Stranded Compute: Two Posts Worth Taking Seriously
Musk says compute will sit stranded in half-built data centers. Chamath maps who eats the loss. The verifiable record already points the same way.
Read the postThe blog
Notes from the practice.
Process notes, practical checks, and observations from building and maintaining the two volumes — written the way the book is written: plain, dated, and checkable.
More from the blog
The Data Has to Get There Somehow
Private connectivity and data transfer are the line items almost nobody puts in the cloud budget. A look at what dedicated links actually cost.
Read the post
Capacity Planning Used to Be Easier
Power, memory, and import controls have become the hard part of a capacity plan. Three factors, each with a dated source, and why the job keeps getting harder.
Read the post
Who Gets a Choice at Scale
Operate capacity at enough scale and you become an internal cloud provider. The next question is which choices you hand your internal customers.
Read the post
The Second Budget
Finance owns the actual budget. Cloud teams end up running a second one. It takes a couple of years for people to stop calling it the pretend budget.
Read the post
Old Chips Are Supposed to Get Cheaper
The H100 rental price peaked near $8 an hour, fell to $2, then rose again — on a chip that is now three years old. What that says about reading a price curve.
Read the post
A SKU Is Not a Machine
The same SKU is not always the same machine. At a fleet of thousands, the differences between otherwise identical servers stop being trivia and start being operations.
Read the post
Spot Instances Were a Lie
Spot pricing was never quite the market it claimed to be. Seventeen years of spot: the auction that was not an auction, the 2017 repricing, the death of the middle product, and where the product goes next.
Read the post
And then your cloud provider started making chips...
The transition started in 2015, ARM made it possible, and today it supplies most of the new CPU capacity at AWS. But Nvidia still owns the accelerator layer, and Google's TPUs show where the road ends.
Read the post
Capacity is not a CPU count
Workloads are gated by different resources — CPU, memory, disk. If you run capacity for databases, a CPU-only view breaks long before you reach scale.
Read the post