The AI Bubble Bull-Bear Debate and the Compute Capital Cycle, as Seen Through Alphabet's Earnings

**Key points**

**1. Alphabet's earnings are a mixed bag**

Cloud revenue rose 103% and Gemini users reached 1 billion, but capital expenditure rose 114%, free cash flow was negative $9.07 billion, and long-term debt rose to $125.1 billion; net profit was dragged down by an $8.66 billion investment loss.

**2. The bubble watershed is external payment and ROIC**

Whether free cash flow is positive or negative cannot be judged on its own; the return on new computing capacity must exceed the cost of financing over the long term; the $561 billion cloud backlog is contracts, not cash, and under circular financing, independent payment by end customers is what matters.

**3. Technological success does not equal valuation safety**

Railways, telecoms and Cisco show that late-arriving demand, profit migration, asset specificity and depreciation can devastate investors; technological success and valuation losses can coexist.

**4. China's computing capacity needs to guard against mismatches**

Intelligent computing is expanding fast, but we need to guard against mismatches between construction and use cases, hardware and software, and general capabilities and industry demand; the focus should shift from investment promotion to asset management, with assessment based on billability, renewability and industrial returns.

**5. The crash stems from future revenue being priced in advance**

This round of crash stems from booking future revenue in today's ledger ahead of time; going forward, watch RPO collections, the gap between capital expenditure and operating cash flow, inference price utilization, Chinese application renewals and power/software constraints. 2026-11-03 07:55 The Great AI Bubble Bull-Bear Debate: The Root of This Crash and the Lessons Jinduan Quick View

This article comes from the WeChat public account: Jinduan, author: Mu Yang

On October 28, Alphabet (Google's parent company) released its third-quarter report. Revenue grew 25% year on year, cloud business revenue doubled, and operating profit grew 30%. Looking only at these figures, this is a report card that could not be better.

But at the close that day, the stock price fell 12% overnight.

The same company, the same set of numbers, produced a huge divergence in the market: some stared at the cloud business's 103% growth rate, some stared at negative $9.07 billion in free cash flow; some calculated revenue, some extrapolated debt; bears said "the bubble is beginning to burst," bulls said "the market is hugely wrong."

Almost no AI participant denies that AI is rewriting the way software, advertising, industry and organizational production work. But the divergence has indeed become abrupt right now: can the money needed for this round of construction ultimately be earned back from the new revenue AI creates; before the answer arrives, who is qualified to keep spending shareholders' money?

01

Why does one earnings report read as two worlds?

The most confusing thing about Alphabet's earnings report is that it contains four mutually contradictory facts at the same time.

First, the growth is real.

Quarterly revenue was $127.577 billion, up 25% year on year; Google Cloud revenue was $30.709 billion, up 103% year on year, cloud business operating profit was $11.1 billion, more than triple the same period last year, and the margin rose to 36.2%. Gemini app users reached 1 billion, and first-party model APIs process nearly 32 billion tokens per minute, more than four times a year earlier. AI is not a story stuck in a slide deck; it is already driving real usage.

Second, the pressure on the traditional cash cow is also real.

Google Search & other revenue was $63.350 billion, up 12% year on year. Management acknowledged on the call that AI Mode queries are growing rapidly, but monetization is still low. Search query volume growth hit a five-year high, but monetization lagged behind.

Third, the pressure on cash flow is also real.

Operating cash flow for the quarter was about $42.3 billion, capital expenditure was about $51.4 billion, free cash flow was negative $9.07 billion, and capital expenditure grew about 114% year on year. The earnings call disclosed that the company's long-term debt had risen from $21.6 billion a year earlier to $125.1 billion. The cash cow is still producing milk, but the bill for building the pasture has, for the first time, grown large enough to change the company's constitution.

Net profit for the quarter was $26.170 billion, lower than the same period last year; other income turned from positive to negative, and the equity investment portfolio recorded a loss of about $8.66 billion. Earnings per share missed expectations, and it is not entirely "computing investment dragged down profit"—investment gains and losses accounted for an important proportion. Therefore, directly equating $2.12 per share with "the failure of the AI business model" is not rigorous.

The bear camp's leaders summarized this set of numbers as: cash flow turned negative, buybacks stopped, reliance on financing, and RPO may not be fulfilled.

The bulls seized on another point: cloud revenue doubled, profit tripled, and negative cash flow is merely an accounting timing difference of investment preceding return. Some bulls even estimate that the return on newly invested capital is about 20%, clearly higher than the cost of capital.

Neither side fabricated facts. They merely chose different stretches of the same river.

In fact, free cash flow is the most easily abused indicator. Its accounting relationship is simple: free cash flow ≈ operating cash flow − capital expenditure. Operating cash flow shows how much money the existing business can collect today, and capital expenditure shows how much the company has invested for the future. A negative difference may mean living beyond one's means, or it may mean it is using today's strong profits to build tomorrow's larger capacity. Looking only at whether it is positive or negative proves nothing.

What really needs to be explored are two other questions: what assets are formed after this money is invested; and for every yuan invested in these assets, how many dimes can be earned back each year in the future? This is precisely the relationship between return on capital and cost of capital.

If the after-tax operating profit generated each year by new data centers, chips, and networks, divided by cumulative invested capital, can remain above the cost of financing over the long term, then today's negative free cash flow is the prelude to value creation; otherwise, it is turning a good business into a bad balance sheet.

There is also an easily overlooked distinction here: ROIC should look at the return on new investment, not the average profit margin of the old business. Alphabet's past high returns from search advertising do not prove that today's new AI capital expenditure will naturally earn the same returns. An advertising company becoming an asset-heavy infrastructure company means valuation can no longer rely only on the price-to-earnings ratio; it must also look at how much sustainable cash each yuan of new capital can ultimately leave behind.

The bulls have grounds for confidence here: cloud business operating profit is rising rapidly, and the cloud backlog disclosed on the earnings call is about $561 billion, with slightly more than 55% of that revenue expected to be recognized over the next 24 months.

The so-called circular financing can be explained clearly with one example: a chip company invests in an AI lab, the lab uses the money to purchase computing power from a cloud provider, and the cloud provider then buys GPUs from the chip company. All three parties can announce cooperation, with revenue, orders, and valuations rising together, but how much real money external customers actually paid becomes harder to identify.

History has already rehearsed this scene.

A hundred years ago, in the 1920s, American electric utilities relied on layered pyramid holding structures to hold stakes in one another and guarantee one another, packing the same capital repeatedly into different companies' financial statements. Samuel Insull's utility empire once magnified the cash flow of a single power grid into the valuations of dozens of holding companies. After the crash of 1929, the Public Utility Holding Company Act of 1935 was enacted specifically to dismantle this layer of structure.

Previously, the Bank for International Settlements warned of risks in its Annual Economic Report 2026, stating that the opacity of financing in the AI industry intensifies financial fragility, with circular financing particularly worthy of vigilance; several types of arrangements already publicly discussed in the market—for example, Nvidia investing in OpenAI and discussing guarantees for its data center leases, Google providing support for Anthropic-related data centers, and AMD pairing chip deployment agreements with warrants—are in essence the same ancient impulse as the electric power holding pyramids: using structure to cash in the future in advance.

But it must be pointed out that the scale, constraints, and execution of these transactions are not entirely the same, and it cannot be said broadly that they are "all fake revenue." Circular financing itself is not a sin. In the history of railway, telecommunications, aviation, and energy construction, it is not rare for suppliers to help customers finance: external demand truly exists, and it can accelerate the formation of infrastructure; if demand relies mainly on continuous transfusions from the financier, then it is delaying the test of demand.

The market panics one moment and buys back the next, and in essence it is also fighting over this point: it is not only repricing infrastructure companies, but also rearranging "who ultimately gets the value of AI": who can survive the peak of capital expenditure, who can turn usage into sustainable charges, and who will be replaced by large models.

One moment the market fears CapEx is too large, the next it fears AI is not fast enough; this is not actually contradictory—it is pricing both "overinvestment" and "being replaced" at the same time. What is truly dangerous is using only one of these answers to explain all companies.

02

Technology can win, and valuations can still lose

When talking about the AI bubble, the easiest case to bring up is the Nasdaq in 2000: the index fell about 78% from its high, and countless .com companies disappeared. But comparing only index declines yields almost no useful conclusion. Today's large technology companies have profits, users, and global cash flow; they are two different species from the companies back then that had only page views and no revenue.

A more reasonable reference is the infrastructure bubble. And before the telecommunications bubble, there was an even older sample: railways.

In the 19th century, the United States paid capital far beyond contemporary imagination to build railways. By 2000, annual capital expenditure on U.S. telecommunications equipment was about $120 billion, construction relied heavily on debt, and total industry debt reached about $1 trillion by 2002. Railways had rehearsed all of this 150 years earlier at greater cost: duplicate construction, competitive track laying, land speculation, multiple waves of bankruptcies, until the Panic of 1893, when J.P. Morgan stepped in to restructure. Railways did ultimately reshape America, but most early investors did not live to see the harvest.

Telecommunications bubble just performed this script once again. Demand did not arrive at the imagined speed, WorldCom and Global Crossing fell one after another, and only a very small proportion of the fiber laid during the boom was actually lit. Cisco's story is especially painful: in March 2000, its share price was $80.06 and its market value about $500 billion; by October 2002, the share price was about $8.60, a decline of about 88%. Over the same period, the company's revenue did not collapse proportionally, and later the internet did change the world as expected.

Being right about technology and suffering heavy stock losses can absolutely happen at the same time.

This history offers three lessons for today:

First, the return on infrastructure depends not only on whether the technology is useful, but also on the time gap between the pace of construction and the arrival of demand. If capacity arrives first while customers are still on the way, financing costs will blow up.

Second, who the profits belong to. Early telecommunications builders were not necessarily the final winners of the internet; equipment makers can be repriced because of customer debt problems, while later platform companies take more value instead. The same is true for AI: chipmakers, cloud providers, model companies, and application companies growing at the same time does not mean they will share final profits in the same proportion.

Third, asset specificity. Once fiber is laid, it is hard to repurpose for other uses; general-purpose computing power, specialized chips, and power facilities in data centers share the same attributes. The faster chips iterate, the more likely depreciation becomes a hidden examiner of profits. AI assets with high book profit margins today may, if their efficiency falls behind and prices decline in three to five years, have an actual economic life shorter than their accounting depreciation period.

From an economic perspective, this is a typical marginal return proposition. That the AI market will be large in the future does not mean all of today's investment will pay off; aggregate demand is one thing, and the marginal return of adding one more data center, one more chip, or one more industry large model is another. Capital cycles are most prone to losing control amid the consensus that "the market is huge," because every company believes that if it does not invest, it will lose the next round.

The result is a prisoner's dilemma: if rivals expand and I do not, I may lose the market; if everyone expands, the industry may end up with overcapacity. The rationality of individual companies adds up to industry-wide overinvestment.

To break it, three indicators should be put into one pool and observed continuously:

Incremental return on invested capital (operating profit from new investment divided by new invested capital, rather than the entire company's historical profit margin);

Price and utilization (if revenue growth comes from price increases, it must be sustained by utilization improvement and cost reduction; a smart computing center with a high idle rate is merely expensive fixed assets no matter how large it is);

Cash payback period (long-term contracts can ease uncertainty but cannot eliminate it; the longer the payback period, the more important customer credit, technology substitution, and financing costs become).

For this reason, the bulls emphasize cloud business profits and vertical integration, while the bears emphasize end demand and financing structure, and both sides can actually be partly right at the same time: the self-built computing power of leading companies may earn reasonable returns, while highly leveraged marginal projects may not. A bubble does not have to occur evenly across the entire industry; it often bursts first in the most fragile projects and the most optimistic valuations.

Putting financial reports, transaction structures, and history together, we do not agree with two extreme views: "AI is just a scam" cannot explain Google Cloud's revenue doubling, model invocation volume growing severalfold, or AI's entry into advertising, office work, and industrial processes; "as long as AI is useful, all investment is reasonable" evades capital costs, depreciation, and payback periods.

A more precise perspective may be: which revenue comes from external customers paying repeatedly, and which revenue mainly comes from capital and contracts confirming each other. The more of the former, the more construction can weather cycles; the higher the share of the latter, the more one must guard against chain reactions caused by changes in financing conditions.

Of course, for a company like Alphabet, one more layer must be added: its core search business is still generating huge cash, so the key question is not "will it collapse," but whether new businesses can earn returns no lower than the old business, and under what framework the market is willing to price it.

03

China mapping: the same capital cycle, two balance sheets

America's contradiction is the relationship among massive capital expenditure, complex financing arrangements, and the realization of end demand; China's computing power construction is also expanding rapidly, but the driving forces also include the national computing power network, East Data West Computing, telecom operator investment, local industrial policy, and domestic substitution, and the sources of funds, project types, and assessment methods are all different.

East Data West Computing has formed 8 major hub nodes and 10 data center clusters, and the intelligent computing scale built at the 8 major hubs exceeds 80% of the national total; new direct investment in computing network construction during the 15th Five-Year Plan period is expected to reach 4 trillion yuan; in the first half of 2026, China's intelligent computing scale reached 2,185 EFLOPS, up 177% year on year.

These figures can show the intensity of construction, but they cannot automatically explain usage efficiency. What China really needs to guard against, besides "borrowing to build computing power," is three types of more local mismatches:

Mismatch between construction and scenarios (projects are obtained first, while models and applications are sought later; smart computing centers become investment-promotion scorecards but lack enough workload to fill them);

Mismatch between hardware and software (growth in domestic chip shipments is important, but if frameworks, toolchains, model adaptation, and developer services cannot keep up, procurement scale does not equal usable computing power);

Mismatch between general capabilities and industry needs (what factories may need is not a larger general model, but specialized models that can connect to quality inspection, production scheduling, energy management, and after-sales service; parameter scale does not directly create cash flow).

Here lies a meaningful contrast:

In the 19th century, American railways were driven by private capital and expanded through land grants and bonds, and ultimately consolidated after repeated construction and debt bankruptcy; China's high-speed rail, power grids, and computing power networks follow a state-led, unified planning path. The driving mechanisms differ, but the capital cycle's exam question is the same: after it is built, is there enough demand to use it and earn the money back?

Historically, state-led infrastructure has also had stories of capacity preceding demand, except the risk is not called "bubble bursting" but "insufficient utilization" and "hidden debt."

Therefore, China's problem cannot be simplified into "whether to follow the investment," and what needs to be assessed is: whether each additional unit of computing power investment can form local industry capabilities that can be billed, renewed, and exported.

Electricity, equipment, and engineering get paid first. The expansion of smart computing centers first brings orders for servers, networks, cooling, power, and construction; this layer most resembles the "shovel sellers" of infrastructure cycles, with high order visibility and the greatest susceptibility to changes in construction pace. It earns money from capital expenditure, not from application renewals.

Chips and computing power operations come next. Domestic substitution provides additional demand, but assessment cannot stop at procurement and shipments; what truly determines asset quality is utilization, unit computing power cost, software compatibility, and customer structure.

Models and platforms come after that. Growth in call volume does not equal platform profitability; inference costs, the share of free users, enterprise contract terms, and the pace of price declines determine whether this layer can generate stable gross margins. Open source can expand the scope of use, but open source itself is not a business model; hosting, enterprise services, and industry delivery are what determine the sources of cash.

Industrial applications in China are most likely to generate differentiated returns in the following areas: China has a complete range of manufacturing categories, dense industrial belts, and large amounts of real production data. In quality inspection, equipment maintenance, supply chains, drug R&D support, and government and financial back offices, there are clear labor costs and error costs. Whether AI can reduce rework, shorten downtime, and lower inventory answers the value question better than model benchmark scores.

Robots and intelligent terminals connect software capabilities to the physical world. They offer the greatest room for imagination, but also bring higher hardware investment, after-sales costs, and safety requirements. The size of the opportunity is no reason to skip the unit economics.

Arranged this way, it becomes clear that if AI policy rewards only "building it," resources will be pushed toward the first two layers; if it begins to assess "using it" and "earning it back," returns will remain in the last three layers. The true maturity of an industrial chain is not that every link is hot, but that cash can flow all the way back from the application end to the computing power and chip ends.

In 2026, the policy language on new quality productive forces already places "fully implementing the 'AI+' initiative" together with "not shifting from the real to the virtual, not seeking excessive scale, and not creating bubbles." The essence of this combination is precisely to shift assessment from investment scale to industrial results.

For local governments, smart computing centers should shift from "project attraction" to "asset management." New projects should at least answer these questions: Are there real local customers? What is the expected utilization rate? Can electricity and network costs be covered? Can domestic software and hardware be used by enterprises with a low threshold? The significance of Shanghai and Zhuhai issuing computing power vouchers is not to hand out one more subsidy, but to channel the real demand of small and medium-sized enterprises to existing computing power and avoid building new capacity while existing capacity sits idle.

For traditional companies in industrial belts, what is most worth doing may not yet be securing a larger general-purpose model, but organizing industry data, process knowledge, and on-site problems into reusable specialized capabilities. Only after one factory proves that AI reduces defect rates, shortens line changeover time, and makes veteran workers' experience replicable will ten factories be willing to renew. The premise for shared prosperity in an industrial belt is not that everyone buys the same expensive system, but that trial costs are lowered, validation results are shared, and specialized models can migrate between similar processes.

For technology companies that are themselves driving forces, self-reliance and controllability cannot be assessed only by "whether there are domestic chips," but also by migration costs: how much code an enterprise must change to move from one chip to another, whether domestic frameworks can retain developers, and how much electricity each effective task consumes after model deployment. China's core AI industry scale has exceeded 1.2 trillion yuan, intelligent computing power has reached 2,592 EFLOPS, and there are more than 6,600 AI enterprises. Scale has already been achieved; the harder task in the next stage is to turn this scale into stable returns.

For investors, it is most important to distinguish "three types of benefit": increased orders, domestic substitution, and operational improvement. The first two can bring revenue; only the third shows that AI has truly entered corporate income statements. Looking only at project wins and computing power scale makes it easy to misread "construction benefits" as "long-term profitability."

04

This debate will not be ended by some article proclaiming a "burst bubble" or by a single stock price rebound. It has observable verification paths:

Watch the external realization of RPO: over the next few quarters, the speed at which cloud backlog converts into revenue, customer concentration, and collections matter more than total contract value.

Watch whether the gap between capital expenditure and operating cash flow narrows: if AI investment begins to generate returns, free cash flow will improve before grand narratives do; if the negative value continues to expand, the market will demand clearer evidence of ROIC.

Watch inference prices and utilization: falling prices are not inherently bad news; if costs fall faster and utilization is higher, profits can still grow; if prices, utilization, and depreciation deteriorate at the same time, revenue growth will not save returns.

Watch renewal rates on China's application side: whether manufacturing enterprises continue paying in the second year, and whether industrial belts move from pilots to regular procurement, are more persuasive than the number of launch events.

Watch power and software constraints: data center electricity use, domestic framework adaptation, and cross-chip migration costs will determine whether China's computing power investment ultimately becomes an asset or a burden.

The lesson left by 2000 was never "do not invest in new technology," but that technological prospects, capital returns, and valuations must be answered separately. AI can continue to change the world while also eliminating some projects with excessively high valuations and excessively low returns. What truly must be guarded against is not technology moving too fast, but booking the revenue of the next decade on today's ledger in advance. The root of this round of stampede-like crashes in global AI concept stocks comes precisely from this kind of "full pricing."

150 years ago, railway investors were also betting on the same future. They were wrong about only one thing: that future arrived late, and they could not wait for it. We may need to take this as a lesson. Channel: Finance and Economics

435 articles have been published on Huxiu. This content is authorized by the author for publication. 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[ ]