How the numbers fit together
The memory story, the two-mask story, and the latency story are not separate arguments. They are different faces of the same economic model.
Back to atlas
Economics
Why it scales
Economics Of Specialization
The repo’s tables make this explicit: less flexibility, less memory overhead, less exotic cooling, and a faster mask-customization loop are all treated as part of a single economic package.
Key points
This is the page that translates architectural specialization into operating leverage.
The public site should treat these values as reference economics, not unqualified project claims.
The strongest message is that the same design choices can affect both serving cost and iteration speed.
Reference cost
$0.0075 / 1M tokens
Presented in the project’s comparison data as a target envelope.
Rack posture
12-15 kW / rack
The reference story avoids ultra-dense cooling assumptions.
Design cycle
~1 week
Automation speed is central to the business case.
Presentation rule
The site should make clear which economics are current reference targets and which are measured project outputs, so the public narrative stays ambitious without becoming careless.
Tie every number to either reference or measured evidence.
Keep cost and infrastructure claims coupled to architecture choices.
Do not flatten the tradeoff against flexibility.