Netflix topline growth slowing to just under 12%, its lowest since the 2021-22 scare, makes the short-form video angle feel more real. Do you think the real constraint is total attention, or the willingness to pay for yet another premium subscription?
With A100 rental rates still climbing, cost per million tokens is worth rerunning on your own serving stack every quarter. Moving a small model's traffic between A100s and H100s can come out cheaper or pricier depending on where rates landed that month.
The GPU section and the electrical equipment section are the same chart. Rental rates holding above a six year straight line is only evidence about the chip if the chip is the scarce input, and your own capacity data says it is not: electrical equipment capacity actually fell while prices surged. What the renter is bidding for is the substation and the shell, with a GPU attached. That premium should compress when switchgear capacity lands, not when the silicon ages, which means it is not really a useful life argument at all. One test that separates the two: split the Ornn series by region and compare rates in power constrained metros against places with spare interconnect. If the surplus over straight line is concentrated where the grid is tight, the depreciation debate is about electricity and real estate.
The GPU chart is a useful reminder that depreciation schedules are accounting conventions, not physics — real asset life gets set by demand, and demand here has clearly outrun the six-year straight-line assumption.
The same error shows up in retirement planning. I keep seeing corporates fix a retirement date and a drawdown rate as if both are contractual, then never stress-test what happens if returns or medical inflation break the assumption.
Netflix topline growth slowing to just under 12%, its lowest since the 2021-22 scare, makes the short-form video angle feel more real. Do you think the real constraint is total attention, or the willingness to pay for yet another premium subscription?
With A100 rental rates still climbing, cost per million tokens is worth rerunning on your own serving stack every quarter. Moving a small model's traffic between A100s and H100s can come out cheaper or pricier depending on where rates landed that month.
The GPU section and the electrical equipment section are the same chart. Rental rates holding above a six year straight line is only evidence about the chip if the chip is the scarce input, and your own capacity data says it is not: electrical equipment capacity actually fell while prices surged. What the renter is bidding for is the substation and the shell, with a GPU attached. That premium should compress when switchgear capacity lands, not when the silicon ages, which means it is not really a useful life argument at all. One test that separates the two: split the Ornn series by region and compare rates in power constrained metros against places with spare interconnect. If the surplus over straight line is concentrated where the grid is tight, the depreciation debate is about electricity and real estate.
The GPU chart is a useful reminder that depreciation schedules are accounting conventions, not physics — real asset life gets set by demand, and demand here has clearly outrun the six-year straight-line assumption.
The same error shows up in retirement planning. I keep seeing corporates fix a retirement date and a drawdown rate as if both are contractual, then never stress-test what happens if returns or medical inflation break the assumption.