USD 1trn Hyperscaler CAPEX sanity check

The assumptions:

• For simplicity i assume no existing AI infrastructure pre 2025
• Consensus CAPEX estimates for Hyperscalers for 26,27 and 28 are USDbn 637, 804 and 850 (just MSFT, Amazon, Meta and Alphabet) Assuming roughly 100 of these are non-AI related.I am assuming the remainder is around USD 60bn (Coreweave, Oracle, X etc.)
• Required return on invested capital 12% (above WACC, but low for these companies)
• Blended depreciation period of 8 years – 5-6 years for chips (as per company accounting assumptions and longer for infrastructure
• EBITDA margin on AI datacenter of 45%
• Effective tax rate 17%
• No CAPEX growth from 2028
• Finally LLM compute cost share of revenue 70%

The result:
Table: https://imgur.com/a/cXHYATs

This means that the needed incremental revenue to justify the current investments in AI would be USDbn 165, 384, 686, 1137 in 2025 to 2028, respectively. As mentioned, aggressively assuming 70% of LLM provider revenue goes to rent compute from these companies. They would need to bring in USDbn 235, 548, 980, 1625 in the same period, respectively.

What it means:

This means that LLM revenue needs to be 1,7%, 3,0%, 4,9% of US GDP in 2026 to 2028, respectively. Another point of comparison – this would represent 10%, 18% and 29% percent of current global IT spend in the same period, respectively.

I am not saying that it wont reach does levels (eventually) i am just saying that it seems highly unlikely that it will happen this quickly.

This would be acceptable (but still not great) if this was long lived assets that could wait for the demand to pick up – like fiber optic cables (dot com), but if it takes lets say 5 years for demand to catch up – then the majority of these assets are already outdated.

Conclusion:
Current CAPEX will most likely not yield great returns for Hyperscalers, and its not unlikely that most if not all neoclouds will go bust (Coreweave). Further, i believe CAPEX will come down – which hits the semiconductor supply chain more or less directly.