AI Companies Are Hiding $1.65 Trillion in Debt. Here Is What the Off-Balance-Sheet Structure Actually Means for the Engineers Who Build on These Platforms.

The number that matters most in the AI industry right now is not the benchmark score, not the context window length, not the parameter count. It is $1.65 trillion. That is the estimated off-balance-sheet debt that five US tech giants — Alphabet, Microsoft, Amazon, Meta, and Oracle — have accumulated and deliberately kept out of their official financial reporting, according to a Nikkei Asia investigation published this week. The five companies officially reported $1.35 trillion in debt for their most recent quarter. The hidden number is 22 percent larger.

Meta alone accounts for roughly $420 billion of that invisible pile. Let that settle for a moment. A company whose market capitalization fluctuates around $1.4 trillion is carrying off-balance-sheet obligations that represent nearly a third of its visible value — and this debt does not appear on the balance sheet investors look at when they price the stock.

I have spent years watching infrastructure investment cycles in tech, and this pattern has a name: it is the same mechanism Enron used before its 2001 collapse. Special purpose vehicles, legally distinct subsidiaries, operating lease structures recharacterized as something else. The goal is always the same: make the income statement and balance sheet look healthier than the underlying cash reality warrants. When accounting consultants start telling Bloomberg that the accounting treatment “itself is in fashion” and then immediately add “but what if one of these companies was a house of cards,” you are past the warning sign and into the territory where the warning sign has been ignored so long it has become wallpaper.

This piece is not a prediction that any specific company will collapse tomorrow. It is an argument that the engineering community — broadly defined to include infrastructure architects, platform engineers, and anyone who builds systems on top of these clouds — is systematically underweighting a structural risk that is hiding in plain sight inside quarterly filings.

What Off-Balance-Sheet Debt Actually Means for Infrastructure

Off-balance-sheet obligations are not fictional accounting entries. The data centers get built. The power purchase agreements get signed. The cooling infrastructure goes in the ground. The GPU clusters arrive in shipping containers. What is fictional — or at minimum, misleading — is the way these obligations are characterized in financial reporting.

The most common mechanism is the operating lease versus capital lease distinction. Under older accounting standards, a company could sign a 15-year data center lease, keep it off the balance sheet as an “operating” arrangement, and report only the current year’s lease payment as an expense. The International Financial Reporting Standards (IFRS 16) and the US FASB equivalent (ASC 842) changed this starting in 2019, requiring most leases to appear on the balance sheet. But there are still ways around it: build-to-suit arrangements where a developer builds a facility to a company’s exact specification and the tech giant commits to a long-term power or space agreement that is characterized as a service contract rather than a lease.

Power Purchase Agreements (PPAs) for data center electricity work similarly. A company signs a multi-decade commitment to buy power at a fixed rate from a nuclear or renewable energy project. That commitment is disclosed in footnotes under “contractual obligations” or “purchase commitments” — not recorded as a liability on the main balance sheet. The obligation is real. The counterparty is depending on the payment stream to finance the power plant construction. But from a balance sheet perspective, it does not exist until the invoice arrives.

The practical consequence for infrastructure engineers is more concrete than it might sound. When Microsoft commits to buying electricity from a nuclear plant for 20 years to power Azure data centers, that commitment shapes what Azure can credibly offer customers over that same 20-year horizon. It also shapes what happens to Azure’s investment trajectory if circumstances change. When those commitments are being obscured in financial reporting, the risk profile of the platform is also being obscured. An engineer deciding whether to build on AWS, Azure, or GCP is implicitly making a bet on the financial sustainability of the underlying infrastructure. If the infrastructure is funded by obligations that are not visible in standard financial analysis, that bet is being made with incomplete information.

Note: Off-balance-sheet estimates from Nikkei Asia investigation (July 2026). Reported debt figures are approximate from most recent quarterly filings. The total across all five companies is estimated at $1.65 trillion in off-balance-sheet obligations versus $1.35 trillion officially reported — a gap of approximately $300 billion, or 22 percent of the combined disclosed liability.

The Enron Comparison Is Not Hyperbole

Every time someone invokes Enron in a financial article, there is a reflexive urge to dismiss it as alarmism. The Enron comparison is not being made to predict imminent collapse. It is being made to identify a structural mechanism that is well-documented, has a clear historical precedent, and is being repeated — at a scale that dwarfs the original.

Enron’s off-balance-sheet entities, the Special Purpose Entities (SPEs) with names like LJM Cayman and Chewco, were used to hide roughly $1.2 billion in debt from investors. The SPEs met technical accounting criteria for independence — Enron did not consolidate them — while in practice being controlled by Enron executives and backed implicitly by Enron’s credit. When the SPEs’ underlying assets fell in value, Enron was on the hook in ways that were invisible to the outside world until it was too late. The company went from being rated investment grade to filing for bankruptcy in approximately 24 days in late November 2001. The speed of the collapse was a direct function of the opacity: nobody had a clear picture of the actual obligation structure, so once confidence cracked, it cracked completely.

The mechanism being employed by today’s tech giants is structurally similar but legally cleaner. They are not using entities controlled by insiders with undisclosed conflicts of interest. They are using legitimate accounting treatments — PPAs, build-to-suit lease arrangements, committed purchase obligations — that are explicitly permitted under GAAP and IFRS. The problem is not that they are breaking rules. The problem is that the rules permit a significant distortion between the financial picture presented to investors and the actual obligation structure of the business.

I went through Microsoft’s most recent 10-K specifically looking at the commitments footnote. The numbers listed under “operating leases and other commitments” are in the tens of billions. But the nuclear power purchase agreement with Constellation Energy for the Three Mile Island restart, which Microsoft committed to in September 2023, is a 20-year, multi-billion-dollar obligation. It does not appear as a liability on the balance sheet. It appears as a disclosure in a footnote. The distinction matters enormously if you are trying to model what Microsoft’s actual obligation structure looks like under stress scenarios — for instance, if AI revenue growth slows significantly over the next five years and the company needs to rationalize spending.

The Enron parallel is not: “Microsoft is committing fraud.” It is: “The mechanism of using legally-permitted structures to obscure real obligations has a specific historical failure mode, and the scale of what is happening now is 1,375 times larger than the number that took down Enron.”

The AI Build-Out Is Making This Structurally Worse

Off-balance-sheet financing is not new. Airlines have been doing it with aircraft leases for decades. Real estate companies use similar structures routinely. Utilities sign long-term fuel supply contracts that are disclosed as commitments rather than liabilities. The reason the AI context is specifically concerning is the combination of three factors that do not typically occur together at this scale and speed.

First, the investment horizon is unusually long and demand is genuinely unknowable. A data center built in 2025 to run current-generation GPU clusters will be 10 years old by 2035. The commitments being signed now — 20-year PPAs, 15-year build-to-suit agreements — extend to 2044 and 2040 respectively. What will AI workloads look like in 2040? What will the competitive landscape look like? What will energy costs look like after 20 years of climate policy evolution? Airlines have reasonable visibility into passenger demand at 5-10 year timescales; the passenger jet is a mature technology. AI infrastructure is not. Anybody claiming they have a credible demand forecast for AI compute in 2040 is either deluded or lying.

Second, competitive pressure is forcing simultaneous overcommitment across all major players. Microsoft, Google, Amazon, Meta, and Oracle are all making these bets at the same time, which means none of them can pull back unilaterally without ceding ground to competitors who continue building. This is a classic prisoner’s dilemma arms race, and arms races routinely produce overbuilding that only becomes apparent after the fact. The telecom fiber overbuild of the late 1990s resulted in approximately 97 percent of installed fiber going unlit for years after the bubble burst, with carriers that had signed long-term capacity commitments facing years of write-downs and restructurings.

The AI data center build is happening faster, at a larger scale, and with less regulatory friction than the fiber build. The fiber companies could not hide the dark fiber — you could actually see it sitting unlit in conduit. Data center utilization rates and the relationship between committed capacity and actual workload can be obscured much more easily in quarterly reporting.

Third, the revenue justification is concentrated in a small number of enterprise contracts that are themselves contingent on AI adoption continuing to accelerate at current rates. Azure’s AI revenue is real — Microsoft’s cloud revenue growth has been genuinely impressive. But a significant portion of the growth is coming from enterprise “AI consumption commitments,” which are agreements by customers to spend a certain amount on AI services over a contract period. These commitments do not guarantee actual usage. Customers commit to spend $10 million on Azure AI over 24 months; whether they actually consume that compute depends on whether their AI applications actually get built and deployed and used. The gap between committed spend and actual consumption is not publicly disclosed but is known to be material across the industry.

Why Engineers Should Care More Than Investors Do

The standard framing of this story is an investor-protection story: retail investors are being misled about the risk profile of companies they hold shares in. That is true and it matters. But I think the engineering angle is equally important and significantly underappreciated in the current coverage.

When you build a serious application on a cloud platform, you are entering a dependency relationship that has real switching costs. Migrating a mature application from AWS to GCP or Azure is not a weekend project. For a production system with substantial state, multiple services, region-specific configurations, and training data stored in native formats, a cloud migration can take 6-18 months of dedicated engineering time and cost millions in labor plus the dual-running period during cutover. Most companies that commit to a cloud platform are making a de facto 5-10 year commitment even if their contract term is shorter.

That commitment is implicitly a bet on the provider’s long-term investment trajectory and service continuity. Cloud providers have deprectated services before — Google has a particularly well-documented history of killing products — and they have also raised prices, reduced free tiers, and changed service terms in ways that imposed real costs on dependent applications. None of those events rose to the level of existential crisis for the providers. But if the financial structure underlying a major cloud platform becomes stressed — not necessarily catastrophically stressed, but stressed enough to force rationalization decisions — the impact on engineering teams who built on that platform can be severe.

The scenario I am specifically modeling is not bankruptcy. It is a 40-50 percent deceleration in AI revenue growth relative to current expectations. That scenario, which I consider plausible within a 3-year horizon, would require a fundamental reassessment of the pace of data center buildout. The companies would need to cancel some committed capacity, extend depreciation schedules on existing assets, and prioritize high-margin workloads over experimental ones. In that scenario, the services that get rationalized first are the ones with the smallest installed bases and the longest payback periods — which describes a significant fraction of the AI-specific services (fine-tuning APIs, AI search, multimodal services) that companies are currently being encouraged to build on.

I am not predicting any specific service gets killed. I am arguing that the financial structure of the entity you depend on is a legitimate engineering input into architecture decisions, and that the current reporting environment makes it systematically harder to evaluate that financial structure than it should be.

The Special Case of Oracle and the GPU Cluster Bet

Oracle’s position in this landscape deserves specific attention because it is structurally different from the other four companies and in some ways more exposed. Oracle Cloud Infrastructure (OCI) came late to hyperscale AI infrastructure — the company’s cloud revenue in fiscal 2022 was small compared to AWS, Azure, and GCP. In 2023 and 2024, Oracle made an aggressive bet to catch up by signing GPU cluster commitments — specifically, large blocks of Nvidia H100 and H200 capacity — that it then resells to AI companies that cannot get direct allocation from Nvidia.

This model is different from the other four in an important way: Oracle is essentially acting as a GPU capacity financier and reseller, not just a platform operator. The company committed to purchasing massive blocks of GPU capacity on behalf of a customer set that includes AI startups and mid-size enterprises that want cloud-scale GPU access without building their own data centers. Oracle’s revenue from these arrangements is real — the company’s remaining performance obligations (RPO) grew significantly in recent quarters on the back of multi-year AI cluster commitments. But the credit risk in the customer base is different. An AWS customer who needs to renew their contract is typically a mature enterprise. An Oracle OCI customer who signed a three-year GPU cluster commitment in 2024 might be an AI startup that raised Series B money during the AI funding wave. The attrition risk in that customer base is not the same.

Oracle reported approximately $88 billion in long-term debt in its most recent filing — already the highest of the five companies in nominal terms relative to its market cap. If the Nikkei estimates of additional off-balance-sheet exposure are proportionally similar to what is estimated for Meta, Oracle’s actual obligation structure could be significantly larger than what appears in its filings.

The Q2 Earnings Window Is the First Test

Four of the five companies analyzed in the Nikkei investigation are reporting Q2 2026 earnings in the coming weeks. These reports will be scrutinized more intensely than usual in light of the reporting, and a few specific things are worth watching carefully.

First, whether any company changes the characterization of its commitments in response to increased scrutiny. Financial reporting language evolves in response to investor and regulatory pressure. If the Nikkei investigation generates enough analyst questions on earnings calls — and given that it hit 490 points on Hacker News and generated Bloomberg coverage, it will — some companies may choose to provide more granular voluntary disclosure about their off-balance-sheet exposure. The presence or absence of that additional disclosure will itself be informative: companies that proactively clarify are signaling confidence in their actual obligation structure; companies that dodge the question are not.

Second, whether the AI consumption figures continue to show the growth rates needed to justify the infrastructure investment. Microsoft’s Q3 fiscal 2026 showed Azure growth of approximately 33 percent year-over-year, with AI contributing meaningfully to that growth. If Q4 shows material deceleration, the financial case for the data center buildout weakens and the off-balance-sheet obligations start looking heavier in relative terms. The question is not whether AI revenue is growing — it is. The question is whether it is growing fast enough to service the obligation structure being built to support it.

Third, analyst behavior on the calls themselves. There is a small set of financial analysts who cover these companies and are specifically focused on infrastructure capital allocation — Mark Moerdler at Bernstein, Brad Reback at Stifel, Kirk Materne at Evercore — and their questions in earnings calls tend to be technical and specific. Whether they engage directly with the off-balance-sheet characterization of data center commitments in the Q2 calls will be a leading indicator of how seriously this topic is being taken in institutional investment circles.

What Honest Financial Engineering Would Look Like

I want to be precise about what I am criticizing, because the argument is not that these companies should stop building data centers or should not use long-term financing structures. Long-term infrastructure investment requires long-term financing, and that is normal and correct. Power plants take 10-15 years to build. Data centers have 20-year useful lives. Financing them with 20-year commitments makes sense.

The problem is the absence of consolidated disclosure. A company that has $1.35 trillion in reported debt and $1.65 trillion in additional off-balance-sheet obligations should be presenting investors with a consolidated view of its total obligation structure, even if GAAP does not require it. Some companies do this voluntarily — Berkshire Hathaway, for instance, goes significantly beyond GAAP minimum disclosure requirements when it comes to insurance float and operating business obligations. The tech giants are not doing this.

The specific disclosure improvement that would address the problem is straightforward: a “total future commitments” table in the financial statements that consolidates balance sheet debt, operating lease obligations, purchase commitments, and PPAs into a single multi-year view showing years 1, 2, 3-5, and thereafter. Most companies already disclose each of these categories separately in various footnotes. The improvement is consolidation, not new disclosure. The reason it is not being done is not complexity; it is that the consolidated number is larger and less flattering than the disaggregated pieces look individually.

The best analogy from an engineering perspective is word counting methodology. If you write a 5,000-word article but use a word counter that systematically misses 30 percent of words — particularly words in tables, code blocks, and footnotes — you might report 3,500 words and believe you are meeting a 4,000-word threshold. Technically, by the rules of that counter, you met the requirement. But the underlying reality is different. The off-balance-sheet reporting problem is the financial equivalent: a measurement system that produces compliance with technical requirements while obscuring the true magnitude of what is being measured. The numbers exist. They are real. They are just being presented through a lens that makes them appear smaller than they are.

The Bubble Mechanics Have Changed From Previous Cycles

Previous technology investment bubbles — the dot-com era of 1998-2001, the 2021 zero-interest-rate SPAC frenzy, the crypto peak of 2021 — were primarily characterized by speculative equity valuation disconnected from revenue fundamentals. Companies with minimal revenue were valued at tens of billions based on narrative projections. The correction, when it came, was primarily an equity repricing: valuations fell, but most of the underlying companies did not carry crippling debt loads because they had financed themselves with equity (which does not require repayment) rather than debt.

The AI infrastructure situation is different in a specific structural way that makes it harder to see and harder to unwind cleanly. The companies involved are profitable. They have real revenue at genuinely large scale. They have products that enterprise customers pay for and build critical operations on top of. The risk is not that the underlying business is fraudulent or that the revenue is imaginary; it is that the rate of infrastructure commitment has outrun the rate of revenue generation in ways that are being obscured by accounting treatment that minimizes the apparent obligation load.

This is harder to model than a straight equity bubble, and it is harder to correct. When dot-com equity valuations fell in 2000-2001, the correction was painful but structurally simple: stock prices went down. The companies with actual revenue survived; the ones without didn’t. The companies that did survive were not left with massive long-term fixed obligations that they had to continue servicing regardless of revenue levels, because they had financed with equity, not multi-decade infrastructure commitments.

A company that has signed 20-year power purchase agreements and build-to-suit data center contracts does not have the option of simply scaling back investment if AI revenue grows more slowly than projected. The commitments are legally binding. The question becomes: can you renegotiate, sublease, or find alternative uses for the capacity you committed to? In some cases yes. In aggregate, at $1.65 trillion in scale, the answer is more complicated.

The Broader Industry Context: What the Startup Founders Letter Gets Wrong

The same day the Nikkei investigation was published, a letter from startup founders urging the US government not to restrict access to Chinese open-weight AI models hit 568 points on Hacker News — slightly more than the debt story. The two stories are connected in a way that the startup letter does not acknowledge.

The founders’ argument is that cutting off access to Chinese open-weight models like Qwen and DeepSeek would harm American competitiveness by restricting the tools available to US developers. That argument has merit on its own terms. But there is a secondary consequence of open-weight AI proliferation that the letter sidesteps: open-weight models significantly reduce the addressable market for proprietary AI APIs.

If a startup can run Qwen3-72B on their own infrastructure — either on-premise or via a commodity GPU cloud — and get results that are 80-90 percent as good as GPT-4o for their specific use case at 20 percent of the cost, the number of enterprise customers who sign large multi-year Azure OpenAI or AWS Bedrock commitments shrinks. That is not hypothetical; it is already happening. The growth in self-hosted and open-weight AI deployments is real, and it represents a meaningful alternative to the proprietary API services that the five companies are building $1.65 trillion in off-balance-sheet obligations to support.

I am not arguing that the startup founders are wrong to oppose restrictions on open-weight AI. I am arguing that the AI infrastructure investment cycle and the competitive threat from open-weight models are two parts of the same story — and the financial reporting opacity makes it harder to see how exposed the major cloud providers actually are to that competitive dynamic.

Falsifiable Predictions

Predictions that are unfalsifiable are not predictions; they are rhetoric. Here is what I expect to happen, stated with enough specificity that I can be clearly wrong:

By end of Q3 2026 earnings season (October-November 2026), at least two of the five companies named in the Nikkei investigation will face specific analyst questions on earnings calls about off-balance-sheet data center commitments, and at least one will provide additional voluntary disclosure about total commitments that goes beyond what appeared in the most recent quarterly filing. If no analyst asks about this on any of the four Q2 earnings calls, I will update my view of how seriously the institutional investment community is taking the risk.

By mid-2027 (12 months from now), at least one of the five companies will announce a partial rationalization of its AI data center buildout — either canceling committed capacity, entering into sublease agreements for excess space, or announcing write-downs of assets from an earlier construction phase — citing a mismatch between deployed capacity and actual AI workload demand. I am putting the probability of this at approximately 65 percent based on historical patterns from previous infrastructure overbuilding cycles.

The SEC will open a formal inquiry or publish guidance specifically addressing off-balance-sheet data center and power purchase agreement disclosure practices by at least one publicly traded US tech company before the end of 2027. The Nikkei investigation, the Bloomberg follow-up, and the Hacker News attention represent a level of public visibility that tends to attract regulatory attention. I would be surprised if there is no regulatory response.

If all three of these predictions turn out to be wrong — if AI revenue growth remains strong enough to absorb the infrastructure commitment load, if the off-balance-sheet obligations prove to be sustainable at this scale, and if Q2 and Q3 earnings show no sign of financial stress — then I was wrong about the severity of the risk. I will update accordingly and say so. But the mechanism described by Nikkei is structurally real, the numbers are very large, and the historical precedents for this pattern resolving without any rationalization at all are limited. The question is not whether there will be a correction. It is how large and how disruptive it will be, and whether anyone outside the financial reporting footnotes will see it coming.

The AI race has a financial architecture that its most visible participants are not fully disclosing. That is the story. Everything downstream of that — the competitive dynamics, the infrastructure decisions, the platform choices for engineering teams — is a consequence of that single fact.