Microsoft’s AI infrastructure strategy is changing. After committing to a very large buildout of data center capacity, the company has slowed parts of that plan and is now putting more emphasis on flexibility, leasing, and GPU spending.
The shift matters beyond Microsoft. According to SemiAnalysis, Oracle, Crusoe, and CoreWeave are gaining ground as OpenAI moves more of its infrastructure needs outside Microsoft’s exclusive orbit.
The scale of Microsoft’s original buildout
Microsoft entered 2023 and 2024 with a major data center expansion agenda. SemiAnalysis says the company signed binding preliminary agreements for over 5 gigawatts of new capacity, with that capacity intended to come online between 2025 and 2028.
The plan was not limited to leases or single-site projects. Microsoft also acquired tens of thousands of acres of land worldwide, sped up construction at sites it already controlled, and arranged gigawatts of power capacity.
The financial footprint reflected the size of the push. In fiscal year 2024, Microsoft’s property, plant, and equipment book value increased by $26 billion. Its land holdings rose 44 percent year-over-year, while construction commitments tripled to more than $35 billion.
At the high point of this expansion cycle, SemiAnalysis said Microsoft represented over 60 percent of all new leasing contracts in the hyperscale segment. That made the company one of the central forces shaping AI data center demand.
Where the pullback is happening
The change began to show in mid-2024, when Microsoft stopped all new leasing activity. Reports pointed to 2 gigawatts in canceled leasing contracts, but SemiAnalysis says those were tied only to non-binding letters of intent.
The larger issue, according to the same analysis, is Microsoft’s freeze on 1.5 gigawatts of self-developed capacity planned for 2025 and 2026. That is a more direct signal about the company’s own construction pipeline.
Several large projects have made limited progress since then. Land development is still moving ahead, but building construction and orders for cooling and electrical infrastructure have been paused or pushed back in some cases.
This does not mean Microsoft is abandoning AI infrastructure. The company is still expected to expand total capacity in 2025 and 2026. The difference is in how that capacity is being pursued.
- Spending on GPUs is increasing significantly.
- New company-owned data center construction is expected to stagnate or decline slightly.
- Microsoft is moving more of its future capacity strategy toward leasing models.
OpenAI changed the demand picture
Microsoft’s earlier expansion was closely connected to OpenAI’s rapid growth after ChatGPT. OpenAI wanted to multiply its computing capacity, and Microsoft initially occupied a central role in that plan.
That relationship changed in 2025, when the exclusive partnership between Microsoft and OpenAI ended. OpenAI shifted much of its training infrastructure to Oracle and Crusoe, which are building a 1.2-gigawatt data center at the first "Stargate" campus in Texas.
OpenAI also signed a five-year, $11.9 billion contract with CoreWeave and invested $350 million in the provider. SemiAnalysis also notes that CoreWeave GPUs could be used in the future for OpenAI inference tasks, a role that had previously been handled exclusively by Microsoft.
For Microsoft, that changes the risk profile of building ahead of demand. If OpenAI is no longer relying exclusively on Microsoft infrastructure, Microsoft has less reason to keep every part of its earlier buildout moving at the same pace.
Microsoft’s own AI demand is another factor
The external shift with OpenAI is only one part of the story. SemiAnalysis also points to weaker-than-expected demand for Microsoft’s own AI products.
The analysis says Microsoft has lagged competitors such as Amazon in developing large language models internally. It also says anecdotal evidence and current order volumes suggest enterprise AI demand at Microsoft is not meeting earlier expectations.
That combination creates pressure to rethink capital allocation. Building company-owned data centers is a long and expensive commitment. If the demand curve is less certain, leasing and GPU-focused spending can give Microsoft more room to adjust.
The result is not a simple retreat from AI. It is a shift from a highly aggressive owned-infrastructure expansion toward a model that appears more cautious about buildings and more focused on compute resources that can support near-term needs.
What suppliers and rivals stand to gain
The impact on suppliers may be more limited than the headline numbers suggest. SemiAnalysis says companies such as Vertiv may not see as severe an effect as some assume, because much of the leased data center space is already under construction and order pipelines remain stable.
The larger change is competitive. As Microsoft slows parts of its own expansion and OpenAI places more work with other providers, Oracle and CoreWeave become more prominent in the AI infrastructure market.
That shift also shows how quickly AI infrastructure plans can change. Capacity, power, land, GPUs, and leasing contracts are all connected, but the demand behind them can move when major partnerships change or product uptake falls short.
For now, Microsoft is still expanding. But the company is doing so with a different mix of commitments than it appeared to be pursuing during the peak of its 2023 and 2024 buildout.