Microsoft has unveiled a structure designed for the age of large models. This AI super-factory connects two centers located in Wisconsin and Atlanta with an optical network intended for very fast transfers.
The idea is to unite distant sites so that they act as a shared supercomputer. Microsoft is betting on this architecture to support the training of increasingly heavy models, including those used by partners like OpenAI and its internal superintelligence team.
An AI super-factory, isn’t it impressive?
The new structure is based on a very concrete idea: bringing chips closer together to reduce latency. Teams have positioned the components across two dense floors to accelerate internal exchanges.
The whole setup is supported by massive cabling and cooling systems capable of dissipating heat from hardware that works continuously. This design allows for continuous calculations, even during peak activity periods, and minimizes wait times between tasks.
The network used connects both sites across thousands of kilometers with dedicated fiber optic. Part of this infrastructure comes from former acquisitions, while another part was constructed specifically for this project.
Teams have equipped these facilities with exabytes of storage and millions of processor cores. These resources are intended for the heaviest workflows.
With tailored protocols, Microsoft shortens data travel times and reduces delays during transfers. Each site can thus partake in the same training task, nearly in real-time.
Why this maneuver?
This distributed operation prevents some GPUs from idling while others are still computing. Microsoft aims to maintain a constant flow for all units to maximize resource utilization.
To bolster this giant mechanism, Microsoft utilizes the Fairwater architecture that manages high-throughput racks. The sites integrate Nvidia GB200 NVL72 units, designed for massive clusters based on Blackwell chips.
Liquid cooling transports heat outside the building before returning a cooler fluid. The company assures that this system consumes very little new water, aside from occasional replacements for chemical control.
The Atlanta site exactly mirrors the configuration of Wisconsin to replicate this recipe across multiple regions.
