AI capital spending
AI capital spending can outrun paying end demand
Circular bookings and cheap-duration assumptions can fund data halls faster than outside customers generate cash.
Verification indicators
- Hyperscaler capex funding
- GPU utilization
- outside token revenue
- data-centre occupancy and power load
Falsification conditions
- New halls fill with independently funded, paying workloads as they complete
- GPU utilization holds through new supply without vendor financing
- Cash returns rise while the 10-year yield remains restrictive
Change history
-
Physical capacity added Evidence holds
New evidence: The follow-up moved the test from valuation to funding, utilization, and paying occupancy.
Interpretation change: The thesis gained measurable operating tests; it was not marked as proven.
-
Valuation thesis Verification in progress
New evidence: The first note focused on long-duration valuation.
Interpretation change: Initial publication.
Related research
AI pulled 2035’s cash flows into 2024’s multiple. That is not a tech story. It is a duration story
Like the better skeptics of this cycle, this desk treats much of the AI complex as a valuation that already spent the future.
AIEmpty halls: Nvidia, overbuilt watts, and this bubble
The last AI note on this desk was duration: 2035 already sat in 2024’s multiple. The escalation is physical. The bid for the chips is walking around the same table.
RatesKorea looks fine from the asset. The stall ticket says the storm is not here yet
KDI: real wages +0.3% in the first half. Asset income +21% in Q2. July retail −2.4% m/m while chip capex jumped. BOK took the base rate to 3%.