Digging Into Microsoft's Datacenter Shortage Issue

News broke early October that Microsoft was experiencing shortages of datacenter capacity to meet ever-increasing demand. The story, broken by Bloomberg (paywalled), went into the how and why, and I'm going to elaborate on the events surrounding it.

The news is that Microsoft is experiencing a datacenter crunch in capacity nationwide. It is experiencing shortages of both physical space and servers in setting up facilities or expanding existing ones. New subscriptions for Azure cloud services are restricted in some key geographic locations, including Northern Virginia and Texas, through the first half of next year.

Microsoft had previously warned of shortages in its last earnings call, where CFO Amy Hood said constraints would continue through the end of 2025. Now it's being pushed back into the middle of next year. This comes just months after Microsoft said it was tapping the brakes on datacenter construction.

The problem is not exclusive to Microsoft. Google and Amazon have also stated that they are having capacity issues.

A Microsoft spokesperson told Bloomberg said that a majority of Azure services and regions in the U.S. "have available capacity so that existing customers with deployed workloads can continue to grow." If there are demand spikes, the company has what it calls "capacity preservation methods," or what the rest of us would call "throttling."

The problem is a perfect storm of events. There's not one cause, there are multiple causes all happening at once.

• Supply chain shortage: this is affecting everybody, including Google and Amazon. Hardware makers just can't make enough equipment for all the demand.
• Tariffs: tariffs on technological products are causing chaos in the supply chain.
• NIMBYs: There is a growing backlash across the country to datacenter expansion. Datacenters take up a lot of space, they use a lot of water and power, and in some cases they have a polluting effect on the nearby area. So datacenter companies are finding it much harder to get land and they are facing increasing resistance from locals, slowing the process down.
• ChatGPT: OpenAI Doesn't have datacenters, Microsoft does. So this alliance is fully dependent on Microsoft datacenters to provide ChatGPT services.

One of the significant new challenges though, is what I refer to as dual builds. There is a massive race to build AI datacenters for ChatGPT, and NVIDIA is benefiting mightily from it. But while everybody, including Microsoft, is building out AI datacenters, they still have to build standard datacenters for things like Azure services and generic services like IaaS.

So you have a situation where hyperscale cloud giants like Microsoft, Amazon, Google and Meta are building out datacenters the size of football fields, both AI and non-AI,  with tens (if not hundreds) of thousands of servers in them, all competing for limited supply.

And mind you, hyperscalers don't use servers from the big three server vendors -- HPE, Dell and Lenovo. They use what are known as white box servers, frequently made in China, which again I will remind you is subject to hefty tariffs.

So you have Microsoft attempting to build out both standard datacenters and AI datacenters at the same time when there is a parts shortage and trade wars going on. Not an enviable position to be in, but it could be worse.

This is likely to be an issue on larger scale customers because they place greater demands on the system. Small to medium-sized businesses are not going to demand as much resources as a Fortune-500 firm.

Microsoft has been on a notorious spending spree building datacenter capacity as quickly as possible. Satya Nadella recently said he's going to spend $85 billion this year on datacenters. But even that isn't enough.
The fact is, datacenters take a long time to build. You're not just talking about putting up four walls and a ceiling and loading it with hardware anymore. There has to be very careful power management and water management because datacenters use an incredible amount of water for cooling, especially AI datacenters, which run extremely hot.

The amount of time to build the datacenter has crept up due to the increasing complexity, according to the Uptime Institute. What once took two years to do now takes more than three years. At least some of that is due to NIMBY pressure and reluctance by local authorities to grant licenses and permits. The licensing and permits part of the datacenter build is becoming increasingly long, longer than the build out itself.

The irony is that people are now talking about how datacenters take so long to build, by the time they go online, they're already obsolete. AI hardware is leaping forward with each generation, and each generation is coming in about one year's time from the last. So with NVIDIA and AMD cranking out new AI chips every year, three years to build the datacenter means it will go live with hardware that may not even be supported anymore.

Not that the hardware is useless. Far from it. It's just not top of the line anymore because the datacenter build took so long.

So for now, this is primarily an issue for Microsoft's largest customers. Individuals and SMBs are less likely to feel the pinch, and if they are, it could just mean a slower response on a ChatGPT query.

All in all, it's a good problem to have.

Posted by Andy Patrizio on 11/03/2025