ilari

We modeled every dollar of $googl cloud capex as a standalone investment. The returns math isn't the problem. The commitment math might be. Here's my walk-through: THE RAMP Quarterly capex: $6.3B (Q1-23) → $44.9B (Q2-26). That's 7.1x in 13 quarters. Two distinct phases — a first step in Q4-23, then a near-vertical move from Q1-25 (+215% in six quarters, no plateau). Capex is now 29.7% of revenue. Revenue per $1 of capex has collapsed from 10.5x to 3.4x — below the 2018 trough. I think that number alone explains a lot of the market's discomfort. THE BREAKEVEN Given a 6-year asset life, 2-quarter deployment lag, 17% cash tax and 10% WACC, a dollar of capex needs to produce just $0.18 of annual revenue to earn its cost of capital. Observed: $0.54. So that's ~3x cover. On margin, breakeven is 25.9%. Google's incremental cloud operating margin is 53.7%. THE RETURN ~$0.18 of NPV per $1 of capex over the asset's 6.5-year life. Discounted breakeven at 5.5 years. That's a 15.1% IRR per $1 of capex. Steady-state ROIC of 23.5% on net invested capital against a 10% cost of capital — a 13.5-point spread. For a business with Google's moat and durability, that isn't marginal. That's an excellent use of shareholder capital. WHAT IT ACTUALLY DOES TO EPS One year's capex cohort ($132B, based on LTM capex) generates ~$15.6B of incremental annual NOPAT at full ramp — ~12.7% of current NOPAT, or ~$1.27/share. But it lands 2.5 years after the cash goes out, and it's a one-time step-up in the level of earnings, not an addition to the growth rate. In steady state with flat capex, retiring cohorts offset new ones and the growth contribution goes to zero even at 23.5% ROIC. Sustained EPS growth requires capex to keep rising. Which it is.   THE PART THAT SURPRISED ME The standard bear case is that hyperscaler margins are manufactured through depreciation. Google did extend server lives 3→4 years in 2021 and 4→6 in 2023. So we stripped out capital not yet in service and measured implied asset life on productive assets only. It peaked at 8.0 years in Q3-2024 and has fallen every quarter since — to 6.7 today. The depreciation rate is rising. Depreciation policy has gotten more conservative through the build, not less. The reported figure only looks stretched because a third of the asset base is a half-built data centre.   DEMAND IS CONFIRMING Backlog: $108B → $520B in five quarters. Coverage from 2.2x to 6.7x LTM cloud revenue. Cloud revenue +82% Y/Y. Cloud is 20.7% of total revenue, up from 14.1%. Cloud operating income: $191M (Q1-23) → $8.8B (Q2-26). And the return metrics have improved as capex ramped, not deteriorated.   SO WHY THE SKEPTICISM FCF turned negative in Q2-2026 — the first negative quarter in 54 quarters of history. Purchase commitments went from $72B to $811B. That's 6.1 years of current-run-rate capex, largely non-cancellable. The return per dollar is excellent. What's gone is the option to stop. That's the asymmetry. Markets price optionality, not just returns.   THE CAVEAT My return metrics lag six quarters — you need forward revenue to measure them. Today's 0.54x describes capital deployed around Q4-2024, when LTM capex was ~$53B vs $132B now. So "returns improved as capex ramped" documents the first ramp. The 2025-26 cohorts aren't measurable until late 2027. That's precisely where diminishing returns would surface first, if they do at all.   Assumptions are mostly conservative: 6-yr asset life (vs 6.7 observed), 45% incremental margin (vs 53.7% actual), 10% WACC, 5% residual. Two lean slightly favorable — revenue per dollar at 0.35 vs 0.33 observed, and a 17% cash tax rate vs 18.4% effective.   Bottom line: per-dollar, this is a no-brainer for management. The risk was never the return. It's the size and irreversibility of the commitment.