ANALYSIS The AI infrastructure race is being financed with increasingly aggressive leverage: SoftBank is tapping junk-bond markets for more than $11 billion to fund a single investment, while Microsoft plans to more than triple its data center fleet at annual spending approaching $200 billion, a trajectory that compresses margins and strains traditional capital structures.
Why it matters
The scale of capital now flowing into AI infrastructure has outgrown the balance sheets willing to carry it transparently. SoftBank Group Corp. is seeking the equivalent of more than $11 billion in what would be one of the biggest junk bond deals ever, earmarked for its OpenAI investment1. Meanwhile, Microsoft is planning a data center expansion from roughly 12 gigawatts of capacity to as much as 38 gigawatts over the coming years2. The financing mechanisms chosen to sustain this buildout, high-yield debt and capital-intensive direct spending, carry risks that do not always surface in headline earnings.
The big picture
Microsoft's ambitions illustrate the sheer physical cost of staying competitive. The company plans to more than triple its data center fleet, with spending "over to close to $200 billion a year on that effort". One gigawatt of data center capacity costs somewhere around $40 to $60 billion right now, and half of that spending goes to Nvidia. The company currently has 12 gigawatts of data center capacity across the world, of which 2 gigawatts is specifically for AI. Microsoft likely has a bigger data center fleet than anybody but Amazon Web Services.
Even at this scale, supply is not keeping up with demand. Microsoft, Amazon, and Google all complain that supply constraints in data centers are limiting cloud revenue growth. Temu could not get more capacity from Microsoft Azure and went to Oracle. Microsoft has more backlog revenue on the balance sheet than it has the capacity to fulfill.
The financial strain is visible. Microsoft's gross margin shrank to 61%, down from 66.3% in the previous quarter. One analyst noted that the company is not tapering CapEx down at this point, and "almost everybody is increasing CapEx next year". Microsoft would still be free cash flow positive this year, but the margin compression reflects the cost of retooling infrastructure from CPU-based older cloud computing to GPU computing.
On the other side of the capital stack, SoftBank's planned $11 billion-plus junk bond issuance represents a different kind of bet. The debt, denominated in both dollars and euros ($10 billion and €1 billion), would fund its OpenAI stake. ANALYSIS Using high-yield debt to finance a position in a private company concentrates credit risk: if OpenAI's valuation trajectory slows, SoftBank services the debt from other cash flows or faces refinancing pressure.
Between the lines
Microsoft is adapting its capital model in ways that shift risk. One analyst observed that the company is "asking customers to think about it, bring it on chips or bring your own hardware. This way they don't have to be ahead of time". ◆ This "bring your own" approach effectively transfers some infrastructure risk to customers, reducing Microsoft's upfront capital commitment per unit of capacity while preserving its role as the platform operator.
The biggest obstacle right now is permits and power for data center expansion, and opposition to data centers is coming from both Democrats and Republicans. Delays have been seen, though not massive ones. Microsoft has data centers going live in Texas and Wisconsin, and the company says the agentic era may require more CPUs than people expected even a year or two ago, because inferencing agents are accessing databases and creating records. ◆ That architectural shift means the capital intensity of AI may not plateau even as GPU-specific spending matures; CPU demand for agentic workloads adds a second cost curve on top of the GPU buildout.
Microsoft later said the reported capacity numbers "were not accurate without giving a whole lot more" detail. As Bloomberg's Brody Ford reported, the plan spans multiple years, and companies tend not to confirm projections beyond a year or so.
What's next
SoftBank's bond sale, if completed at the reported scale, would test appetite for AI-linked high-yield debt at a size rarely seen in junk markets. Microsoft's capacity constraints and margin pressure could play out over the next 6 to 12 months. Whether the industry's leverage, explicit in SoftBank's case and embedded in Microsoft's CapEx trajectory, proves sustainable depends on whether AI revenue growth keeps pace with the infrastructure commitments already locked in.