Microsoft Prepares to Install Gigawatts of Its Own Chips in Data Centers

Microsoft is preparing to install the newest generation of its own chips in its data centers. The scale is no longer counted in cards and racks but in gigawatts of power draw. According to Bloomberg, the company's goal is to run its AI services mostly on hardware it designed itself. The available excerpt of the report does not make clear the exact chip name, the schedule or the sites. But the unit of measurement alone shows how serious the company is.

Why gigawatts

Gigawatts are the language of AI infrastructure today. Growth is limited not only by the number of available accelerators, but also by the electricity delivered to a campus, the cooling and the network fabric. By our rough estimate, one gigawatt, counting all the supporting infrastructure, corresponds to roughly half a million accelerators of the current class. That is the editors' estimate, not a Microsoft figure, and the real number depends on the chip's power consumption and the efficiency of the site.

So the phrase "gigawatts of its own chips" means one thing: Microsoft is planning not pilot clusters for internal experiments, but industrial-scale operation.

What is known and what is not

• Stated goal: to power Microsoft's AI software mostly with its own hardware.
• Scale: gigawatts of installed capacity.
• What is being installed: the latest version of chips designed by Microsoft itself.
• Not disclosed: process node, HBM memory volume and type, number of racks, timeline, regions.
• Not disclosed: what share of purchases Nvidia and AMD will keep. The word "mostly" does not mean abandoning them entirely.

Some background. Microsoft showed the Maia 100 accelerator and the Arm-based Cobalt 100 processor in November 2023. At the time, it looked like groundwork for the future: most training and inference in Azure ran on Nvidia GPUs, and Microsoft remained one of their largest buyers. Satya Nadella later said the company has access to OpenAI's system-level intellectual property in chips. For a company designing its own silicon, that is a serious help.

Why Microsoft wants this

There are three reasons, and all of them are about money and control.

First, cost. Every token that Copilot and services built on OpenAI models produce is paid for with an accelerator, and the supplier's markup is built into that price. A custom chip tuned for specific models and inference can lower the cost per unit of compute.

Second, supply predictability. The queue for accelerators and HBM memory determines when Microsoft can bring new capacity online. A custom design does not remove the dependence on TSMC and memory makers, but it reduces dependence on a single vendor.

Third, competition with other clouds. Google has been building TPUs for years, Amazon is developing Trainium, and Meta makes MTIA. Microsoft is catching up, and it cannot do that without mass deployment of its own chip.

The strategy has weak points. The main one is the software stack: a huge ecosystem has grown up around CUDA, and moving workloads to a different architecture takes compilers, libraries and engineers. The second problem is power: gigawatts have to be sourced somewhere and delivered to the site.

What to watch next

For Azure customers, the main question is when instances on Microsoft's chips will appear and how that will affect inference pricing. If the savings are real, they will show up in model pricing and in terms for large customers.

Developers should watch for PyTorch and Triton support on the new hardware. The fewer changes a port requires, the faster workloads will move off Nvidia.

For investors, the capital expenditure line and the structure of purchases matter more. If Nvidia's share of new contracts starts to fall, that will back up Bloomberg's report with numbers. For now, the key point is known: Microsoft has announced its intention to move to its own hardware, and it has chosen a scale that requires power plants.