2. ~30% of the spend is shells and power that last 25-30 years. 15 is v conservative. 3. **Not every buyer pays cash**. The hyperscalers mostly do, but neoclouds, labs and SPVs are borrowing 75-85% against 5yr contracts (much of which is hyperscaler risk). Freda’s $2.5-3T is more than double the ~$1.2-1.5T the hyperscalers spend, and a lot of the rest is levered. Once it's financed like infrastructure, ROE is actually the metric that matters. Easiest to think about it per GW. On her numbers a GW costs ~$55B. Pay all cash and it needs ~$17B/yr of rent to make 15%. Use 80% debt, even at 10%, and it only needs ~$14B to make 15% ROE (13bn at 7.5%). So the bar is $13-17B depending on who's buying. Her own math has a GW bringing in ~$20B today tho (prices appear to be going up…) On the whole 50GW that's ~$700-850B a year once it's live, against her ~$1T.

Freda Duan@FredaDuan

2026.10.17 22:14

The Trillion-Dollar Question Nobody Really Has a Good Answer To - Where Is the ROI? We spend a lot of time debating how many GW of AI compute will come online. But the bigger question is: who ultimately pays for all of it, and can the economics justify trillions of dollars in investment? I find it helpful to think about AI economics as a chain: CapEx → Compute / Inference Revenue → Model / API Revenue → End Demand One important nuance: the market often uses model company revenue (@OpenAI , @AnthropicAI , etc.) as a proxy for AI demand. But outside direct consumer subscriptions, much of that isn't necessarily ultimate end demand. A company paying for Claude Code is still an intermediate customer in the economic chain. The real economic value comes when that company uses AI to generate incremental reve or reduce costs. --- Let's put some numbers around 2027. 1. CapEx: US$2.5-3.0T Assuming ~50GW of new global AI compute capacity, we're looking at roughly US$2.5-3.0T of total infrastructure investment. ~70% goes toward chips and related compute equipment, with the rest covering buildings, cooling, power and supporting infra. Importantly, this is much broader than the often-quoted ~US$1.2T of hyperscaler CapEx. Neoclouds, colocation providers, private developers and others are also funding the buildout. 2. Compute / Inference Layer: US$1T+ Revenue Potential If all 50GW were deployed and commercially utilized, the capacity could theoretically support over US$1T in annualized compute infrastructure revenue. For context, combined CSP cloud rev = sub US$600B on consensus est. 3. Model / API Layer: How Much Revenue Do We Actually Need? @OpenAI and @AnthropicAI are rumored to be approaching US$200B of combined ARR by YE26 (source: public data). My instinct is that we need to see something closer to US$400B+ by FYE27 to begin supporting the scale of investment we're talking about. Rough sanity checks: 1) US$2.5-3.0T of CapEx implies roughly US$400-480B in annual depreciation. 2) Total GW online grows +70% in FY27E, hence ARR needs to grow at least 70% (to $350B)? --- So, was the ROI ever a slam dunk? Never, really. And I think we should be intellectually honest about that. Don't get me wrong. I believe AI is one of the biggest technology supercycles we've ever seen. But when it comes to justifying trillions of $$$$ of CapEx, even the bull case rests on extrapolating the early success of AI rev. Some extrapolate linearly. Others exponentially or logarithmically. The resulting forecasts can be wildly different. And that's precisely why the market tends to overshoot in both directions. It is not really backed by a solid set of numbers, but rather, extrapolations. --- Which brings me to how I think about trading this cycle. The market isn't necessarily trading the absolute level of AI CapEx or revenue. It's trading the rate of change, particularly revisions to expectations. Think of it as two cars racing: 🏎️Car 1: AI Infrastructure Investment How quickly are CapEx expectations being revised upward? For example, since the beginning of 2026, CSP CapEx expectations have been revised upward by roughly 40-50%. 🏎️Car 2: AI Monetization How quickly are AI rev expectations being revised upward? How quickly is actual end demand catching up? The market wants to see 🏎️Car 2 outpacing 🏎️Car 1. If CapEx estimates keep moving higher while AI monetization fails to keep pace, investors will increasingly question incremental returns, even if absolute AI revenue is still growing rapidly. AI can be the biggest technology supercycle of our lifetime, and investors can still overpay for the infrastructure along the way. Both can be true imho.

커뮤니티로 전달