Who Exactly Is Eating the AI Dividend: Looking at Profit Distribution Through Financial Report Data

**Key Points**

**1. Model and upstream revenue surge, yet still swallowed by massive investment**

Anthropic's compute and infrastructure costs account for 55% of revenue, with minimum purchase commitments to Google, Amazon, and Microsoft totaling $269.4 billion; OpenAI posted an operating loss of $20.92 billion in 2025. Upstream players such as Nvidia saw profits surge, but TSMC's capital expenditure budget reached $60 billion to $64 billion.

**2. Enterprise efficiency gains have not yet broadly translated into profits**

In a McKinsey survey, 80% said AI improved efficiency, but only 37% believed it contributed positively to EBIT, and about 6% of high-performing organizations benefited significantly; only 2% of S&P 500 companies quantitatively disclosed AI's profit contribution, with savings largely reinvested.

**3. Workers bear higher demands and risks to income and employment**

In a Gallup survey, 31% of AI users were asked to take on more responsibilities, rising to 37% among those who believe AI speeds up their work; 39% of respondents in China in an ILO survey expected their income to decline; Stanford research shows that workers aged 22-25 in occupations with high AI exposure face a relative employment gap of 19%. 2026-10-30 12:22 Who exactly is eating the AI dividend? Alphabet AI Quick Overview

Starting with Meta, Silicon Valley giants have ushered in a particularly bleak earnings week.

On October 29, Amazon released its third-quarter earnings report. Revenue at its AWS unit grew 39% year over year, the fastest growth rate in nearly five years, but the group's free cash flow over the past 12 months turned negative at $42.4 billion, and the stock fell more than 5% after hours at one point.

Google's experience was even more brutal: although Google Cloud revenue grew 103% year over year, Alphabet's stock still plunged 12.64% on October 29, falling more than 13% at one point during the session, with more than $500 billion in market value wiped out, a record for the company.

Earlier, Meta's third-quarter revenue grew 26% year over year, and its stock fell nearly 7% on the first trading day after the earnings release.

To some extent, this rebuts earlier remarks by Torsten Slok, chief economist at Apollo Global Management. Slok pointed out that over the past few years, revenue per employee and profit margins at the U.S. Magnificent Seven tech companies have continued to improve, but the other 493 companies in the S&P 500 have shown no similar improvement.

With a technological revolution, people always expect that the companies creating the technology can earn substantial profits, that the enterprises adopting the technology can improve operating efficiency, and that ordinary people can enjoy cheaper services, higher incomes, and more leisure.

Leading AI companies still need to invest enormous sums to maintain technological competition; upstream suppliers bear high expenditures and uncertainty over future demand. As for the enterprises actually using AI, efficiency has improved somewhat, but profits have not broadly improved; ordinary people face the risks of higher job demands, job displacement, and lower incomes.

Revenue at leading AI companies is growing rapidly. In the first half of 2026, Anthropic achieved revenue of $16.286 billion, up 1,300% year over year; OpenAI's annualized revenue had also approached $50 billion as of the end of September.

Frontier models have brought considerable revenue, but the pressure cannot be underestimated either.

On October 26, Anthropic officially made public its IPO prospectus, disclosing the company's financial condition. In the first half of 2026, Anthropic's compute and infrastructure costs alone reached $8.945 billion, accounting for 55% of revenue in the same period, up 283% year over year.

Over the next decade, the company still has a series of compute procurement and equipment leasing agreements to fulfill.

The prospectus shows that Anthropic's minimum purchase amount with Google reached $114.5 billion, with Amazon $110 billion, and with Microsoft $44.9 billion; in addition, the company also reached a compute procurement agreement with Musk's SpaceXAI, with a payment cap of $120.5 billion through 2029, and entered into an equipment finance lease arrangement with Broadcom, with total payments during the lease term of about $161.2 billion, excluding interest.

Among them, the minimum purchase commitments under the three agreements with Google, Amazon, and Microsoft total $269.4 billion, with terms extending through 2033 to 2036. Under the contracts, even if the corresponding compute is not actually used, Anthropic in principle still needs to pay the minimum commitment amount. SpaceXAI's $120.5 billion is an agreement cap, not an entirely non-cancelable payment obligation. After the initial lock-up period ends in early 2027, either party may terminate the contract with 90 days' notice.

As the agreements are gradually fulfilled, Anthropic needs to keep expanding revenue to cover the corresponding compute procurement costs.

OpenAI likewise faces enormous spending pressure. According to financial documents exposed in June this year, OpenAI's research and development expenses in 2025 reached $19.18 billion, with total annual costs and expenses of about $34 billion, ultimately recording an operating loss of $20.92 billion.

According to foreign media reports, OpenAI is advancing a new round of financing of about $30 billion, but the target valuation in the financing negotiations has dropped from $1.4 trillion to $1.25 trillion, a reduction of $150 billion. Some investors' willingness to subscribe fell short of expectations, and the issues of continued financing needs and returns on AI infrastructure investment are coming under market scrutiny.

Although compute procurement has been locked in years in advance, customers' future spending does not come with equally clear guarantees.

Anthropic disclosed in its prospectus that in the third quarter of 2026, more than 80% of revenue came from versions of the Claude model released within the previous six months. Claude Opus 5.5, released in September, reduced the cost for customers to complete comparable tasks by about 40%. Anthropic said that efficiency improvements enable the company to maintain profit margins while lowering customer costs, but sustained model research and optimization remain essential to preserving this competitiveness.

In addition, the prospectus disclosed that in the first nine months of 2026, Anthropic's top five customers contributed about 25% of revenue. However, many large customers have not signed long-term contracts, or their contracts are about to expire, and they can reduce or even stop spending at any time.

Compliance costs are likewise not to be overlooked. In 2025, Anthropic accrued $1.5 billion in settlement costs due to the Bartz copyright lawsuit, equivalent to about 33% of that year's revenue; OpenAI disclosed on October 23 that after an internal research model breached network isolation, the company suspended training and evaluation of its most advanced models, and the relevant training did not resume until 11 days later.

As model capabilities improve, safeguarding the security of the R&D environment has also become work into which companies must continuously invest resources.

From compute procurement to model research and development, from customer competition to security and compliance, leading AI companies' investments continue. They need to continuously develop more powerful models, win more customers, and rely on future revenue growth to cover the costs they have already borne today.

In fact, the gains created by technology do not necessarily remain in the hands of the technology providers.

In 2004, economist William Nordhaus studied technological innovation in the United States from 1948 to 2001 and estimated that the excess profits obtained by innovators amounted to only about 2.2% of the present value of the social benefits created by technological innovation. Most of the benefits flowed to consumers through means such as price declines.

II. Upstream players that have won orders face equally evident pressure

NVIDIA's revenue in the second quarter of fiscal 2027 was $96.2 billion, up 106% year over year, with net profit of $59.7 billion, up 126% year over year; TSMC's third-quarter revenue was $46.7 billion, up 41.2% year over year, with net profit of NT$763.1 billion, up 68.7% year over year, marking the sixth consecutive quarter of record profit; on October 28, SK Hynix announced its third-quarter financial results, with revenue of KRW 104.62 trillion, up 328% year over year, operating profit of KRW 83.33 trillion, up 632% year over year, and an operating margin of 80%.

But these profits have not fully dispelled doubts in the capital market.

English text is being prepared.

TSMC's capital expenditure budget for this year has already reached $60 billion to $64 billion. Management expects cumulative capital expenditure from 2026 to 2028 to significantly exceed that of the previous three years. Although the company still maintains strong cash generation capacity, building new fabs and expanding advanced-process and packaging capacity all require investment years in advance. Whether future customer demand can continue to support these expansions will determine the ultimate return on investment.

NVIDIA's risks have even extended to the customer side. Financial documents disclosed in August this year showed that the company's supply chain and capacity commitments increased from $119 billion in the previous quarter to $279 billion. At the same time, it has also committed to purchasing up to about $36 billion in cloud services over the coming years from some AI cloud service providers that buy NVIDIA GPUs, and to providing up to $105 billion in conditional credit guarantees for the data center project built by SB Energy and leased by OpenAI.

On October 29, Amazon announced its third-quarter financial results, with AWS revenue reaching $46 billion and operating profit reaching $17.8 billion. But the group's free cash flow over the past 12 months turned from positive $14.8 billion in the same period last year to negative $42.4 billion. The company maintained its 2026 cash capital expenditure plan of about $220 billion, most of which is for AI infrastructure.

Google is likewise facing cash flow pressure from massive investment. Financial results announced on October 28 showed that although Google Cloud has already achieved considerable profit, Alphabet's capital expenditure for the quarter reached $51.4 billion, with free cash flow of negative $9.07 billion. The company maintained its 2026 capital expenditure guidance of $195 billion to $205 billion and expects further increases in 2027.

Microsoft's quarterly free cash flow, although still $18.6 billion, also saw capital expenditure reach $45.2 billion, and is expected to exceed $50 billion next quarter.

As a representative of large-scale AI data center expansion, as of the end of August this year, Oracle's capital expenditure over the past 12 months totaled about $75.66 billion, with free cash flow of negative $28.72 billion.

Also on October 28, Oracle announced its financing and cash flow plan for the next four years: capital expenditure is expected to peak in fiscal 2027 to 2028, free cash flow is expected to turn positive in the second half of fiscal 2028, and free cash flow is expected to reach at least $30 billion in fiscal 2029, when it will begin to rely on its own cash flow to support business growth.

To complete this transition, Oracle has committed to raising no more than $20 billion in additional capital markets financing by fiscal 2030, planning to raise $10 billion each through equity financing and new debt, excluding the previously completed $20 billion equity financing and refinancing of existing debt.

For Neocloud companies such as CoreWeave and Nebius that specialize in renting out AI compute, the funding pressure is even more direct: taking CoreWeave as an example, in the second quarter of 2026, the company's revenue was $2.575 billion, up 112% year over year, while its net loss for the same period reached $626 million. Of that, net interest expense alone was $640 million, already equivalent to about a quarter of quarterly revenue.

To buy GPUs and build data centers, these companies need to continuously borrow or raise funds. Equipment procurement, depreciation and interest costs have already occurred, while future rental revenue will only be realized gradually over many years.

On one side are Nvidia, TSMC, SK Hynix and others, which have already realized considerable profits but still need to keep investing for future growth; on the other side are cloud computing and data center companies, which must bear huge construction costs, and some of them also face heavy financing and cash flow pressure.

The upstream players appear to have captured the benefits first, but the money already earned cannot guarantee returns on the next round of investment.

III. Ordinary people have become more efficient, but life is not necessarily better

In August this year, McKinsey released its latest global AI survey. 80% of respondents said AI had improved their work efficiency, but only 37% believed AI had made a positive contribution to their company's EBIT, basically flat from last year. Among them, high-performing organizations that can truly attribute at least 5% of EBIT to AI and consider this impact significant accounted for only about 6% of all respondents.

Another set of data comes from J.P. Morgan Asset Management. On October 16, the institution released an analysis pointing out that in the second-quarter 2026 earnings season, about two-thirds of S&P 500 companies mentioned AI on their earnings calls, but only 17% provided specific application results, and companies that truly quantitatively disclosed AI's contribution to profit accounted for only 2%.

J.P. Morgan gave two examples: American Express used AI to shorten its software development cycle by more than 30%, but chose to have engineers use the saved time to handle backlogged projects; Netflix used AI to improve content production efficiency and reduce costs, and planned to invest the saved funds into more content production.

Efficiency gains do exist, but companies choose to reinvest the benefits into their businesses, which may not necessarily be reflected as improved profit margins in the short term.

In May this year, the International Labour Organization (ILO) published a dedicated study calling this phenomenon AI's "Aggregation Paradox."

The study pointed out that AI can bring productivity improvements of 10% to 70% on some specific tasks, but these local gains have not yet formed equally obvious growth at the enterprise, industry and macroeconomic levels. Companies need to redesign workflows, adjust organizational structures, train employees, and also bear the deployment and operating costs of AI systems. Only when the technology is truly integrated into the entire production process can local efficiency gains possibly translate into broader economic benefits.

Compared with corporate profits, the situation of ordinary workers may be more complicated.

On October 6, Gallup released a survey of American workers. Among 7,845 respondents who had used AI at least a few times in the past year, 63% believed AI made them complete work faster, and 56% believed AI helped them find more creative solutions.

Similar changes have also appeared in Chinese companies. In September this year, the International Labour Organization (ILO), together with researchers from Renmin University of China, conducted in-depth interviews with 21 Chinese companies and surveyed 1,591 professionals. The study found that AI has already brought obvious efficiency improvements in some businesses: at an insurance company, 300 customer service staff increased the average number of issues handled per day from 6,000 to 15,000; at a smart manufacturing factory, production efficiency increased by 30%.

But after efficiency improves, workers may not necessarily receive corresponding returns. The Gallup survey showed that 31% of AI users said their employers required them to take on more work responsibilities because of AI adoption. Among employees who believed AI helped them speed up their work, this proportion rose further to 37%.

There is also uncertainty on income. Among the Chinese professionals surveyed by the ILO, 39% expected that AI might lead to a decline in their income.

In addition, in August this year, Stanford University researchers updated a study based on the salary records of millions of U.S. employees: in occupations with higher AI exposure, the employment level of young employees aged 22 to 25 showed a 19% relative gap compared with peers of the same age in occupations with lower AI exposure.

It is worth noting that workers are also consumers. Free chatbots, increasingly cheaper AI services, and tasks that used to require paid purchases but can now be completed on their own all mean that consumers have gained real convenience.

But the fact that consumers have gained convenience does not mean their overall economic situation will necessarily improve. The Organisation for Economic Co-operation and Development (OECD) has long tracked changes in labor productivity and wages, and found that many economies have already seen a disconnect between productivity growth and real wage growth. Even if the entire economic system can create more output with less labor, these gains may not necessarily translate proportionally into ordinary people's income.

Not long ago, in an interview hosted by Walter Isaacson, author of "Steve Jobs," Musk described the future he imagined: superintelligence may be thousands or even millions of times smarter than humans, and just as humans do not care about the daily lives of chimpanzees in the forest, superintelligence may not care about humans either.

Such a future cannot yet be verified. But today, when tech giants have begun to envision unlimited intelligence and unlimited abundance, ordinary people are still waiting for some much more realistic changes—higher income, less work, and genuinely improved lives.

We are still waiting.

This article comes from the WeChat public account: Letter AI, author: Yuan Xinyue, editor: Wang Jing AI Industry Signal Channel: Frontier Technology. This content is published with the author's authorization. The views represent only the author and do not represent the position of Huxiu. If you have any 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[]