$1.65 Trillion in Off-Balance-Sheet Commitments and a 166-Fold Token Price Gap: Three Endgame Scenarios for the AI Capital Game

[[**US tech giants' profits rely on unrealized gains from equity investments, with off-balance-sheet hidden commitments reaching $1.65 trillion; the pace of AI commercialization lags far behind the expansion of capital expenditure, potentially triggering chain financial risks.**

**Key points**

**1. Profit magic masks cash outflow**

More than 85% of Alphabet's and Amazon's quarterly profits come from unrealized fair-value gains on equity investments, while free cash flow has turned negative; Microsoft is relatively healthy. GAAP accounting allows unlisted equity to be revalued at financing valuations, spawning a self-reinforcing loop of "invest → purchase computing power → valuation rises → financial statements look good → refinance"; once financing cools, the income statement will reverse.

**2. $1.65 trillion in off-balance-sheet debt is comparable to Enron**

The five major tech giants' off-balance-sheet long-term procurement, lease and SPV guarantee commitments total $1.65 trillion, exceeding on-balance-sheet liabilities. Meta's off-balance-sheet commitments are 2.8 times its on-balance-sheet liabilities; Oracle's have ballooned more than 30-fold in four years. Unlike Enron's fabricated revenue, the giants' cloud revenue genuinely exists, but profits are beautified, and circular flows within the industry chain deepen systemic risk.

**3. The AI monetization gap is as high as 5 times**

In 2026, the five giants' capital expenditure is expected to reach $769.2 billion, while annual AI application revenue is only about $200 billion; CICC estimates that $1 trillion is needed to cover costs. Sequoia points out that every $1 invested in computing power needs to generate $4 in application revenue to close the loop, and the current level is far from that. On the financing side, AI bond issuance has doubled, and underlying funding is under pressure.

**4. China-US Token price gap is 166 times, monetization paths diverge**

OpenAI output costs $30 per million tokens, while DeepSeek costs only $0.18, a price gap of 166 times. Overseas players take the high-priced enterprise customer route (75% of Anthropic's revenue comes from large enterprises), while domestic players take the low-price, high-volume calls plus overseas expansion route (73% of MiniMax's revenue comes from overseas). The competition between the two models is unresolved, but domestic large models can command a premium overseas.

**5. Most likely endgame: localized blowups with chain transmission**

English text is being prepared.

This article is from the WeChat public account: Skai Suisui Kan, author: Skai, original title: "In-depth | The $1.65 trillion hidden bombs and cabbage-price Tokens: an endgame simulation of an AI capital game"

Our Singapore partner Eric is a leading expert in AI industry research. Recently, after going through the financial reports of US tech giants, he said to me: "The more I look, the more something feels off."

In the summer of 2026, Alphabet and Amazon successively reported record quarterly profits of $112.1 billion and $62.6 billion, respectively. But after the reports were released, both companies' share prices fell instead.

A careful breakdown of the notes revealed: of Alphabet's hundred-billion net profit, about $98 billion came from upward revaluation of its equity investments in SpaceX and Anthropic; of Amazon's $62.6 billion profit, $53.4 billion came from fair-value upward adjustments on its Anthropic investment.

None of these profits involved a single cent of cash flowing into the company's accounts; they were all paper gains on the books.

Alphabet's Q2 2026 net profit was $112.1 billion, about 87% from paper gains on equity investments. Amazon's $62.6 billion net profit was about 85% from fair-value adjustments on its Anthropic investment. Meta's paper gains accounted for a relatively lower share, but its free cash flow plunged 91%, leaving only $784 million. Microsoft is the healthiest of the four, with paper gains accounting for less than 8% and free cash flow of $19.6 billion.

This profit magic is concentrated in Alphabet and Amazon.

It is not that companies earned cash and are unwilling to pay dividends; rather, book profits are merely numbers brought about by valuation changes, while the companies continue to make large capital expenditures on computing power. Excluding paper gains, Alphabet's core business operating profit was about $40.8 billion, and quarterly free cash flow turned negative for the first time, at -$5.9 billion, the first such occurrence in the company's decades since listing. Amazon's free cash flow turned negative over the past twelve months, with a funding gap of about $7.6 billion.

This playbook relies on the fair value measurement rules in US GAAP: when an unlisted company's equity undergoes a new round of financing, the investor can simultaneously revalue that equity; listed equity is adjusted at market price each quarter. This is fully compliant at the accounting level, but it has spawned a self-reinforcing capital cycle:

Cloud service providers invest in AI labs → AI labs use the financing to purchase cloud computing resources → cloud service providers recognize revenue → the valuation of the AI company equity they hold rises → the financial statements recognize large investment gains → impressive financial statements push up the share price → it becomes easier to obtain financing and continue to increase investment in AI companies.

2. $1.65 trillion in hidden commitments: leverage hidden outside the statements

Paper gains are only the first layer of appearance; the deeper risk comes from huge long-term contractual obligations outside the main body of the financial statements.

Nikkei Asia Review reviewed the annual reports of the five major tech giants Alphabet, Microsoft, Amazon, Meta, and Oracle and found that the total explicit debt recorded on their balance sheets is about $1.35 trillion, while off-balance-sheet commitments from various long-term leases and equipment purchase contracts are estimated at as much as $1.65 trillion.

The scale of such off-balance-sheet contractual obligations has already exceeded the debt recorded on the statements. If the statistical scope is expanded to include nine tech companies—Nvidia, Broadcom, SpaceX, and AMD—the combined long-term procurement and lease commitments total about $3 trillion.

It should be clarified here: this portion is not rigid arrears that have already come due and must be repaid, but long-term contracts signed by companies that will require performance payments over the coming years.

Hidden commitments mainly come from three categories:

First, data center leases that have not yet taken effect. The computer rooms have not yet been built and put into use, so the leases are temporarily not included on the balance sheet; the nine giants total about $1.2 trillion.

Second, long-term purchase orders for GPUs and servers. The hardware has not yet been delivered, but the companies have already signed irrevocable purchase agreements, totaling about $1.9 trillion.

Third, SPVs, special purpose vehicles. The giants set up separate legal entities to build data centers and purchase hardware, with the parent company contributing only a small amount of capital but committing to bear backstop risk. Meta's Hyperion data center partnership in Louisiana saw project costs balloon from $27 billion to $50 billion; Meta contributed only a 20% equity stake but committed to guarantee all losses if the partnership breaks down, and this obligation is not reflected on the balance sheet.

Meta has the largest scale of off-balance-sheet commitments, about $420 billion, nearly 2.8 times its on-balance-sheet liabilities. Oracle's hidden commitments exceeded $270 billion by the end of May 2026, expanding more than 30-fold in four years. Alphabet's off-balance-sheet commitments are estimated at about $250 billion.

Many people will associate this set of practices with Enron in 2001. Enron used SPVs to move huge debts off its balance sheet, while using mark-to-market accounting to recognize expected revenue for many years ahead in the year the contracts were signed, and ultimately collapsed.

There is an essential difference between the two: Enron's revenue was almost entirely fabricated. The cloud computing revenue of today's tech giants genuinely exists; Google Cloud and AWS really are selling computing power to customers. The problem is that profits are heavily beautified by unrealized gains on equity investments, real free cash flow continues to deteriorate, and buyers and sellers across the industrial chain are highly bound to each other, forming an internal loop.

3. An arithmetic problem that cannot be balanced

If we separate the input side and the revenue side, the gap is obvious at a glance.

Input side: the five giants' combined capital expenditure in 2026 is expected to reach $769.2 billion, almost double that of 2025, and institutions estimate that capital expenditure in 2027 may hit as high as $1 trillion.

Revenue side: CICC estimates that the AI application track needs to generate about $1 trillion in sustainable and stable revenue each year to cover all capital expenditure and debt costs. Currently, the combined annual recurring revenue, ARR, of leading U.S. model makers is about $200 billion. There is a 5x gap in between.

London Stock Exchange Group forecasts that the five tech giants' operating cash flow is expected to increase by $340 billion in 2027, but capital expenditure is expected to increase by $534 billion over the same period. For every additional $1 of operating cash flow, $1.57 of capital expenditure must be invested.

Sequoia Capital's calculation is even more direct: for every $1 invested in AI infrastructure, the application side needs to generate $4 in revenue for the commercial loop to work. The current ratio of industry investment to revenue is far from reaching this balance.

Pressure on the financing side is also amplifying. Morgan Stanley forecasts that global AI-related bond issuance in 2026 will approach $570 billion, doubling from 2025.

Most of the buyers of these bonds are pension funds and insurance institutions. A large amount of U.S. tech giant stocks is allocated in 401(k) retirement plans and index funds. AI bonds are transmitted layer by layer through institutions, and the ultimate underlying funding comes from ordinary people's pensions. The Bank for International Settlements calls this model shadow lending.

UBS has issued a warning: once the AI investment boom loosens, the probability of a U.S. recession reaches 50%.

4. The China-U.S. pricing gap: the same token, a price difference of 166 times

Having discussed the capital side, let us look at the monetization side. There is also a highly contrasting phenomenon in the AI industry: for generating the same million tokens, the price gap between Chinese and U.S. models reaches as much as 166 times.

Leading overseas vendors: Anthropic's annualized ARR is about $65 billion, and its adjusted operating profit turned positive for the first time in Q2 at $559 million. OpenAI's annualized ARR is about $60 billion, but for every $1 of revenue it creates, it burns $1.8 in costs.

Leading domestic vendors: ByteDance's Doubao has annualized ARR of about $4.2 billion and an annual loss of about 3.8 billion yuan; Zhipu's annualized ARR is about $1.6 billion, with an adjusted net loss of 1.964 billion yuan in the first half; DeepSeek is about $700 million and is the Chinese large-model company closest to breaking even.

The revenue gap between leading companies is about 10 to 15 times, but this is not simply a gap in user scale; the core difference comes from pricing power.

Price comparison: OpenAI GPT-5.5 output is $30 per million tokens, and the Pro version is $180. Anthropic Claude flagship input is $10 and output is $50. DeepSeek V4 Flash input is $0.09 per million tokens and output is $0.18 per million tokens. Zhipu GLM-4-Plus is about $0.7 per million tokens.

For the same consumption of 1 million tokens, OpenAI charges up to $30, while DeepSeek charges only $0.18—a price gap of 166 times.

OpenRouter data shows that domestic large models' weekly call volume has exceeded that of the United States for five consecutive weeks, but revenue scale is far behind. The customer structure differs enormously.

Anthropic derives 75%–85% of its revenue from large-enterprise API calls; financial, healthcare, and government clients are insensitive to price and willing to pay a premium for stability and compliance. As of June 2026, 34.4% of U.S. enterprises had procured Anthropic services, and its market share rose to 41% in June.

Domestic customers are mostly developers and small and medium startup teams, price wars are white-hot, and enterprise procurement budgets are limited. But there is a counterintuitive fact: when domestic large models go overseas, they find it easier to obtain high premiums.

In 2025, 73% of MiniMax's revenue came from overseas. San Francisco AI company Lindy switched its business traffic from Anthropic to DeepSeek, with no drop in performance, saving millions of dollars in computing costs each year; Coinbase selected Zhipu GLM-5.2 and Moonshot AI's Kimi K2.7 as its engineers' default models, cutting AI spending by nearly half.

China's large models do not lack a paid market; rather, domestic market prices have been pushed down to near the cost line, while overseas markets can command higher premiums.

Two monetization routes, with the outcome still undecided: the overseas route, pursuing high unit prices plus large-enterprise customers; the domestic route, pursuing low-price massive calls plus overseas monetization.

V. To whom the bill is passed: gains concentrate upward, costs are shifted downward

First, job displacement effects continue to occur. Goldman Sachs predicts that AI adoption will force about 9% of American workers, roughly 15 million people, out of their original jobs. In technology, consulting, and design industries, 10,000 to 15,000 new positions are being eliminated each month. In January 2026, the United States cut 108,000 jobs in a single month, with Amazon, Salesforce, and IBM all citing AI efficiency gains for layoffs. Stanford research shows that the employment rate of AI-related fresh graduates fell 16%, and the number of employed software developers aged 22–25 fell 20% relative to its peak.

Second, K-shaped consumption divergence is intensifying. Stocks account for 37% of U.S. household wealth, a record high. High-income groups rely on the tech-stock bull market to support consumption; low-income groups bear inflation pressure, with real incomes stagnating. Excluding healthcare and transfer payments, nonfarm payrolls shrank by an average of 41,000 per month in late 2025.

Third, computing power is becoming a new digital land rent. CICC estimates that the marginal production cost of tokens continues to decline, from $60 in 2022 to $0.06 in early 2026. But tokens are not homogeneous goods; for the same number of tokens, model output quality varies enormously. Closed-source models still hold a market advantage, and users are willing to pay a premium for high-quality models. Computing power vendors can rely on quality barriers to maintain pricing power and will not completely become public utilities.

At present, capital-driven AI has formed a self-reinforcing loop: Nvidia invests in AI companies, AI companies purchase Nvidia GPUs, Nvidia's revenue rises, and it continues investing in more AI startups; cloud vendors invest in AI labs, the labs purchase cloud computing power, cloud vendors recognize revenue, and unrealized gains on equity investments push up stock prices, enabling more financing.

Funds circulate within the industrial system. The business model has changed from "selling products to end consumers" to "selling computing power to each other within the industrial chain." This closed loop has a fatal precondition: real external enterprise demand continues to grow. MIT research shows that 95% of enterprises have not obtained quantifiable business returns after implementing AI. Once enterprise implementation falls short of expectations, the entire loop chain will break.

VI. Endgame scenarios: three possibilities

A killer product is born at the application layer, enterprise implementation accelerates on a large scale, and revenue growth catches up with capital expenditure. In 2030, the AI industry's annual revenue exceeds $1.4 trillion, the industry as a whole achieves break-even, and massive off-balance-sheet commitments are absorbed through growth. CICC judges that AI today is not a typical bubble, but capital investment is clearly running ahead of real demand.

AI implementation is slower than capital expansion, off-balance-sheet contracts mature in concentrated fashion, and asset impairments combine to erode corporate profits. In 2028, the external financing gap reaches $700–800 billion, tech giants are forced to cut capital expenditure, and upstream chips and servers come under simultaneous pressure. Deutsche Bank estimates that OpenAI may not turn cash-flow positive until 2030 at the earliest.

Endgame three: localized blowups, with risk transmitted in a chain (most likely scenario)

Not all giants will collapse simultaneously; instead, a problem will first emerge at some link: an AI lab's valuation is sharply cut, a data center project defaults, or a credit event occurs in AI bonds. The deep binding among giants through cross-shareholdings and mutual procurement will amplify localized risk and form chain transmission. UBS's 50% recession probability is a judgment made for this scenario.

My judgment: the third scenario has the highest probability. The problem is not whether AI technology has value, but whether the pace of AI commercialization can catch up with the pace of capital expenditure expansion. This is essentially an arithmetic problem, and the answer to the arithmetic is far more brutal than the narrative.

Final note

In this system, AI is both a product and a financing tool, an valuation anchor, and an accounting item on the balance sheet. It may ultimately give birth to technology that changes human civilization; but it is also possible that, before the technology matures, it first detonates financial-level risk.

This is a cold arithmetic problem.

$1.65 trillion in off-balance-sheet commitments and rock-bottom token prices are the two most glaring sets of numbers in this problem right now. Channel: Finance and Economics. This content is published with the author's authorization; the views represent only the author and do not represent Huxiu's position. If you have objections or complaints about this manuscript, please contact tougao@huxiu.com. Everyone who is changing and wants to change the world is on the Huxiu APP [ ]