The AI infrastructure buildout is increasingly financed through mechanisms designed to keep its true scale off corporate balance sheets, from residual value guarantees that shift hardware risk to Wall Street, to one of the largest junk bond sales ever proposed to fund a single AI investment. Together, these moves suggest the industry's capital appetite has outgrown conventional corporate finance, pushing participants toward structures that obscure the full extent of their exposure.
Why it matters
The sums are staggering. Microsoft plans to more than triple its data center fleet to as much as 38 gigawatts, up from 12 gigawatts today3. The company is spending close to $200 billion a year on that effort. SoftBank is looking to issue $10 billion and €1 billion in debt for its OpenAI investment, in what would be one of the biggest junk bond sales ever2. And Big Tech is increasingly using residual value guarantees for off-balance sheet AI spend, giving Nvidia and Broadcom a way to support their customers' purchases1. Each of these financing channels carries distinct risk, but they share a common logic: the principals are betting that AI revenue will arrive fast enough to justify commitments that, if booked conventionally, would alarm shareholders and credit analysts.
The big picture
Wall Street has found a new way to turn tech giants' credit strength into cheaper funding for AI build-out. Residual value guarantees let hyperscalers acquire GPU clusters and data center hardware without recording the full liability on their books, effectively pledging that the equipment will retain a stated value at lease end. For Nvidia and Broadcom, the arrangement functions as a demand accelerant: their customers can buy more hardware, sooner, without the balance-sheet drag that might trigger internal spending reviews or analyst downgrades.
SoftBank's approach is blunter. The conglomerate is turning to the junk bond market to raise roughly $11 billion in combined dollar- and euro-denominated debt to fund its OpenAI stake. The scale of the issuance, among the largest ever in high-yield markets, reflects both the size of SoftBank's AI ambitions and the limits of its existing capital structure.
Meanwhile, Microsoft's infrastructure trajectory shows why these financing innovations exist. The company currently has 12 gigawatts of data center capacity across the world, with 2 gigawatts specifically for AI. Its plan targets 38 gigawatts, more than tripling the fleet through the coming years. One gigawatt costs somewhere around $40 to $60 billion right now for a data center. Microsoft has data centers going live in Texas and Wisconsin, and the company likely has a bigger data center fleet than anybody but Amazon Web Services.
Between the lines
Demand is not the constraint; supply is. Microsoft, Amazon, and Google all complain that supply constraints in data centers are limiting cloud revenue growth. Timu 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. That backlog creates a powerful internal justification for aggressive spending, but it also means the industry is building to a demand signal that has not yet been fully tested in production revenue.
The financial strain is already visible. Microsoft's gross margin shrank to 61%, down from 66.3% in the previous quarter. Yet the company would still be free cash flow positive this year, and almost everybody is increasing CapEx next year. Microsoft is also asking customers to bring their own chips or bring their own hardware, reducing the need to front capital for every workload.
ANALYSIS The "bring your own hardware" shift and the off-balance-sheet guarantees reported by the Financial Times are two sides of the same coin: hyperscalers are distributing financial risk away from their own books, whether to customers, to Wall Street intermediaries, or to junk bond investors.
The biggest obstacle right now is permits and power for data center expansion. Opposition to data centers is coming from both Democrats and Republicans. Delays have been seen, though not massive ones. Permitting friction could slow the buildout enough to test whether the financial structures underwriting it can tolerate extended timelines between capital deployment and revenue generation.
What's next
Microsoft says the agentic era may require more CPUs than people expected even a year or two ago, as inferencing agents access databases and create records. The company is changing the infrastructure of its data centers for GPU computing, which is different from CPU-based older cloud computing. The permitting and power issue could play out over the next 6 to 12 months. SoftBank's junk bond sale, if completed, will price the market's appetite for AI-linked credit risk at a scale not previously tested. Microsoft later said the reported capacity numbers were not accurate, without providing alternatives.