The AI boom is racing its balance sheet

In October 2026, Microsoft opened a voluntary separation buyout program to about 60,000 employees. It covered the United States, Canada, the United Kingdom, and India, more than a quarter of its global workforce. The Copilot division was included too, the same division that was still hiring this spring when other departments paused recruitment. The package included severance, health insurance, and stock vesting arrangements. Employees did not need to be bottom-ranked in performance reviews, and they did not need to meet a certain age or tenure threshold. General managers and some vice presidents were also eligible to apply.

On the same balance sheet, another thing was happening at the same time: Microsoft expected capital expenditure of about $175 billion for 2026. Later, because some GPU contracts were delayed, it lowered the figure to $170 billion. The order of magnitude did not change.

The workforce budget was contracting while compute investment was expanding.

Putting these facts together produces an uncomfortable equation: Microsoft earned $133.7 billion in net profit in the past fiscal year, while planning to pour $170 billion into data centers. The money saved by the buyout program is a rounding error next to capital expenditure. Its signal value exceeds its financial value. The company is freeing up room for AI investment, and the way it frees up room includes asking people to leave.

This is not just Microsoft’s problem.

One

OpenAI is the best sample for this investment. According to audited financial documents, OpenAI’s 2025 revenue was about $13.07 billion, total costs and expenses were $34 billion, and its operating loss was $20.92 billion. If one includes the one-time non-cash accounting charge arising from the company’s transition from a non-profit structure to a for-profit entity, the net loss expands to $38.5 billion. Excluding that non-cash item, the underlying operating loss was about $8 billion.

So “burning more than $20 billion a year at the operating level” holds. Equating that directly with “losing $20 billion in cash” would overstate the actual pace of cash burn. The difference between the $8 billion underlying operating loss and the $20.9 billion operating loss consists of depreciation, stock-based compensation, and other non-cash items. These items do not directly consume cash, but they reflect resource commitments.

Even so, the scale is still striking. In 2022, the total output value of Japan’s anime industry was about $20 billion. OpenAI’s operating loss for one year is equivalent to the global output value of a mature content industry. AI has value. The problem is speed: leading large models are commercializing far more slowly than capital is being spent.

Large language models themselves do not directly make money. Their economic significance appears more in pulling the chip industry, indirectly driving cloud service revenue, and supporting valuations of traditional businesses. The question is whether this indirect pull can cover trillion-dollar-scale capital expenditure. That remains unanswered.

Two

Video generation models are currently the category with the shortest commercialization path and the most visible revenue.

Kuaishou’s Kling AI broke $20 million in monthly revenue in December 2025, equivalent to an annualized revenue run rate of about $240 million. The veteran player Runway had ARR of about $90 million in mid-2025, with year-end expectations of $265 million. Kling and Runway were roughly in the same tier in terms of 2025 revenue scale.

But the key is scale. Video generation models do make money directly, but even leading products generate only hundreds of millions of dollars in annual revenue. OpenAI’s operating loss for one year is in the tens of billions of dollars, and Microsoft’s annual capital expenditure alone is in the hundreds of billions.

Video models do make a profit. But that profit is far from sufficient to become the economic pillar supporting the entire AI narrative. It is more like a showroom proving that AI can charge money, not an engine that can pull the entire industry to payback.

As for indirect profitability, the logic does require bundling with industrial manufacturing. In the Chinese market, some deployment cases are visible: AInnovation’s AI plus manufacturing revenue accounted for 80.9% in 2025; TCL created more than 1 billion yuan in total benefits through AI applications in 2025; Shandong Haihua’s salt chemical large model first-phase investment was more than 32 million yuan, with annual benefits of about 23 million yuan.

But placing these cases in the U.S. context produces a different picture. In 2025, data center construction spending surged 18%, while new factory construction shrank 2.5%. U.S. AI capital expenditure has not flowed into AI-driven upgrading of manufacturing. It has flowed into computing infrastructure itself. Deindustrialization means that AI bundled with industry, as an indirect profitability path, lacks enough industrial carriers in the United States.

Three

The clearest part of the chain is the hardware geography: the beneficiaries of U.S. AI capital expenditure are concentrated in East Asia.

The most profitable and most capacity-constrained links in the AI chip supply chain are in Asia. TSMC controls nearly three-quarters of the global wafer foundry market. SK Hynix and Samsung together hold about 80% of the global HBM chip market. Nvidia’s AI chip production depends heavily on TSMC’s advanced packaging, and Nvidia alone has booked more than 70% of TSMC’s capacity in this category. Nvidia’s earnings releases can directly drive linked stock price gains for Asian suppliers such as Samsung Electronics, TSMC, and SK Hynix.

Broadcom’s custom AI chip business is also based in the United States and is sizable: AI-related revenue was about $20 billion in fiscal 2025, up 65% year over year; its custom AI chip order backlog has exceeded $73 billion. AMD’s data center AI GPU revenue is far smaller than Nvidia’s, about $4.3 billion in Q3 2025 versus Nvidia’s $41.1 billion in single-quarter data center revenue, but it is also ramping up.

So among U.S.-based AI chip design companies, Nvidia captures the vast majority of profits, but Broadcom and AMD are also taking a share of the incremental growth. Profits in manufacturing and memory are mainly left in South Korea and Taiwan.

This means the domestic multiplier effect of U.S. AI capital expenditure is overestimated. Money flows out of the balance sheets of U.S. tech giants, turns into TSMC wafers, SK Hynix HBM, Samsung packaging capacity, then into Nvidia GPUs, and finally into U.S. data centers. The parts of the U.S. economy that steadily profit are a small number of chip design companies and cloud service giants, not a broad manufacturing ecosystem.

Four

Back to Microsoft.

60,000 buyouts, covering 27% of employees, with the Copilot division included. In 2025, Microsoft had already cut about 15,000 positions. At the same time, capital expenditure was $170 billion to $175 billion.

This cannot be summed up with the phrase “cost reduction and efficiency improvement.” The premise of cost reduction is that the cost structure has a problem, but Microsoft earned $133.7 billion in net profit in the past fiscal year. Its profitability is not the problem. The logic of the buyout program is to free up workforce budget and shift it into compute budget. The fully loaded cost of an employee at a company like Microsoft is roughly $200,000 to $300,000 a year. If all 60,000 took buyouts, assuming average compensation of $100,000, that would be a one-time expense of $6 billion. The workforce cost saved would be about $12 billion to $18 billion a year. That sounds like a lot, but next to $170 billion in capital expenditure, it is still a rounding error.

So the money saved by the buyout program is not the point. The point is the signal: the company is telling the market that it is willing to change its workforce structure to make way for AI investment. For shareholders, that signal is positive. It shows management is serious about controlling costs. For employees, it is another matter.

Using cash flow and profits from traditional businesses, Azure and Office, to pay for AI infrastructure investment that cannot yet sustain itself, while offsetting part of the cost pressure through workforce contraction. This is not just Microsoft’s problem. It is the common situation of all giants heavily invested in AI. They cannot stop investing, because once they stop, shareholders will flee and the bubble will burst. They also cannot prove that current investment can generate enough returns within an acceptable time window.

Five

AI-related debt is expanding.

As of the second quarter of 2026, AI-related investment-grade bond issuance during the year had reached $330 billion, 10 times the full-year 2023 level. The combined debt of the Magnificent Seven U.S. tech giants rose from about $399 billion in 2025 to $578 billion by mid-2026, an increase of nearly 45%.

The financing structure is shifting from spending your own money to borrowing. In the past, tech giants mainly used operating cash flow to pay for capital expenditure. The scale of AI data centers has become so large that they must increasingly rely on the debt market. Bloomberg reported that record debt accumulation is the biggest difference between this round of AI investment and previous cycles. The use of debt and the widespread circular nature of financing transactions, where suppliers and customers finance each other, have introduced risks that did not exist before.

The credit market has already sent warning signals. Since the summer of 2025, CDS spreads for AI-related companies have widened significantly. Default risk at CoreWeave and Oracle has drawn particular attention. Market commentary has turned cautious: lenders are getting timid.

A more accurate formulation of “the debt market has no money left” may be that cheap money is drying up, rather than there being absolutely no money to lend. UBS’s estimate at the end of 2025 still held that 80% to 90% of tech giants’ planned capital expenditure comes from operating cash flow. Debt financing is incremental rather than the main body. Borrowing costs are rising and conditions are tightening, while the return timetable for AI investment remains vague.

Gartner predicts that by 2030, AI data centers will need to generate cumulative revenue of $8.5 trillion to cover capital expenditure and achieve expected returns. Revenue during the same period is expected to be no more than $3.6 trillion. This gap is the core quantitative basis for the bubble argument.

Six

Stringing the clues together: huge losses at leading large model companies, narrow directly profitable scenarios, hardware profits flowing to East Asia, workforce costs being squeezed, and financing shifting from cash to debt. All of these have support in public information.

The better description is a race between the return timetable and financing costs.

Capital expenditure is still running at high speed. Stopping would mean the immediate collapse of shareholder confidence and valuation support. The patience and pricing of the debt market are becoming tighter constraints in this race.

Video models are making money, but the scale is too small. AI plus manufacturing is being deployed, but the United States lacks manufacturing carriers. Chip demand is exploding, but manufacturing and memory profits remain in South Korea and Taiwan. Microsoft is making money, but it is paying people to leave while pouring hundreds of billions into data centers.

Each fact looks reasonable on its own. Together, they point to one question: can the returns created by the U.S. AI industry catch up with the investment it has already committed before financing costs crush the balance sheet?

Wall Street has no answer. Silicon Valley has no answer.

Balance sheets will answer. Microsoft’s next earnings report, with capital expenditure versus headcount, Oracle’s debt maturity schedule, and CoreWeave’s CDS trend: these specific numbers will give signals earlier than any analyst’s judgment.