Microsoft's Azure Compute Rationing, Explained Before Earnings

Two days before it reports fiscal 2026 fourth-quarter earnings, Microsoft is still telling anyone who will listen that it cannot build data centers fast enough. Not "won't." Cannot. The company has spent much of the past nine months disclosing, quarter after quarter, that demand for Azure and its own AI products has outrun the physical capacity it has to serve it, and that it is making explicit choices about who gets compute and who waits. That is the backdrop for July 29, 2026, when Microsoft reports results for the quarter ended June 30. This piece looks at what Microsoft has actually said about rationing Azure compute, why a company spending an estimated $190 billion this year still cannot keep up, and what a capacity-constrained hyperscaler means for the enterprises that depend on it. It does not report on the July 29 results themselves, which had not been released as of this writing.

Microsoft

Azure

OpenAI

AWS

Google Cloud

NVIDIA

Microsoft's compute crunch touches its own AI products, its Azure customers, and now its two biggest cloud rivals.

What Microsoft has actually disclosed

The core disclosure is not speculation. Over three consecutive quarterly earnings calls in fiscal 2026, Microsoft executives have said, in some form, that demand for Azure and Copilot exceeds available capacity. Reporting on the company's fiscal third-quarter results (reported April 29, 2026) described Microsoft defending its rising capital spending by pointing out that "demand for Azure, Copilot, and AI infrastructure continues to exceed available capacity," according to DigiTimes.

That is not a new problem. On the fiscal first-quarter call on October 30, 2025, CFO Amy Hood was asked directly by a Deutsche Bank analyst how much higher Azure revenue could have been without the shortage. Her answer: "It is safe to say that the number could be higher." Pressed on which part of the business absorbs the shortfall, she said "Azure probably does bear most of the revenue impact." Azure still grew 40% year over year that quarter despite the constraint, a detail that says more about underlying demand than about Microsoft's ability to meet it.

$190BMicrosoft's 2026 capex guidanceCalendar-year data center and AI infrastructure spend

$625BCommercial bookings backlog (RPO)Disclosed at fiscal Q2 2026 earnings, Jan. 2026

45%OpenAI's share of that backlogAbout $281B tied to a single customer

40%Azure revenue growth, fiscal Q3 2026Reported April 29, 2026, above analyst estimates

Jul 29, 2026Next earnings dateFiscal Q4 2026, not yet reported

The priority order Microsoft has described

When Hood and CEO Satya Nadella have been asked how scarce compute gets allocated, the answer has been consistent: internal AI products and the largest customer commitments come first, and general Azure demand is served with what is left over. On the fiscal Q1 call, Hood said internal AI products, Microsoft 365 Copilot, GitHub Copilot, and the company's security Copilots, get priority because "that is where, and it is a priority for us, to allocate resourcing there first." On the following quarter's call, in January 2026, she described the order more bluntly: internal workloads and major commitments are served first, and "then what you end up with is the remainder going towards serving the Azure capacity," as reported by The Next Web.

Aerial view of Microsoft's Redmond, Washington headquarters campus — Microsoft's Redmond, Washington headquarters campus, photographed in 2009. Capacity-allocation decisions described on recent earnings calls are made here and at Microsoft's data center sites worldwide, not at this specific location. Photo by Jelson25 via Wikimedia Commons, public domain.

The scale of the internal draw is not small. Microsoft's own AI business had surpassed a $37 billion annual revenue run rate as of the April 2026 report, up 123% year over year, and that is before counting the compute Microsoft dedicates to Microsoft 365 Copilot seats, GitHub Copilot, and internal research. Meanwhile the company's commercial backlog, its remaining performance obligations, reached $625 billion as of the January 2026 earnings call, up 110% year over year, with OpenAI alone accounting for roughly 45% of that figure, or about $281 billion, according to Business Insider reporting relayed by Yahoo Finance. Every dollar of that backlog eventually has to be served by a GPU somewhere.

Renting from the rivals

The clearest sign of how tight the situation has gotten is that Microsoft has reportedly started buying capacity from the companies it competes against. Business Insider reported in June 2026 that Microsoft added Amazon Web Services capacity for GitHub after a string of AI-driven reliability incidents there, and that it had separately evaluated capacity from Google Cloud as it looked for compute wherever it could find it. GitHub's own numbers explain some of the pressure: agent-opened pull requests rose from about 4 million in September 2025 to more than 17 million by March 2026, and GitHub's chief operating officer said commit volume was on pace to hit roughly 14 billion for 2026, up from about 1 billion in all of 2025, according to Tech Times' coverage of the incident.

We are shopping for capacity everywhere.
Microsoft executive, quoted by Business Insider

A Microsoft spokesperson framed the AWS arrangement as tactical rather than strategic, saying the company was "both accelerating our move to Azure and continuing to explore a multi-cloud strategy," per Windows Central's summary of the reporting. That framing matters, because Microsoft has not abandoned Azure-first infrastructure; GitHub was already migrating to Azure and remains on track to finish that migration. But an internal executive quoted by Business Insider put the underlying math more starkly: "All of the supply is gone once you solve for frontier labs and our internal businesses like M365 and Microsoft AI." Reporting also indicated Microsoft explored, then walked away from, leasing Oracle infrastructure over compliance concerns, underscoring how many options it is working through at once.

Why compute scarcity happens even with $190 billion of spending

It is worth being precise about why more money does not immediately fix this. Microsoft's calendar-2026 capital spending is guided at roughly $190 billion, up 61% from 2025, and is tracking toward the top of a group of four hyperscalers spending a combined $725 billion this year. That spending buys land, buildings, power infrastructure, networking, and chips, but each of those has its own lead time that money alone cannot compress. Grid interconnection and substation buildouts can take years to permit and construct. Advanced GPU and HBM memory supply is capacity-constrained industry-wide, not just for Microsoft, since every hyperscaler is competing for the same TSMC packaging slots and memory fabs. And commissioning a data center, wiring it for the power density modern AI racks require, and validating it for production is a multi-quarter process even once the building exists.

Microsoft is not alone in outspending its own buildout

2026 calendar-year capital-spending guidance for the four largest hyperscalers. Ranges shown are midpoints; see tooltip for the full guided range.

Source: Yahoo Finance hyperscaler capex roundup, June 3, 2026. Combined 2026 guidance across the four companies totals roughly $725 billion, up about 77% from 2025.

That is why Microsoft's own guidance has repeatedly framed the constraint as durable rather than a one-quarter blip: the company has said it expects capacity constraints to persist at least through the end of fiscal 2026 (its fiscal year ends June 30), even as it adds roughly a gigawatt of new capacity per quarter and works toward doubling its global data center footprint within two years. The result, visible in Azure's own numbers, is growth that has stayed remarkably steady in the high 30s to 40% range for four straight quarters rather than accelerating, even as demand indicators like the RPO backlog have more than doubled.

Azure growth has held near 40%, capacity limits and all

Azure year-over-year revenue growth by Microsoft fiscal-2026 quarter. The last bar is Microsoft's own guidance for the quarter it reports on July 29, 2026, not an actual result.

* Q4 FY26 is Microsoft's guidance range, given ahead of the July 29, 2026 earnings report, not a reported figure.

What rationing means for enterprise buyers

For a large enterprise with an existing Enterprise Agreement or a major capacity commitment, the practical effect of this has reportedly been manageable, if noticeably tighter provisioning windows. For smaller and mid-size customers trying to reserve GPU quota for model training or fine-tuning, the effect has been more direct: quota denials, waitlists, and in some cases multi-month waits for specific GPU SKUs in specific regions, with certain Azure regions effectively full for some instance families. At the same time, Microsoft has reportedly raised Azure sales quotas by roughly 30% even while acknowledging it cannot fully supply existing demand, a tension The Next Web highlighted between growth targets set by the sales organization and the delivery capability of the infrastructure organization.

That tension is the real strategic story for buyers to watch, separate from any single quarter's revenue number. A cloud provider that is simultaneously the operator of the marketplace, a first-party seller of AI products on that marketplace, and short on the underlying resource has an obvious incentive conflict: every GPU-hour that goes to a Copilot seat or an OpenAI training run is an hour not sold to an external Azure customer, including customers building products that compete with Copilot. Microsoft has been transparent that this trade-off exists and that its own products are winning it. Whether that allocation reflects strict revenue-maximizing logic, strategic bets on its own AI products, or straightforward operational necessity during a scarce period is something Microsoft has not fully separated in its public disclosures, and reasonable analysts read it differently.

Portability as a hedge, not a workaround

None of this is unique to Microsoft. Every major cloud and model provider is capacity-constrained to varying degrees in 2026, and the specific provider that is tightest changes by quarter and by region. That is the practical argument for not routing all AI work through a single vendor's stack by default. A team whose workflows depend on one model or one cloud inherits that provider's capacity cycle as its own; when the vendor prioritizes its own roadmap, the customer has no lever to pull. Model-agnostic tooling does not solve a data center shortage, but it does mean a team is not stuck waiting on one company's GPU allocation decisions when an alternative exists. That portability logic is the same reason Metir AI is built to give teams access to leading models across providers in one workspace rather than tying a workflow to a single lab or cloud, so a capacity crunch at any one provider is an inconvenience rather than a blocker.

What to watch for on July 29

Since Microsoft has not yet reported fiscal Q4 2026 results as of this writing, the useful frame is what to watch rather than what happened. Analysts have modeled revenue of roughly $87.7 billion for the quarter, toward the top of Microsoft's own guided range of $86.7 billion to $87.8 billion, with Azure growth guided at 39% to 40%, according to GeekWire's earnings preview. The more informative numbers will likely be qualitative: whether Microsoft says the capacity constraint is easing, holding steady, or getting worse; whether capital spending guidance for the next fiscal year moves up again; and whether management offers any update on the RPO backlog and OpenAI's share of it. CEO Satya Nadella, speaking on CNN two days before earnings, framed the broader AI buildout as needing to show real economic payoff, saying "if we don't see that, we're not going to have this movie end well." Whether Azure's growth number on July 29 reads as vindication of the capex or as a sign the constraint is finally biting into revenue is likely to be the central debate the market has that day.

Sources:

Image credits

Header image: exterior of a Microsoft data center in Middenmeer, the Netherlands, photographed in 2023, via Wikimedia Commons, photo by Hay Kranen, licensed under CC BY 4.0. In-body photograph: an aerial view of Microsoft's Redmond, Washington headquarters campus, photographed in 2009, via Wikimedia Commons, photo by Jelson25, public domain. Neither image depicts a specific 2026 capacity-allocation decision; both show real Microsoft-operated sites referenced in the reporting above.