Great read, one of those pieces I wish I had written. I think the majpr takeaway is the role of the engine in data center financing. In aircraft ABS the engine's sheer cost makes it the determinative factor of the financing terms, just like GPUs in In AI data centers' with 2 main difference: 1. The duration of the residual value Engines lose value with use, and lessors collect maintenance reserves to cover it. GPUs lose value through obsolescence, and nobody reserves for that. Every flight a plane makes impacts the residual value of the remaining parts of the plane. Every few flights the lessor takes the plane for maintenace with expenses financed by the maintenance reserves the lessor collected. The lender’s collateral changes in value according to the engine cycles & the fixes that resetored some value that was lost. But the engine itself can serve the lessor for 25 years & be sold off later in a secondary market. GPU's OTOH don't jus go obsolete, it happens a lot faster on a 4-6 year time span. That implies the entire financing instrument for data centers is a short duration risk with more tailwind risk that depends on Nvidia, TPU or new SRAM. No one really knows what the value will be in 2 years of B200, let alone in 5 years. its all a guessing game. 2. GPU Engines impact revenues directly. The price of a flight is determined by market price alone, irrespective whether the flight from JFK to Heathrow has a GE GEnx-1B or a Rolls Royce Trent 1000. The price of the flight will be the same, regardless. No one knows what engine was on the plane they flew in & it has no impact on the price. Compute price is derived directly from which GPU is being rented out . Each chip performs differently & delivers a different number of tokens, but theres also a variance across data centers with the same GPU that can impact the number of tokens per hour available.