Off-Balance-Sheet Financing is an Infrastructure Staple

The AI infrastructure boom has changed the nature of hyperscaler financial commitments. Traditionally, investors focused on on-balance-sheet debt (bonds, bank loans and finance leases). Today, a growing issue is contractual obligations that are legally binding but are not yet recognized as liabilities under U.S. generally accepted accounting practices.

These commitments include:

• Multi-decade data center leases

• Colocation agreements

• GPU purchase commitments

• Power purchase agreements (PPAs)

• Network infrastructure contracts

• Joint venture funding commitments

• Capacity reservation agreements

The accounting treatment, while legitimate, represents future fixed cash commitments not reflected on balance sheets. 

According to Moody's Ratings, at year-end 2025, about $1.9 trillion in such future commitments were incumbent on a few hyperscale high-performance-computing suppliers::

Item	Amount
Total undiscounted future data center lease commitments	$969 billion
Not yet commenced (therefore largely off balance sheet)	$662 billion
Portion already commenced	about $307 billion

According to Moody's, $662 billion of uncommenced lease obligations exceeded the group's adjusted reported debt by roughly 113 percent. 

Company	Major off-balance-sheet commitments	Approximate current scale (2026)	Likely growth through 2028
Alphabet	AI data center leases, TPU infrastructure, power contracts	roughly $120B-$170B	$200B+ possible
Amazon	Data center leases, AWS capacity, equipment commitments	roughly $180B-$250B	$300B-$400B possible
Microsoft	Azure leases, OpenAI infrastructure, private investment funds	roughly $150B-$220B	$250B-$350B possible
Meta Platforms	AI campuses, Hyperion leases, networking	about $279B disclosed lease commitments after Q2; additional July commitments announced	$350B-$450B possible
Oracle	OCI data centers, GPU commitOfments, leased facilities	roughly $50B-$100B	$100B-$150B possible

Meta provides one of the clearest examples:

• future AI lease obligations appr;;;;;;\\q12aaching $279 billion

• another approximately $68 billion of new leases signed after quarter-end

• lease terms extending as long as 30 years

• commitments tied to multiple multi-gigawatt AI campuses.

Many obligations never appear as debt but nevertheless commit future cash flow:

Commitment type	Typical accounting treatment	Economic effect
Long-term data center leases	Recognized when lease commences	Future fixed payments
GPU purchase contracts	Usually not recognized until delivery	Locked capital spending
Power purchase agreements	Often disclosed as commitments	Long-term electricity costs
Network capacity reservations	Footnote disclosure	Fixed operating expense
Joint venture funding	Often partially off balance sheet	Future capital contributions
Colocation agreements	Lease recognition delayed until commencement	Multi-year cash obligations

Such practices are common in other capital intensive industries, or parts of industry infrastructure, including airline transportation providers, because doing so:

• preserves reported leverage ratios

• matches accounting recognition to asset availability

• allows capacity to be secured years before facilities open

• enables landlords and infrastructure funds to finance construction

• reduces the need to issue corporate debt immediately. 

If AI capital spending continues near today's pace, a reasonable scenario for Alphabet, Amazon, Microsoft, Meta and Oracle is:

p[;;;;;

Year	Estimated aggregate off-balance-sheet commitments (five hyperscalers)
2025	~$662B (Moody's estimate of uncommenced leases)
2026	~$800B-$1.0T
2027	~$1.1T-$1.4T
2028	~$1.3T-$1.7T

The issue for some observers is whether such future commitments are a problem, or not. Even if not considered “debt” in a GAAP sense:

• They require future cash payments.

• They reduce future financial flexibility.

• Many cannot easily be canceled without substantial penalties.

• Credit analysts increasingly incorporate them into leverage analysis.

The issue is impact on firm leverage. 

But that seemingly often is an issue with the financing of infrastructure. Railroads, electric utilities, pipelines, airports, seaports, cellular networks, and fiber networks all required enormous up-front investment years before meaningful revenues arrived. 

Each wave developed financing techniques that shifted risk away from the sponsoring company while securing long-term capital.

Infrastructure era	Typical financing	Off-balance-sheet elements	Primary revenue support	Similarity to AI infrastructure
Railroads (1800s)	Corporate bonds, land grants	Limited	Freight and passenger traffic	Moderate
Electric utilities (1900s-present)	Utility debt, project finance	Power purchase agreements, independent power producers	Regulated utility revenues	Very high
Seaports	Public authorities, revenue bonds	Long-term terminal concessions	Shipping fees	High
Airports	Municipal bonds, PPPs	Airline gate leases	Passenger and airline fees	High
Toll roads	Project finance, PPPs	SPVs, concession agreements	Toll revenues	Very high
Mobile networks	Corporate debt, tower leasing	Tower REIT leases	Wireless subscriptions	Very high
Fiber networks	Project finance, infrastructure funds	Long-term IRUs, dark fiber leases	Wholesale and retail access	Extremely high
AI data centers	Corporate debt, project finance, lease finance	Long-term data center leases, PPAs, equipment commitments	Cloud and AI services	Highest

Electricity infrastructure evolved from vertically integrated utilities financing everything on their own balance sheets to today's mixture of:

• utility-owned assets,

• independent power producers,

• project-financed generation,

• long-term power purchase agreements (PPAs), and

• infrastructure funds.

The important innovation was separating ownership from usage. A utility could commit to buying electricity for 20–30 years without necessarily owning the generating plant. That resembles today's hyperscalers signing 15- to 30-year leases for AI campuses built by third-party developers.

Modern project finance emerged because infrastructure became too expensive for sponsors to fund entirely on their own balance sheets.

Instead:

• a special-purpose vehicle (SPV) owns the project,

• lenders are repaid primarily from project cash flow,

• the sponsor's liability is limited,

• long-term customer contracts reduce lender risk.

This structure became common for:

• toll roads,

• airports,

• ports,

• power plants,

• pipelines.

AI data centers resemble these projects.

Wireless carriers originally owned virtually everything:

• towers,

• land,

• buildings,

• backup power.

Beginning around 2000 they sold towers to companies such as American Tower, Crown Castle, and SBA Communications.

Instead of ownership they signed:

• 10- to 20-year leases,

• automatic renewals,

• inflation escalators.

From an economic perspective, tower lease obligations became debt-like commitments while freeing carriers' balance sheets for spectrum purchases and network equipment. Hyperscalers appear to be following almost exactly the same path.

Debating Amazon Leo Objectives

Amazon’s objectives with Leo are debated. 

Is this a standalone telecom business or a strategic infrastructure layer feeding higher-margin businesses (likely AWS)?

The possible motives are complicated as Amazon often talks like a “margin hunter,”  but often acts like a scale builder that tolerates thin margins for a time. 

The trick is that Amazon usually tries to separate where value is created from where it is captured. 

Amazon repeatedly enters markets characterized by low margin and high margin, so “margin” is not the primary consideration.

The effort to find “moats” or bottlenecks where value is extracted, and sometimes a low-margin business can lead to a high-margin moat. 

Layer	Characteristic	Amazon Behavior
Customer-facing layer	Huge TAM, fragmented, price-sensitive	Compete aggressively, often low margin
Infrastructure / platform layer	High fixed cost, scalable, defensible	Invest heavily, aim for high margin long term
Data / control layer	Feedback loops, optimization	Build moats that others can’t replicate

The point is that Amazon doesn’t mind entering a low-margin market if it helps it own a high-margin layer underneath or adjacent to it.

Also, “high capital investment” can be a feature, not a bug:

• High CapEx deters competitors

• Once built, marginal costs drop sharply

• Scale converts fixed costs into a profit flywheel

• Infrastructure can support multiple businesses

• Pricing power eventually comes, once dominance is achieved.

So huge capex commitments are consistent with Amazon’s playbook, if Amazon believes it can control a bottleneck layer.

“Is this a high-margin or low-margin business?” might not be the right question for Amazon leaders, who likely are asking:

• Can we own a critical layer?

• Does this scale globally?

• Does it reinforce our existing flywheels?

• Can we improve cost structure vs incumbents?

• Is there a hidden high-margin component?

So the larger picture is often not the immediate or obvious business, but the ability to create leverage elsewhere. Consumer initiatives such as e-commerce; devices or streaming then can be viewed as demand aggregators and ecosystem lock-in creators that drive revenue indirectly (advertising, cross selling, subscriber lock in).

Enterprise infrastructure plays such as AWS or logistics might be better examples of direct, high margin initiatives.

The thing about Leo is where it fits. From one point of view, consumer telecom is a low-margin, highly-competitive business with high regulatory conditions, low innovation and low growth rates. 

So why even consider it?

Amazon probably envisions non-obvious leverage points:

• Edge computing

• Private 5G networks

• Carrier cloud

• Where Amazon captures high-margin compute, not connectivity

• With different value drivers in consumer and business markets.

Owning a connectivity service could:

• Reduce internal costs

• Improve performance (latency, reliability)

• Be bundled with Prime and devices to

• Drive usage of AWS, the advertising platform and e-commerce

• Support IoT connectivity (devices, logistics, smart home). 

Framed that way, Leo might be viewed as a platform layer supporting:

• Edge cloud

• AWS (compute plus connectivity)

• Telcos as customers

• Prime average revenue per user or account

• Customer retention and acquisition

To be sure, execution will matter. But, in theory, Leo is not directly about high margin. It is about control of what is likely to be a low-margin feature of a higher-margin ecosystem. 

Amazon’s explicit framing is straightforward:

• Create a global broadband access business

• Serving “tens of millions of customers” globally

• in “unserved and underserved” markets

• Offers private connectivity directly into AWS

• for enterprise, government, and telecom customers.

So AWS integration, enterprise and government use cases and private networks might be key, not “consumer telecom.”

Leo then is a connectivity extension of AWS. 

But there are clear risks and some skeptics. 

Optimistically, Leo extends AWS to the edge of the network. 

On the other hand, it is a near-term drag on earnings, in a business with tough economics and financial returns that could take some time to develop.

So it might matter hugely whether Leo can generate AWS pull-through; enterprise demand and other ecosystem upsides. 

Also, how long this takes could matter. 

Layer	Role	Margin Potential
Consumer broadband (Leo ISP)	Distribution / scale	Low
Enterprise connectivity	Premium services	Medium
AWS integration layer	Data + compute + control	High
Ecosystem effects (Prime, commerce, ads)	Indirect monetization	Very high

Sure, it’s risky. But some will point to past Amazon initiatives based on entry into low-margin businesses that provided moats:

• Retail → low margin → enabled AWS scale

• Devices → low margin → enabled ecosystem lock-in

• Logistics → low margin → enabled marketplace dominance.

Leo arguably fits the pattern, optimists will argue. It’s about high-margin AWS, not low-margin telecom. Skeptics will worry about the execution risk.