The More Money AI Makes, the Less Money Entrepreneurs Have?

[[**AI is simultaneously creating valuation wealth and a cash-flow gap: tech giants' revenues are growing rapidly, yet Meta, Alphabet, and Alibaba have seen free cash flow turn negative. Whether this wealth can be realized depends on whether computing power investment can be converted into qualified output, profit, and cash.**

**Key points**

**1. Valuation wealth diverges from cash flow**

NVIDIA's FY2027 Q2 revenue was $96.2 billion, up 106% year over year, but Meta, Alphabet, and Alibaba recently posted free cash flow of approximately -$2.12 billion, -$5.9 billion, and -RMB 44.670 billion, respectively.

**2. AI adoption is high, but profit conversion is low**

Organizational adoption has reached 88%, and 80% acknowledge productivity improvements, but only 37% report a positive contribution from AI to EBIT, and only 6% of high-performing enterprises do. Saving work hours is not profit; it must be embedded in processes and converted into revenue or cost savings.

**3. Asset-heavy investment creates a time lag**

Five major tech companies' capital expenditure exceeded $400 billion in 2025 and is expected to rise another 75% in 2026; data center electricity use may grow from 485 TWh to about 950 TWh by 2030. Chips are paid for first, while profits are realized later; whether bought or leased, both affect depreciation and gross margins.

**4. Financing amplifies risk, and the market shifts toward return on capital**

This article is from the WeChat public account: Zhengjing She, author: Zhizhe Zhengshu

On one side, valuations of chip, cloud computing, and model companies keep rising, and entrepreneurs' equity wealth rises with them; on the other side, investment in data centers, GPUs, and power is surging, and the free cash flow of some tech giants has begun to turn negative.

NVIDIA's revenue in the second quarter of fiscal 2027 reached $96.2 billion, up 106% year over year, with data center business revenue of $89 billion and a gross margin of 75%. AI is certainly not a business with only stories and no revenue. [1]

But also within the AI wave, Meta's free cash flow in the third quarter of 2026 was negative $2.12 billion, Alphabet's free cash flow in the second quarter was about negative $5.9 billion, and Alibaba's net free cash outflow in the June quarter was RMB 44.670 billion. [2][3][4]

Companies are worth more, yet have less cash on hand.

Entrepreneurs' wealth is soaring, while companies instead need more bonds, leases, and equity financing.

Analysts at Zhengjing She note that the answer is not in the rich list, but in the three financial statements: the income statement, the cash flow statement, and the balance sheet.

Stanford University's 2026 AI Index Report shows that organizational AI adoption has reached 88%. On OSWorld, which tests real computer operations, the agent success rate rose from about 12% to about 66%. [5]

But conversely, even in structured test environments, agents still fail about one-third of tasks. They can achieve breakthroughs in advanced math and programming tasks, and they can also make mistakes on simple visual or common-sense tasks.

It is no longer just a chat tool and has begun to execute tasks in scenarios such as coding, customer service, marketing, and knowledge retrieval; but it is also far from universally becoming a "general employee" that requires no supervision.

Equally critical is that employees finding it useful and companies actually making money are two different things.

McKinsey's 2026 survey of 1,719 respondents shows that 80% acknowledge that AI improved personal productivity, but only 37% report that AI has a positive contribution to corporate EBIT. High-performing enterprises that truly meet both "AI contributes at least 5% of EBIT" and "obtain significant value" account for only 6%. [6]

Why is there such a large gap in between?

Because an employee spending one less hour a day making materials does not equal the company earning one more hour of money each day.

Only if that time is used to serve more customers, shorten delivery, reduce outsourcing, or truly reduce labor costs can it possibly enter the income statement. Otherwise, it may merely make employees a little more relaxed, or be filled up again by new low-value tasks.

Saving work hours only releases capacity; it is not profit itself.

Stage one: turn a seemingly good output into a work result that can be accepted;

Stage two: embed a single-point result into the complete process, rather than remaining forever in a pilot;

Stage three: turn the released capacity into new revenue or real cost savings.

In the same McKinsey survey, nearly three-quarters of high-performing enterprises claimed to have fundamentally restructured workflows because of AI, while among other enterprises this proportion was only about one-quarter. [6]

The gap lies not only in who bought stronger models, but also in who dares to change roles, permissions, assessments, and delivery methods.

This is also the "productivity J-curve" common to general-purpose technologies: processes, data, manpower, and organizational transformation cost money first, and efficiency gains are realized later. Early profit pressure does not necessarily mean the technology is useless, but always using "the future" to explain the present cannot replace financial verification either. [7]

Once a product is built, the cost of adding one more user can be very low, and what entrepreneurs value most is traffic, users, and network effects.

GPUs, servers, data centers, land, grid connections, and long-term power contracts have all become competitive thresholds. In the past, entrepreneurs competed for traffic; now they also have to compete for computing power, electricity, and delivery time.

According to the International Energy Agency, the capital expenditure of five major tech companies exceeded $400 billion in 2025 and is expected to grow another 75% in 2026. Global data center electricity consumption is expected to rise from 485 TWh in 2025 to about 950 TWh in 2030. [8]

Thus, entrepreneurs face their first difficult question: should computing power be bought or rented?

]

Buying gives you control over supply and iteration speed, but it also means bearing the risks of equipment upgrades, utilization, and depreciation; renting keeps you flexible, but may sacrifice gross margin and leaves you dependent on suppliers.

The second difficult problem is the time lag.

Microsoft offers a relatively smooth example: in the fourth quarter of fiscal year 2026, Azure revenue grew 43%, paid Microsoft 365 Copilot seats exceeded 30 million, and free cash flow for the quarter remained positive. This shows that AI commercialization and cash generation can proceed in parallel, but its capital expenditure has also risen significantly. [9]

Quarterly data from Meta, Alphabet, and Alibaba, however, show the other side: core businesses can still grow, but capital expenditure may run even faster.

Negative free cash flow does not necessarily mean investment failure; it may also be building future capacity.

But it will certainly change a company's margin of safety.

Companies that could previously expand using retained earnings have begun to rely more on bonds, leases, project companies, and private credit. Entrepreneurs who previously only needed to explain growth to shareholders now also need to explain repayment to creditors.

In its analysis related to the October 2026 Global Financial Stability Report, the IMF pointed out that hyperscale cloud providers' share of global investment-grade bond issuance has risen from less than 1% in 2024 to about 10%. The 14 publicly disclosed circular financing arrangements it examined involve about $300 billion in equity investment, $650 billion in debt, and about $1.4 trillion in future revenue commitments. [10]

So-called circular financing does not necessarily mean fake transactions.

But when investors are also suppliers, creditors, or customers, risks become intertwined. Once end-market commercialization falls short of expectations, equity value, loan repayment, and order revenue may come under pressure together.

The wealth an entrepreneur can truly dispose of personally also depends on shareholding ratio, dilution from subsequent financing, equity pledges, personal guarantees, restrictions on share reductions, tax costs, and asset liquidity.

Stock prices can price in ten years of expectations in a single day, but debt and fixed expenses will not disappear because of that.

But what entrepreneurs and investors should focus on most is no longer how many points a given model release has added, but three more specific changes.

First, the shift from "can do" to "can reliably finish."

What companies need is not a one-time successful demo, but continuous delivery of qualified results under specified permissions, cost, and time constraints.

Long-chain tasks amplify single-step errors, and exception recovery, human takeover, log auditing, and responsibility boundaries will all become part of the cost.

Therefore, closed and verifiable processes will be commercialized first.

Code can run tests, invoices can be checked against fields, industrial quality inspection can be compared against standards, and customer service refunds can be validated against rules. By contrast, open-ended strategic decisions and high-risk medical and financial judgments still require more human oversight.

Second, the shift from "how much does one token cost" to "how much does one qualified result cost."

A low token price is of course a good thing, but it is not the entire cost.

If a task requires multiple rounds of reasoning, human review, and repeated rework, even the cheapest call fees may not add up. The real cost should include model calls, computing depreciation, human review, system integration, security compliance, and error losses, divided by the number of accepted outcomes.

This means that the most commercially valuable capability in the future may not be always using the largest model.

Model routing, small models, specialized models, caching, and task decomposition may all improve profits more than simply stacking parameters.

Third, the shift from "did revenue rise" to "did capital earn its return."

Real demand does not mean shareholder returns have already been realized.

Next, the capital market will gradually change its set of exam questions.

In the early construction phase, the market rewards capital expenditure, computing power reserves, and revenue growth; after supply increases and depreciation rises, the market will focus more on utilization, renewal rates, gross margin after depreciation, ROIC, and free cash flow.

Chinese companies will not completely replicate the high-capital-expenditure path of the United States.

According to the China Academy of Information and Communications Technology, domestic average daily token calls have increased from 100 billion in early 2024 to 500 trillion in June 2026, and domestic model service prices are significantly lower than mainstream overseas models. [11]

Low prices help applications spread, but may also compress model service profits. Call volume proves "used more," but cannot by itself prove "earned more."

What truly gives China an opportunity to differentiate is the combination of open-source models, engineering capability, platform channels, and manufacturing scenarios: embedding AI into R&D, quality inspection, production scheduling, supply chains, and after-sales service, and measuring value by yield, inventory turnover, and delivery cycles, rather than merely comparing model quotes.

Is the cost per qualified unit of output falling, and does profit cover depreciation and review costs?

Does pressure on free cash flow correspond to verifiable high-return capacity, or to expansion with unclear utilization?

Are net debt, interest, leases, guarantees, and off-balance-sheet commitments controllable?

Do entrepreneurs' equity pledges, personal guarantees, and corporate financing stack the same risk twice?

Analysts at Zhengjing She believe that in the next three to five years, what is most likely is not simultaneous realization across the entire industry, but partial differentiation.

Scenarios where reliability and cost both cross the threshold will be the first to move from auxiliary positions to restructuring complete processes; in links where payment and process transformation lag behind capacity construction, even if revenue continues to grow, asset impairment and financing pressure may still emerge.

Technological progress can continue, and poor returns at some companies can occur at the same time.

In the end, what determines an entrepreneur's wealth is not how much computing power he bought, but how much qualified output, how much profit, and how much truly discretionary cash that computing power ultimately turned into. [Produced by Zhengjing She]

Main sources

[1] NVIDIA, "Second Quarter Fiscal 2027 Results," August 26, 2026.

[2] Meta, third-quarter 2026 results materials; the free cash flow measure includes capital expenditure and principal payments on finance leases.

[3] Alphabet, second-quarter 2026 results materials.

[4] Alibaba Group, "June Quarter 2026 Results."

[5] Stanford HAI, "The 2026 AI Index Report."

[6] McKinsey, "The State of AI in 2026: On the Road to ROI"; the survey covered 97 markets and 1,719 respondents, and the EBIT contribution is based on respondents' attribution.

[7] NBER, "The Productivity J-Curve: How Intangibles Complement General Purpose Technologies."

[8] IEA, "Key Questions on Energy and AI"; the 2030 data center electricity consumption figure is a forecast.

[9] Microsoft, fourth-quarter fiscal 2026 results and earnings call materials.

[10] IMF, October 2026 Global Financial Stability Report; the circular financing data reflect its examination of 14 publicly disclosed transactions and do not represent total industry financing.

[11] China Academy of Information and Communications Technology, related release of the "Token Economy Development Research Report (2026)"; token call volume does not equal revenue, and service prices of different models should not be mechanically compared without regard to capability definitions.

The data in this article are as of October 27, 2026. Forward-looking judgments in this article are research analysis and do not constitute securities trading advice. AI industry signal channel: Frontier Technology. This content comes from the internet and the original link; the views represent only the author and do not represent Huxiu's position. For copyright issues, please contact hezuo@huxiu.com, and we will verify and handle it promptly. Everyone who is changing the world and wants to change the world is in the Huxiu APP [ ]