Microsoft's cloud has always sold the same promise as every other hyperscaler: capacity on demand, whenever an enterprise needs it. According to reporting published the weekend of July 26, that promise now comes with an asterisk. When compute runs short inside Microsoft — and executives say it runs short constantly — the company's own AI products get served first, and Azure customers get whatever is left.
The account, first detailed by Business Insider's Ashley Stewart in a July 26 investigation of what she called Satya Nadella's "hardest year," describes a company that cannot build data-center capacity fast enough and has been forced into triage. First-party products such as Microsoft 365 Copilot and GitHub Copilot, along with frontier-lab training workloads, eat first. Everything else waits.
What the reporting says
The prioritization is not, strictly, a secret — Microsoft has described the mechanics of it on the record, even if it has never framed it as choosing itself over paying customers. On the company's January earnings call, Chief Financial Officer Amy Hood laid out the pecking order plainly. Microsoft solves first for M365 Copilot and GitHub Copilot, then for internal research and development.
"Then what you end up with is the remainder going towards serving the Azure capacity that continues to grow in terms of demand," Hood said. Had those same chips been pointed at Azure instead, she added, the segment's growth would have topped 40% rather than the 39% it reported.
What is new in the latest reporting is that insiders say the squeeze has intensified since. "All of the supply is gone once you solve for frontier labs and our internal businesses like M365 and Microsoft AI," one executive told Business Insider. Microsoft has been buying capacity from rivals to plug the gap — Amazon reportedly bailed it out after a run of GitHub outages, and the company has evaluated leasing from Oracle, Amazon and Google. "We are shopping for capacity everywhere," one person familiar with the talks said.
Microsoft has not officially confirmed a formal policy of deprioritizing Azure customers, and the sharpest characterizations come from unnamed executives rather than company statements. But the underlying capacity crunch is well documented. "We have been short power and space," Hood told analysts on an earlier call, when Microsoft was carrying close to $300 billion in cloud contracts it had not yet been able to fulfill.
The trust problem
The awkward part is what Microsoft is doing at the same time: raising quotas for its Azure salespeople. Some quotas are up roughly 30% this year even as the underlying supply tightens, according to people familiar with the change — meaning the field is being pushed to sell more of a resource the company is struggling to deliver.
Inside Microsoft, the logic is understood even where the messaging is not. One executive reportedly framed the trade-off bluntly: why would Nadella prioritize growing Adobe, an Azure customer, over growing Microsoft's own M365? "I have no idea how we're going to land that message with customers," the person added.
That is the strategic vise. Serving Azure customers lifts revenue now and protects the share price, which is already down about 19% this year — the worst performance among the megacap tech names. Serving Microsoft's own AI products is a longer bet that those products eventually win. Starve the first and cloud growth disappoints. Starve the second and Microsoft falls further behind in the race that justified spending roughly $190 billion on infrastructure this year in the first place.
An industry running on empty
Microsoft's rationing is the most pointed example of a shortage now visible across the entire AI economy. In late June, Google told Meta it could not sell it all the Gemini compute Meta wanted, capping a major customer despite a reported backlog worth hundreds of billions — and this from a company spending more than $180 billion on capex this year and paying SpaceX roughly $920 million a month for access to about 110,000 Nvidia GPUs. Anthropic signed its own SpaceX arrangement earlier in the spring. Nvidia, meanwhile, is effectively rationing its own chips among buyers, and has been in talks to help backstop hundreds of billions in OpenAI data-center financing.
The through-line is that model quality is no longer the binding constraint. Power, land, and the physical supply of accelerators are. Even the best-capitalized companies on earth cannot conjure a data center faster than the grid and the fab lines allow.
For enterprises, the lesson punctures a core assumption of the cloud era: that capacity is a utility you can draw on at will. When the provider is also your competitor for that capacity — and has told its own field to keep selling — the guarantee of elastic, on-demand compute starts to look conditional. Expect more customers to pursue multi-cloud commitments, reserved-capacity contracts, and written service-level terms rather than assuming a region will simply have room when they need it.
Nadella, for his part, was busy the same weekend reframing the whole boom as a test. Asked on CNN whether AI is a bubble, he did not deny it. "Unless we see that broad economic growth," he said, "we're not going to have this movie end well."
What to watch
Microsoft reports fiscal fourth-quarter results on Wednesday, July 29, with Amazon following on Thursday. Watch three things: whether Azure growth reaccelerates or stays capacity-capped; whether Hood signals when supply finally catches demand (past guidance has repeatedly slipped); and whether any named enterprise customer goes public about being throttled. The first customer to say out loud that Azure made it wait will do more damage to the on-demand promise than any anonymous executive quote.
"All of the supply is gone once you solve for frontier labs and our internal businesses like M365 and Microsoft AI."— Unnamed Microsoft executive, Speaking to Business Insider