Infrastructure

Copilot Is Served Before Azure Customers

Microsoft's finance chief told analysts Azure would have grown past 40 percent if new GPUs had gone to customers rather than first-party Copilot. Capacity is now a physical limit, not a commercial one.

InfrastructureBy Servola Tech Desk2026-07-284 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

Copilot Is Served Before Azure Customers

Key takeaways

Microsoft's finance chief told analysts that Azure growth would have exceeded 40 percent if the accelerators deployed in the first half of the fiscal year had gone to customers instead of M365 and GitHub Copilot. When a cloud vendor publishes the size of the demand it chose not to serve, that number is a measurement of your place in the queue.
Microsoft executives have said the company is short enough of compute that first-party Copilot is served ahead of Azure customers, while cloud sales quotas were raised 30 percent. The sales force is being paid to sell capacity the company has told investors it cannot fully deliver.
Satya Nadella has said the binding constraint is no longer a shortage of chips but power, and the ability to finish builds close to it. A power constraint does not yield to a larger commitment or a better renewal rate.
Before the next renewal, get the delivery term in writing per region and per accelerator family, with a date and a stated remedy. Microsoft reports fiscal fourth-quarter results on 29 July.

The counterfactual a finance chief volunteered

On a Microsoft earnings call, Amy Hood did something companies almost never do: she told analysts what Azure's growth would have been if the company had made a different internal choice. Microsoft was solving first, she said, "for the increased usage in sales and the accelerating pace of M365 Copilot, as well as GitHub Copilot, our first-party apps", and allocating what remained to Azure customers. Had the accelerators brought online in the first half of the fiscal year gone entirely to Azure, she added, the reported growth rate would have cleared 40 percent rather than the high-thirties figure the company printed.

Why it matters. Rationing is normally invisible. A vendor short of supply serves who it serves and says nothing about the rest, and the customers who waited never learn how long the line was. Hood quantified it, on the record, to investors. The gap between the printed number and the counterfactual is not an analyst estimate. It is the company's own measurement of demand it had the customers for and not the capacity to serve. She also told analysts Microsoft expected to remain capacity constrained through at least the end of the fiscal year.

This sits on top of a capital programme of roughly 190 billion dollars this year, most of it data-centre capacity for AI workloads. Spending at that scale and still rationing is the part that should register with any buyer. The shortage is not a budgeting failure that a bigger cheque next year will correct.

Quotas went up while supply was being rationed

Microsoft executives told Business Insider in late July that the company is short enough of compute that Copilot is served before Azure customers, and that cloud sales quotas had been raised 30 percent over the same period. Those two statements were made to different audiences. One went to investors, the other to the field.

For a buyer the practical consequence is straightforward: the person selling you Azure capacity has a number to hit and no control over the constraint that decides whether you receive it. That is a structural mismatch rather than bad faith, and it is worth naming because of what it does to the value of assurances. A verbal commitment on availability from an account team is worth less this year than it was last year. The assurance worth having is a written delivery term tied to a named region and a date.

The ordering itself is the tell. First-party applications are the highest-margin use of a scarce input, and they also produce the usage data that improves the product. A customer workload does neither for the vendor. For as long as the input stays scarce, serving Copilot first is the rational choice, which is precisely why it will persist and why no amount of escalation will reverse it.

Power is what your renewal is competing against

Nadella has been direct about where the limit now sits. "The biggest issue we are now having is not a compute glut, but it's power", he said, pointing to the difficulty of getting builds finished fast enough close to generation. Reporting across the sector has described accelerators sitting in inventory for want of the electrical capacity to install them. The bottleneck moved from the fab to the substation, and it moved quietly.

That changes what negotiation can achieve. A commercial constraint yields to price, volume or contract term. A grid connection does not. In Europe the effect is sharper than in the United States, because the regions European buyers are steered towards sit in exactly the metropolitan clusters where connection queues are longest. The question for a European buyer is no longer only where data may legally sit, but whether the region chosen can be energised on the timetable the business is planning against.

What to settle before the next renewal. Get the delivery commitment in writing per region and per accelerator family, with a date and a remedy if it is missed, because service credits for downtime do not cover capacity that never arrives. Ask what happens to your allocation if the region is constrained, and get the answer in the contract rather than the email. Price what a two-quarter slip costs your own roadmap, since that is the exposure you are carrying whether or not you have quantified it. Then run one real workload on a second provider or on portable infrastructure this year, so that the answer to a capacity refusal is a migration rather than a conversation. Microsoft reports fiscal fourth-quarter results on 29 July, and the number to listen for is not the Azure growth rate. It is whether the capacity-constrained language is extended into the new fiscal year.

Frequently asked questions

Does this mean Azure cannot take on new workloads?

No. It means allocation is being actively managed, and that first-party products sit ahead of customer workloads within it. Microsoft has continued to sign and serve customers throughout. The constraint shows up as slower access to particular accelerator types in particular regions rather than as an outright refusal, so the risk you are carrying is one of timing rather than of availability in principle.

Is this specific to Microsoft?

The ordering is visible at Microsoft because its finance chief quantified it. The underlying constraint is industry-wide. Every large provider is building against the same shortages of power and grid connection, and every one of them also runs first-party AI products that compete with customer workloads for the same accelerators. The reasonable assumption is that the same priority ordering applies elsewhere unless a contract states otherwise.

What should actually be in the contract?

A delivery term naming the region, the accelerator family and a date, plus a stated remedy if the capacity does not appear. Downtime credits are the wrong instrument here, because they compensate for a service that failed rather than for capacity that never arrived. If a provider will not commit in writing to a region and a date, that refusal is itself a clear answer about how firm your allocation is.

The useful thing Microsoft did was tell the truth in a measurable form. A company that publishes the gap between what it sold and what it could serve hands every customer the information needed to plan around it. The mistake would be to file this as a Microsoft problem. Power and grid connection now sit between every cloud commitment and the hardware behind it, and the buyers who come through the next two years intact will be the ones who treated capacity as a physical supply chain with lead times, rather than as a line item that arrives because it was purchased.

MicrosoftAzureCopilotAmy HoodAI capacitycloud procurementgrid connectionvendor concentration

2028 Is an Executive Order, Not a Product Date

Antares raised $470m after its Mark-0 reactor went critical in Idaho. The 2028 deployment date is set by an executive order and a customer no European operator can be.

5 min read

Energy

The Money Buys Test Rigs, Not Permission

A BAE Systems spin-out raised 15 million pounds to put a reactor behind an industrial meter. Its own stated use of funds stops short of licensing, and that is the honest timeline.

6 min read

Law & Regulation

The EA Buyout Now Hinges on Brussels

The largest take-private in history has cleared EA shareholders and US antitrust. The last real gate is the EU - and not classic merger review, but the Foreign Subsidies Regulation, facing its first blockbuster test.

3 min read

← All articles