The AI investment boom is increasingly being financed in ways that do not show up as conventional debt on Big Tech’s balance sheets. A Nikkei study published on 21 July 2026 found that five major technology companies, Alphabet, Microsoft, Amazon, Meta and Oracle, have accumulated $1.65 trillion in purchase commitments and leases tied to future AI infrastructure.
The table below was collated based on the newest filings available: the Q1-2026 10-Qs (period ended 31 March) and Oracle’s FY26 10-K (31 May). For each company, it adds the purchase and other commitments the company disclosed to the leases it has signed but not yet started.
| Company | Period | Purchase / other commitments | Leases not yet commenced | Total off-BS | Total liabilities | Interest-bearing debt |
|---|---|---|---|---|---|---|
| Alphabet | 31 Mar 26 | 332.4 | 75.6 | 408.0 | 225.2 | 79.5 |
| Microsoft | 30 Jun 2025* | 142.1 | 196.6 | 338.7 | 279.9 | 45.0 |
| Amazon | 31 Mar 26 | 103.8 | 106.4 | 210.1 | 474.7 | 122.4 |
| Meta | 31 Mar 26 | 237.7 | 182.9 | 420.6 | 151.6 | 58.7 |
| Oracle | 31 May 26 | 13.3 | 260.0 | 273.3 | 218.7 | 129.5 |
| TOTAL, rebuilt | 829.2 | 821.4 | 1,650.7 | 1,350.0 | 435.2 | |
| Nikkei published | 1,650.0 | 1,350.0 | - | |||
| Variance | 0.7 | 0.0 | - |
*Microsoft discloses these only annually; its Q3 FY26 10-Q contains zero occurrences of “purchase commitment” or “construction.”
Note that what the $1.35tn stated in the Nikkei Asia article compares against is total liabilities, not debt. Measured against actual borrowings of $435bn, the ratio is 3.8 times, not 1.2.
The largest technology companies are now spending an increasing share of their operating income on new capacity, forcing them to borrow and issue shares to cover the gap. Add the capital coming from institutional investors, and the result is a building boom that may be running well ahead of the returns it can generate.
The BIS highlighted this risk in March, with economists using the term “shadow borrowing” for structures that funnel institutional capital into these projects while keeping the associated debt off the balance sheet, and warned that stalled construction could turn a local problem into a broad loss of confidence across the AI trade.
Source: Bloomberg, S&P Capital IQ, PIMCO as of 15 May 2026. LTM = last 12 months.
Capital spending took up about 53% of operating cash flow at the start, fell to 40% in 2023, and has risen consistently since: 49% in 2024, 68% in 2025, and 75% on a trailing twelve-month basis through the first quarter of 2026. The more striking number is the forecast, shown in the hatched bars: close to 94% for both 2026 and 2027.
At that level, the hyperscalers would be putting almost all their operating cash flow into capital projects, mainly AI data centers and compute capacity. That leaves little for dividends, share buybacks, or paying down debt. This is why the trend matters for AI-related credit and for the wider economy: spending at this level means leaning more on outside financing rather than on the cash the businesses generate themselves.
Where is the money coming from?
The hyperscaler takes a 20% equity stake, signs a long-term lease, writes a residual value guarantee, and argues that it does not direct the activities that most significantly affect the vehicle’s economic performance, so it is not considered the primary beneficiary and does not have to consolidate the vehicle on its balance sheet. The debt is real, sits at the SPV, and does not appear on the hyperscaler’s balance sheet. Keeping ownership below 20% allows them to leave both the entity and its debt off their own balance sheets. The table below sets out several recent deals:
| Vehicle | Sponsors / split | Project cost | Debt raised | Verification |
|---|---|---|---|---|
|
Beignet Investor LLC (Hyperion, Louisiana) |
Meta 20% / Blue Owl 80% | $27.0bn | $27.294bn | 6.581% fixed, matures 30-May-2049, CUSIP 076912AA2, S&P A+. |
|
Abilene DC 1, LLC (Stargate Ph.1) |
Crusoe 7.7% / Blue Owl 92.3% | $3.4bn | $2.288bn | Full KPMG-audited financials filed as Ex-99.3 to Blue Owl NLT's 10-K. |
|
“Project Sopaipilla” (El Paso, Texas) |
Meta 20% / BlackRock 80% | ~$14bn | $12 to 12.5bn | This was reported only, not verified. Name returns zero EDGAR hits. Exclusivity only; close expected Q3-26. |
| Broadcom AI XPV Platform | Apollo lead / Blackstone | n/a | $35.0bn | Offtaker is Anthropic, not Google. |
| Related Digital (Michigan) | Related / Blackstone | $16bn | $14.0bn | This was reported only. 7.50%, 2045. PIMCO bought ~$10bn. Blue Owl declined to fund. |
All figures in US$bn. Source: company 10-Q and 10-K filings, EDGAR.
The Hyperion structure is worth a closer look, because the debt does not sit where the description suggests. The $27.294bn bond was issued not by the property joint venture but by Beignet Investor LLC, the entity holding the 80% interest in it. Blue Owl’s equity in the venture is therefore itself around 92% debt funded.
Meta put in assets rather than cash, and actually took money out at closing. It transferred $4.30bn of assets net of liabilities, the land and the part-built campus, and received a one-time payment of $2.55bn back, leaving a net book investment of about $1.75bn. Its total commitment over the life of the project is around $5.6bn, or 20% of the $27bn development cost. Blue Owl Real Estate Net Lease Trust has disclosed a $1.533bn equity commitment for its 63% indirect interest in Beignet Investor, which implies about $2.43bn of total equity at that level.
So the real first-loss equity sitting beneath the notes is about $2.4bn, or roughly 9% of the bond. The bond is larger than the whole development cost, a debt-to-cost ratio of about 101%, with the extra covering the interest bill until rent starts in 2029. Conventional data center project financing runs at 25% to 40% equity. Blue Owl’s disclosed remaining build-to-suit funding obligation across all its joint ventures was $1.887bn at December 2025 and $1.732bn at March 2026, which suggests very little of that equity had actually been paid in, even though the full bond had been outstanding since October.
What protects the bondholders sits outside the capital structure entirely: a residual value guarantee worth up to $28bn in total, larger than the bond it stands behind. That guarantee, together with a $12.31bn lease commitment, is what S&P rated A+, one notch below Meta itself. The rating rests on Meta’s promise to pay, not on Louisiana real estate or on the equity cushion. Meta’s auditor flagged the decision not to consolidate the venture as a Critical Audit Matter. Then the question to ask is this: if the rating is based on Meta’s ability to pay, does that not count as debt? Additionally, the bond raised to fund this project has already been paying interest since October 2025, probably out of the bond’s own proceeds rather than from any income, since the campus earns nothing until 2029.
Is the spending justified?
Technology writer Ed Zitron makes the most alarming case, arguing that more than fifteen times the required capacity is being planned. Three parts of his analysis are shaky: his cost figures contradict each other, his revenue target does not match the one real market price he cites, and his demand estimate of $120 billion a year looks low when OpenAI and Anthropic alone already run at about $70 billion. A steadier version lands nearby. Bain, which advises the companies doing the building, sees the industry needing $2 trillion of annual revenue by 2030 and can find only $1.2 trillion of it.
Three factors argue against such a conclusion. Most of the money is still a promise, with about $1.15 trillion of the $2.64 trillion actually spent, so the rest can be slowed. What has been spent is plainly real: $49.7 billion of data centres built in 2025 on government figures, $384 billion of chips banked by NVIDIA, and Broadcom’s order book up from $25 billion to $165 billion. And the industry cannot build fast enough anyway, since GE Vernova delivered only 7.5 of 20 gigawatts ordered and grid connections now take over five years.
History suggests the likely outcome. Fibre optic cable and shale oil both produced something genuinely valuable while wiping out many of the investors who funded the build-out. Expect large losses without a repeat of 2008, since the borrowing is far smaller relative to the companies involved and none of the layered betting that turned housing losses into a banking crisis exists here. Investors have already taken the hit, with Oracle and CoreWeave each down about 62% without a missed payment. More tellingly, the growth rate of US data centre construction has slowed from 77% a year in 2024 to 23% in 2026.