The Factory Frame: From Data Center to Production Line
Retire the word "data center." The building that trains and serves large models is a factory, and a factory has exactly one job: convert electricity and data into finished tokens at the lowest cost and highest sustained rate. That reframing is not a metaphor for color — it imports a full discipline. Throughput is set by the slowest station, not the fastest chip. Profit is cost per finished unit, not the price of the most impressive machine. The line must be balanced so nothing starves or floods. Idle silicon is a paid-for press gathering dust. Once you see the building as a production line, the right questions replace the wrong ones, and the wrong ones were about components.
Knowledge check
Why does buying the fastest available accelerator not guarantee the lowest cost per token?
Show the answer
Because output is a property of the whole line, not one station. The slowest resource — memory bandwidth, network fabric, power, cooling, or the scheduler — sets sustained throughput, so a fast chip that idles waiting to be fed produces fewer realized tokens per dollar than a modest chip in a balanced, well-utilized line. Cost per token is total line cost divided by tokens actually delivered at real utilization, not peak specification.