Microsoft Plans to Double AI Capacity by 2028

Anyone who has tried to provision GPU capacity on a public cloud in the last 18 months knows the answer is usually "next quarter, maybe." Microsoft just told Wall Street it intends to double the size of that constraint by 2028, while burning through $31.9 billion of capex in a single thirteen week window. For platform leads betting on Azure as their AI substrate, this is the most important number in the quarter.

The doubling pledge from Satya Nadella sits on top of an installed base that already looks absurd on paper. The question is whether the unit economics hold up when two thirds of that capex has the shelf life of a mid-range laptop.

Key Details

The headline commitment, as The Next Platform reported, is that Microsoft added 1 gigawatt of capacity in the March quarter alone, and Nadella reiterated the company is on track to double total capacity (measured in gigawatts, not flops) over the next two years. The current footprint is already enormous: more than 80 Azure regions, over 500 datacenters, 190-plus point of presence locations, and over 800,000 kilometers of network fiber. The whole estate burns roughly 10 gigawatts of power.

Doubling that means Microsoft is signing up for another 10 gigawatts of compute, cooling, and grid interconnect by 2028. For context, that is the output of roughly ten nuclear reactors.

The financials underwriting this are loud. Q3 fiscal 2026 sales hit $82.89 billion, up 18.3 percent year on year. Operating income rose 20 percent to $38.4 billion. Net income climbed 23.1 percent to $31.78 billion, or 38.3 percent of revenue. Intelligent Cloud, which carries Azure, posted $34.68 billion in sales (up 29.6 percent) with $13.75 billion in operating income at a 39.7 percent margin. Microsoft Cloud as a whole did $54.55 billion in sales with $36 billion in gross profit. Productivity and Business Processes cleared just over $35 billion in sales with just under $21 billion in operating profit, and the enterprise app business inside it grew 16.9 percent.

On the silicon side, Cobalt Arm server CPUs are now in half of Azure regions, and the Braga Maia 200 XPU, revealed in late January, is live in Iowa and Arizona datacenters. Capex was $31.9 billion in the quarter, with about two thirds going to CPUs and GPUs (Microsoft's "short-lived assets") and the rest to datacenter shells with a fifteen year-plus lifespan. Cash on hand at quarter end: $78.27 billion.

Why This Matters for AI Development

The capex split is the line every CTO should stare at. Roughly $21 billion of one quarter's spending went to chips Microsoft itself classifies as short-lived. That is a depreciation schedule closer to consumer electronics than traditional infrastructure. Production incidents I've seen at fintechs that misjudged hardware refresh cycles usually start with someone assuming a five year life on gear that lasts three. Microsoft is being honest about the math, which is more than most.

The operational read is that Azure's AI capacity is being financed against the assumption that GPU generations turn over fast and the revenue from each generation has to clear the bill before the next one lands. At $31.9 billion a quarter, that is roughly $127 billion annualized if the pace holds. The 39.7 percent operating margin in Intelligent Cloud says, for now, the model works.

The other shift is strategic. Microsoft no longer has exclusive rights to OpenAI GPT models, the deal that started in September 2020 has loosened, and OpenAI is spending tens of billions to source its own hardware via Nvidia, AMD, and Cerebras. OpenAI still has substantial commitments to rent Azure capacity, so the revenue stream isn't gone, but the lock-in is. For teams building on Azure OpenAI Service, this is mostly good news. Microsoft now has a direct incentive to host competitive models alongside GPT, which means more optionality at the API layer. Engineers comparing Azure-hosted models to direct calls against the OpenAI API or Anthropic's docs will see the gap on features narrow, not widen.

My take: the unwinding of OpenAI exclusivity is the single most underrated piece of this quarter. It frees Microsoft to act like a neutral cloud again, and it forces Azure's AI roadmap to compete on infrastructure quality rather than model exclusivity.

Industry Impact

For iGaming, fintech, and ad-tech teams running real-time inference, the doubling of capacity is the practical takeaway. Rate limits and quota refusals on Azure OpenAI have been a recurring headache for teams I've worked with on European fintech rollouts, particularly around launch windows when traffic spikes coincide with global capacity squeezes. Another 10 gigawatts of capacity, even spread over two years, materially loosens that constraint.

The Cobalt and Maia 200 deployments matter for cost-sensitive workloads. Cobalt is now in half of Azure regions, which means Arm-based VM SKUs are no longer a curiosity, they are a procurement option. Teams running stateless backend services, message queues, or feature stores can probably trim 10 to 20 percent off compute bills by switching, based on patterns we've seen with Graviton on AWS. The Maia 200 in Iowa and Arizona is more limited, but it signals Microsoft is willing to compete with Nvidia on inference price per token in its own house.

The uncomfortable read: $31.9 billion of capex in one quarter is roughly the entire annual revenue of a top-five European bank. On a 200 person platform engineering team at a mid-market operator, the equivalent of one quarter of Microsoft's GPU capex would fund the entire payroll for several centuries. That kind of asymmetry means smaller players cannot rationally try to own the stack. They have to rent it, and they have to pick their cloud partner with the same care they used to pick a core banking vendor.

The 38.3 percent net margin tells you Microsoft has pricing power. Don't expect Azure AI list prices to fall sharply. Expect committed-use discounts and reservation contracts to be the lever instead.

What to Watch

Three signals will tell us whether the doubling pledge is real or aspirational. First, watch the quarterly capex figure. If it stays north of $30 billion for the next four quarters, the build-out is on schedule. If it dips and Microsoft starts talking about "capital efficiency," the timeline slips. Second, watch Maia 200 region expansion. Two datacenters today, Iowa and Arizona, is a pilot. Six to ten regions inside twelve months would mean Microsoft is serious about reducing Nvidia dependence. Stalling at two or three means the silicon isn't ready for general workloads.

Third, watch what Microsoft does with frontier models. The company built a hybrid Megatron-Turing model with Nvidia before the November 2021 GPT-3 API moment changed the game. With OpenAI exclusivity gone, Microsoft has both the cash ($78.27 billion in the bank) and the silicon to re-enter frontier model building. If a Microsoft-branded frontier model lands in the next twelve months, the Azure AI story stops being about hosting other people's models.

For platform teams: lock in capacity reservations now if you have visibility into 2027 demand. The doubling is coming, but so is the demand to absorb it.

Key Takeaways

• Microsoft added 1 gigawatt in Q3 fiscal 2026 and committed to doubling its roughly 10 gigawatt AI footprint by 2028.
• Q3 capex hit $31.9 billion with two thirds spent on CPUs and GPUs that Microsoft classifies as short-lived assets.
• Intelligent Cloud delivered $34.68 billion in sales at a 39.7 percent operating margin, funding the build-out without strain.
• Cobalt Arm CPUs cover half of Azure regions; Maia 200 XPUs are live in Iowa and Arizona, giving Microsoft real silicon optionality.
• OpenAI exclusivity is over, but rental commitments remain, freeing Microsoft to host competing models and rebuild its own frontier capability.