While the sky-high demand for GPUs has made NVIDIA the most valuable company in the world, many of these GPUs simply aren’t being used.
In a remarkably candid admission, Microsoft CEO Satya Nadella revealed that the company’s biggest bottleneck in scaling its AI infrastructure isn’t the availability of chips—it’s power and construction timelines. Speaking about the current state of Microsoft’s AI buildout, Nadella painted a picture of a supply chain crisis that few outside the industry have fully grasped: advanced AI chips sitting idle in warehouses, unable to be deployed because the infrastructure to power them doesn’t exist yet.
“The biggest issue we are now having is not a compute bottleneck, but it’s power and it’s the ability to get the builds done fast enough close to power,” Nadella explained. “So if you can’t do that, you may actually have a bunch of chips sitting in inventory that I can’t plug in. In fact, that is my problem today, right? It’s not a supply issue of chips. It’s actually the fact that I don’t have warm shelves to plug into. And so that’s how some supply chain constraints emerge.”
The Microsoft chief’s comments underscore a dramatic shift in the AI infrastructure race. For the past two years, the narrative has centered on an acute shortage of GPUs, with companies scrambling to secure allocations from NVIDIA and other chipmakers. But Nadella’s remarks suggest the industry has moved past that phase into a new, equally challenging era: the infrastructure deficit.
The implications of Nadella’s admission are significant for the broader AI industry. Microsoft has committed tens of billions of dollars to building out AI infrastructure, including a landmark partnership with OpenAI and data center projects worldwide. Yet even with that level of investment, the company is hitting fundamental constraints around electrical capacity and construction speed—issues that can’t be resolved simply by writing larger checks.
This power crunch is becoming a defining challenge across the tech industry. Data centers running AI workloads consume exponentially more electricity than traditional computing facilities, with a single AI training cluster potentially requiring as much power as a small city. Major tech companies have begun signing power purchase agreements with utilities, exploring nuclear energy options, and even delaying data center projects in regions where grid capacity is insufficient. Google and Amazon have both announced investments in nuclear power to meet their AI infrastructure needs, while Meta has faced similar constraints in expanding its AI training capabilities.
Nadella’s comments suggest that the race to dominate AI may ultimately be won not by whoever secures the most chips, but by whoever can most effectively navigate the complex challenges of power infrastructure, regulatory approvals, and construction logistics—a far cry from the purely technological competition many observers expected.