Hobite Capital Research Note
AI Infrastructure Financing
Hidden debt, real commitments, and who carries the risk
The hidden-debt story is real, but most of the exposure is not literally undisclosed debt. It is a mixture of future leases, purchase commitments, guarantees, project debt, and long-dated infrastructure contracts that are disclosed in footnotes but are not yet recognized as conventional balance-sheet debt.
The bearish claim that Big Tech secretly owes trillions is too aggressive. The opposite claim that these are harmless footnotes is also wrong. A large share of the obligations behaves economically like debt because future cash flows are committed years, and sometimes decades, in advance.
01 / The shift
From software economics to infrastructure economics
The hyperscaler model is moving from software, advertising, cloud, and abundant free cash flow toward a hybrid that also resembles telecom, utilities, real estate, and infrastructure finance. These remain excellent businesses, but the economics are becoming more capital intensive.
Future leases
$1.09T
Estimated for five major hyperscalers in the source analysis.
Recognized leases
$285B
Currently recognized lease liabilities in the comparison.
2026 capex
$750B
Estimated spending, equal to roughly 38% of combined revenue.
Old model
Software + advertising + cloud
High incremental margins and large free-cash-flow conversion.
Emerging model
Compute + power + real estate + finance
Higher fixed costs, longer commitments, and greater sensitivity to utilization.
02 / Capital structure
Five ways the obligations are created
| Structure | Economic reality | Examples |
|---|---|---|
| Uncommenced leases | Future rent is committed, but the liability is not recognized until the lease begins. | MSFT, AMZN, ORCL, META, GOOG |
| Purchase and take-or-pay commitments | The company commits to GPUs, servers, power, or capacity before demand is fully proven. | All major hyperscalers |
| SPV and joint-venture financing | An outside vehicle borrows while the hyperscaler becomes the anchor tenant or guarantor. | META, especially Hyperion |
| Guarantees and backstops | The company guarantees payment obligations tied to third-party data-center projects. | GOOG |
| Direct corporate bonds | Conventional debt is increasingly used to fund the AI capital cycle. | GOOG, AMZN, META, ORCL |
The key analytical adjustment is to look beyond GAAP debt and estimate total committed capital usage. The disclosures generally exist, but the headline debt figure alone does not capture the full economic burden.
03 / Case studies
How the structures differ by company
MSFT
The lease pipeline is the main issue
Microsoft disclosed roughly $196.6B of additional leases, primarily for data centers, that had not yet commenced as of March 31. The terms can extend as long as 21 years. This is enormous beside roughly $40B of conventional debt, although Microsoft's AAA-quality balance sheet makes it a capital-return issue rather than a near-term solvency concern.
GOOG
Guarantees create contingent leverage
Alphabet backstops some third-party data-center payment obligations and accounts for those guarantees as credit derivatives. The source answer cites about $43.8B of potential credit-derivative exposure, $7.6B of financial guarantees, and $24.1B of potential future infrastructure backstops as of June 30, 2026.
AMZN + ORCL
Pending leases reveal different risk levels
Amazon's future lease pipeline is meaningful but supported by a diversified cash engine. Oracle's commitments are more concerning because they sit beside roughly $130B of debt, negative levered free cash flow, lower credit quality, and leases that can run for 15 to 19 years.
META
Hyperion is project finance, not disappearing debt
Meta and Blue Owl formed a joint venture for the Hyperion campus in Louisiana. The project vehicle raised about $27.3B of senior secured debt maturing in 2049. It is not Meta corporate debt, but Meta is the anchor tenant whose lease payments support the project economics.
Why finance indirectly?
Project structures preserve corporate borrowing capacity, match debt with assets, share construction and residual-value risk, and protect credit ratings. They also allow investors to own AI infrastructure risk separately from hyperscaler equity risk.
04 / Core risk
Long debt, short technology life
The strongest part of the AI-subprime argument is duration mismatch. A data-center building or grid connection may remain useful for decades, while the GPUs inside can become commercially obsolete in five to seven years. Twenty-year financing can therefore outlive the most valuable collateral by a wide margin.
20-25 years
Possible project-financing duration
5-7 years
Potential commercial life of a GPU generation
90%
Illustrative utilization assumption that may prove fragile
If compute efficiency improves rapidly, inference pricing collapses, utilization falls, or newer accelerators consume much less power, project cash flows can deteriorate while the debt remains fixed. The collateral loses value faster than the financing amortizes.
That still does not make the system equivalent to 2006 subprime housing. The anchor customers are among the world's strongest cash generators. The closer analogy is the 1998-2001 telecom and fiber buildout: the technology was real, demand eventually became enormous, and too much capital was still deployed too early.
05 / Transmission
Where financial stress is likely to appear first
01
Private credit and infrastructure funds
Long-duration claims increasingly sit with private lenders, insurers, pension funds, and infrastructure vehicles. Refinancing losses and lower collateral values would appear here first.
02
Data-center developers
Operators without hyperscaler balance sheets are highly exposed to construction delays, lower lease pricing, and utilization shortfalls.
03
GPU-backed lenders and neoclouds
This is the most subprime-like part of the chain: aggressive borrowing against fast-depreciating hardware and short-duration demand.
04
Utilities and power developers
Gigawatts of generation and transmission are being built against long-duration demand assumptions that may not all be realized.
05
The corporate bond market
Heavy hyperscaler issuance competes for capital and can lift financing costs for other investment-grade borrowers.
The realistic bearish case is not that Microsoft or Alphabet suddenly fails. It is that returns on incremental capital disappoint, free cash flow remains depressed, and equity valuation multiples compress even while AI adoption and reported earnings grow.
06 / Portfolio view
Risk-adjusted investment ranking
META
Low to moderateMost attractive
The advertising engine still produces exceptional cash flow, while the market has already discounted a meaningful portion of the AI spending risk. Project finance adds complexity, but the valuation offers the best risk-adjusted setup in the group.
MSFT
LowQuality at a reasonable price
The lease pipeline is enormous, but Microsoft's credit quality, cloud backlog, and enterprise software ecosystem provide unusually strong visibility. Economic leverage is higher than headline debt suggests, yet solvency risk remains remote.
AMZN
ModerateBuy on meaningful pullbacks
AWS momentum supports the monetization case, but the stock is not pricing a disaster and free cash flow is obscured by a very heavy investment cycle. Operating quality is strong; valuation support is less obvious than at Meta.
GOOG
ModerateAttractive, monitor commitments
Alphabet combines direct debt with guarantees, credit-derivative treatment, and infrastructure backstops. The business remains attractive, but returns on the expanding AI capital base matter more than a superficially reasonable P/E.
ORCL
HighHigh-beta special situation
Oracle has the most difficult mix of debt, negative levered free cash flow, long lease commitments, and execution concentration. A large drawdown creates rebound potential, but this is not the same quality of opportunity as the other four.
This ranking is based on risk-adjusted prospective return, not maximum upside. Oracle could rebound more sharply than Microsoft, but it does not represent the same quality of opportunity. Position sizing and portfolio concentration remain more important than small estimates of near-term undervaluation.
07 / Monitoring
Five signals that would weaken the AI capital cycle
- 01
Cloud growth slows while capital spending continues to grow above 30%.
- 02
AI infrastructure utilization falls and empty GPU hours begin to resemble vacant real estate.
- 03
GPU rental prices collapse faster than performance per dollar improves.
- 04
Hyperscaler credit spreads widen even as reported earnings remain strong.
- 05
SPV and project bonds become difficult to issue, forcing more financing onto corporate balance sheets.
Central conclusion
This is not yet an AI subprime crisis. It is the financialization of AI infrastructure.
The decisive question is no longer who has the best model or the most GPUs. It is who can earn a return on trillions of dollars of capital above the cost of financing it. AI can succeed technologically while shareholders still receive disappointing returns if capital intensity stays high and incremental returns fall.