Azure capacity: Copilot gets the GPUs before you do

Microsoft reports fiscal Q4 results on Wednesday. Before the number lands, it's worth re-reading what the company already told investors last quarter about Azure capacity — because it's the rare vendor disclosure that says out loud where you sit in line. Short version: when GPUs come online, Copilot and internal R&D get first claim, and Azure customers get what's left.

What actually happened

On the FY26 Q3 earnings call, Microsoft said "broad and growing customer demand continues to exceed supply" and that it expects "to remain constrained at least through 2026." The allocation logic was stated plainly: incoming supply gets balanced "against our other high ROI priorities: first party applications, R&D, and end of life server replacement." CFO Amy Hood added that she'd "expect the pressure between first-party usage and being able to meet Azure demand will persist."

The demand side explains why. Microsoft 365 Copilot passed 20 million paid seats, up 250% year over year, with the number of customers running more than 50,000 seats quadrupling. Those seats consume inference in the same data centers your VMs run in.

The spend side is the part that doesn't fix this quickly. Microsoft expects roughly $190 billion in calendar-2026 capex — including about $25 billion attributable purely to higher component pricing — with over $40 billion in the fiscal fourth quarter alone. That's a company spending at record scale and still telling customers to expect a queue. Q4 results land July 29.

Why cloud capacity constraints matter for your business

Most small operators read "capacity constrained" as a hyperscaler problem. It isn't. It shows up in your world as a specific SKU unavailable in your region, a quota request that sits for weeks, or a reserved-instance renewal that quietly moves you to a different instance family. If your automation depends on one GPU SKU in one region, you have a single point of failure that no amount of uptime SLA covers — allocation isn't downtime, so it doesn't breach anything.

Three things to actually do. Know which of your workloads need a GPU and which just need an API. Most don't need dedicated silicon; they need a model endpoint, and endpoints are far easier to re-point than instances. Then make the endpoint a config value, not a hardcoded client — base URL, model name, and key in environment variables, with one adapter in front. That's a couple hours of work that turns a capacity problem into a fifteen-minute switch.

And stop treating your cloud vendor as neutral. Microsoft sells Copilot and sells you the compute Copilot runs on. When those compete, the disclosure above tells you which one wins. That's not a betrayal — it's a rational allocation of a scarce asset, published in an earnings call. Just build like you read it.

Sources: Microsoft FY26 Q3 earnings call, Microsoft FY26 Q4 earnings event.