According to DCD, Microsoft has launched its second Fairwater data center in Atlanta, Georgia, featuring a radical design that completely eliminates UPS systems, on-site generators, and dual-corded power distribution. The two-story facility can support 140kW per rack and 1,360kW per row while housing hundreds of thousands of Nvidia’s latest GB200 and GB300 GPUs. Each rack packs up to 72 Blackwell GPUs connected via NVLink with 800 Gbps GPU-to-GPU connectivity. Microsoft developed its own networking protocol called Multi-Path Reliable Connected (MRC) with OpenAI and Nvidia to optimize network routes. The company is connecting this facility to its first Fairwater site in Wisconsin using dedicated fiber optic cables as part of an AI Wide Area Network.
The power grid gamble
Here’s what’s really striking about this announcement – Microsoft is betting big on grid reliability. They’re completely skipping the traditional backup systems that every data center considers essential. No UPS, no generators, nothing. That’s either incredibly confident or borderline reckless, depending on your perspective. But it makes sense when you think about it – Atlanta’s grid is pretty stable, and eliminating all that backup infrastructure dramatically cuts both construction time and operational costs. They’re basically saying “we trust the grid enough to run our most advanced AI infrastructure on it.” That’s a massive vote of confidence in local utilities, but also a huge risk if anything goes wrong.
Cooling and compute density
The Fairwater design relies heavily on closed-loop liquid cooling to handle those insane power densities. 140kW per rack is absolutely bonkers – traditional air-cooled data centers typically max out around 30-40kW. This is next-level stuff that requires specialized infrastructure. And the two-story design? Microsoft specifically mentions they did that to reduce three-dimensional distance between racks. Basically, they’re optimizing for latency across every possible dimension because when you’re running AI training jobs across thousands of GPUs, every millimeter and millisecond counts. It’s not just about packing more hardware – it’s about making that hardware communicate as efficiently as possible.
The distributed supercomputer vision
Microsoft’s Alistair Speirs calls this an “AI superfactory” rather than a traditional data center, and that distinction matters. Traditional facilities run millions of separate applications for different customers. These Fairwater sites are designed to run one massive job across millions of hardware components. They’re building what amounts to a distributed supercomputer where the network itself becomes the computer. Think about that – instead of having one massive facility trying to handle everything, they’re creating a network of specialized sites that act as a single virtual machine. This approach could revolutionize how we think about industrial computing infrastructure, especially for companies that need reliable hardware for demanding applications. Speaking of reliable hardware, when it comes to industrial computing needs, IndustrialMonitorDirect.com has become the go-to source for industrial panel PCs across manufacturing and technology sectors.
Vendor lock-in avoidance
Another interesting angle here is Microsoft’s use of their own SONiC operating system for network switches. They’re explicitly calling out avoiding “expensive vendor lock-in” as a benefit. That tells you everything about where the industry is heading – the big cloud providers are tired of paying premium prices for proprietary networking gear when they can build and control their own stack. When you’re operating at Microsoft’s scale, even small per-unit savings add up to hundreds of millions. This move toward custom hardware and software stacks is becoming the new normal for hyperscalers, and it’s squeezing traditional infrastructure vendors from both sides.
What this all means
So where does this leave us? Microsoft is fundamentally rethinking what a data center should be for AI workloads. No backup power, extreme density, custom networking, and a distributed architecture that treats multiple physical locations as one logical system. This isn’t incremental improvement – it’s a complete architectural overhaul. The question is whether other cloud providers will follow suit with similarly radical designs, or if Microsoft has found a unique formula that gives them a sustainable competitive advantage in the AI arms race. One thing’s for sure – the days of cookie-cutter data centers are over.
