Aluminum to Silicon, Part I: Why GPU Neoclouds Run on Airline Economics

gist

Former ABS trader Alex Song argues GPU neoclouds work like airlines and their debt like aircraft engine ABS, not like software businesses. Both borrow against equipment from concentrated makers, park it in scarce infrastructure, and rent it by the hour before amortization catches up. By his estimates, fixed costs exceed 85% of a neocloud's cost base and a debt-financed cluster needs about 65% to 75% utilization to break even. His warning is that most GPU debt is sized like whole-aircraft ABS, while the collateral behaves like engine ABS with no maintenance reserve.

ideas

• Idle GPU-hours expire like empty seats. An unsold hour on an energized H100 or B200 cluster can never be sold later, and debt service accrues whether the rack is busy or not.
• Utilization is the metric that matters. Song puts the marginal cost of running a GPU at roughly 5 to 7 cents per kWh of electricity, so nearly all return on invested capital comes from the band between breakeven and full utilization.
• Neoclouds sit between the OEM and the landlord. NVIDIA plays Boeing and Airbus, capturing most gross margin and setting the upgrade cadence, while powered shells and grid interconnects are the new airport slots.
• Three cabins of capacity. Multi-year take-or-pay contracts with labs and hyperscalers underwrite the debt, on-demand rentals are the price-shopped economy cabin, and spot capacity is sold at a 60% to 80% discount as standby, in Song's telling.
• GPU SPVs copy the aircraft financing template. Lenders such as Blackstone, Magnetar and Carlyle fund GPUs through bankruptcy-remote SPVs at 70% to 80% LTV and 9% to 14%, amortized over 36 to 48 months, by his account; he treats the data center as the long-lived airframe and the GPU as the short-lived engine.

quotes