The AI Factory

10 atomic recall cards. Export to Anki and let spaced repetition do its slow work.

In Anki: File → Import, choose this TSV, set field separator to Tab, deck = Sand to Silicon · Ch 16, note type = Basic.

Front	Back
How many NVL72 racks make up one DGX SuperPOD?	14 NVL72 racks.
What is the total GPU count and FP4 compute of a DGX SuperPOD?	1,008 Rubin GPUs delivering 50.4 exaflops of FP4 compute.
How much HBM4 memory does a DGX SuperPOD carry?	1,046 TB of fast HBM4 memory.
What networking fabrics connect racks within a SuperPOD?	Quantum-X800 InfiniBand or Spectrum-X Ethernet; BlueField-4 DPUs handle protocol offload.
By how much does Rubin reduce inference token cost compared to Blackwell?	Roughly 10×.
By how much does Rubin reduce the GPU count needed to train a comparable MoE model?	4× fewer GPUs required.
What is the fundamental output unit of an AI factory?	Tokens — pieces of language, pixels, video frames, or molecular structures generated by the models running on the racks.
How long does it take to build a new AI factory from groundbreaking to first GPU online?	Years — it is a binding constraint on how fast the industry can grow.
What does the chapter say the reinforced concrete floor must support?	Tens of thousands of pounds per square meter, to bear the weight of the racks and coolant infrastructure.
What is the chapter’s description of the journey a Rubin GPU takes from raw material to running rack?	Roughly six months, crossing four continents, involving perhaps eighty suppliers, twenty governments, and several technologies no single nation can build alone.