Read together, the period's announcements describe three distinct answers to the same problem: reducing dependence on whoever manufactures the hardware the models run on.
The first is contractual and financial: lock in gigawatt-scale capacity with a manufacturer and buy equity in it, turning a customer relationship into a shareholder one, with what that implies for delivery priority.
The second is designing your own chip, co-engineering hardware and model to run together, with efficiency gains nobody buying off the shelf can match.
The third is acquiring the company that etches models directly into silicon, giving up flexibility to eliminate traffic between memory and processor.
All three share a diagnosis and differ on timeline. None solves this quarter's problem, which explains why, in parallel, all of them keep buying capacity on the market at whatever it's asking.
For smaller operations, the applicable version of that lesson is simpler and doesn't cost billions: have a tested alternative supplier before you need one.