google cloud next kicks off this week and every hyperscaler bull is about to cheer the exact move that historically kills the margin they're long

before: google paid nvidia a markup on every h100. margin flowing to santa clara.
now: google's in talks w/ marvell on custom silicon plus a new inference-specific TPU variant. the capex is moving internal.
amazon did it with graviton in 2018. microsoft announced maia in 2023. each time the hyperscaler stopped paying its supplier's markup because the markup became the biggest line item on the income statement.
inference pricing is already under the same pressure.
when a hyperscaler starts competing with its own supplier, the spot price goes first. the incumbent goes second.
a hyperscaler building its own silicon is a utility company building its own turbines, because the electricity rate card stopped making sense.