Who exactly has reaped the AI dividend?

Starting with Meta, Silicon Valley giants have ushered in an exceptionally dismal earnings week.

On October 29, Amazon released its third-quarter earnings report. Its AWS revenue increased 39% year-over-year, hitting the fastest growth rate in nearly five years. However, the group's free cash flow over the past 12 months turned negative at $42.4 billion, and its stock price once dropped by more than 5% in after-hours trading.

Google's situation is even more brutal: despite Google Cloud's revenue surging 103% year-over-year, Alphabet's stock price still plummeted 12.64% on October 29, once falling by more than 13% during the trading session, with its market value evaporating by over $500 billion and setting a new record in the company's history.

Even earlier, Meta's third-quarter revenue rose 26% year-over-year, and its stock price fell by nearly 7% on the first trading day after the earnings release.

To some extent, this is a refutation of the previous remarks made by Torsten Slok, chief economist of Apollo Global Management. Slok pointed out that in the past few years, the per capita revenue and profit margin of the seven major U.S. tech giants have continued to improve, but the remaining 493 companies in the S&P 500 have not seen similar improvements.

When a technological revolution breaks out, people always expect that companies that create technologies can reap substantial profits, enterprises that adopt technologies can improve operational efficiency, and ordinary people can also enjoy cheaper services, higher income and more leisure time.

However, when we look for the beneficiaries of this technological revolution along the AI industrial chain, we find that every party has its own difficulties.

Leading AI companies still need to invest huge sums of money to maintain technological competition; upstream suppliers are facing high expenditures and uncertainty about future demand. As for enterprises that actually use AI, their efficiency has improved, but no widespread improvement in profits has been seen; ordinary people are facing the risks of higher work requirements, job replacement and falling income.

AI has undoubtedly created huge economic value. But who on earth has taken all these dividends?

People building AI are paying for AI

The revenue of leading AI companies is growing rapidly. In the first half of 2026, Anthropic achieved a revenue of 16.286 billion U.S. dollars, a year-on-year increase of 1300%; OpenAI's annualized revenue as of the end of September has also approached 50 billion U.S. dollars.

Cutting-edge models have brought considerable revenue, but the pressure cannot be underestimated.

On October 26, Anthropic officially released its IPO prospectus, disclosing the company's financial status. In the first half of 2026, Anthropic's computing power and infrastructure expenses alone reached 8.945 billion U.S. dollars, accounting for 55% of the revenue in the same period, with a year-on-year increase of 283%.

In the next ten years, the company still needs to fulfill a series of computing power procurement and equipment leasing agreements.

The prospectus shows that the minimum procurement amount signed by Anthropic with Google reaches 114.5 billion U.S. dollars, 110 billion U.S. dollars with Amazon, and 44.9 billion U.S. dollars with Microsoft; in addition, the company has also reached a computing power procurement agreement with SpaceXAI under Elon Musk, with a payment cap of 120.5 billion U.S. dollars by 2029, and has reached a financial leasing arrangement for equipment with Broadcom, with a total payment of about 161.2 billion U.S. dollars during the lease term, excluding interest.

Among them, the total minimum procurement commitments of the three agreements with Google, Amazon and Microsoft amount to 269.4 billion U.S. dollars, with the term extending from 2033 to 2036. In accordance with the contract, even if the corresponding computing power is not actually used, Anthropic in principle still needs to pay the minimum committed amount. The 120.5 billion U.S. dollars for SpaceXAI is the upper limit of the agreement, not all non-cancellable payment obligations. After the initial lock-up period at the beginning of 2027 ends, the two parties can terminate the contract by giving 90 days' notice in advance.

As the agreements are fulfilled one after another, Anthropic needs to continuously expand its revenue to cover the corresponding computing power procurement costs.

OpenAI is also facing huge expenditure pressure. According to financial documents exposed in June this year, OpenAI's R&D expenses in 2025 reached 19.18 billion U.S. dollars, the total annual costs and expenses were about 34 billion U.S. dollars, and it finally recorded an operating loss of 20.92 billion U.S. dollars.

According to foreign media reports, OpenAI is pushing forward a new round of financing of about 30 billion U.S. dollars, but the target valuation in financing negotiations has dropped from 1.4 trillion U.S. dollars to 1.25 trillion U.S. dollars, a decrease of 150 billion U.S. dollars. Some investors' willingness to subscribe is lower than expected, and the continuous financing demand and the return on investment of AI infrastructure are being scrutinized by the market.

Although computing power procurement has been locked in for many years in advance, there is no equally clear guarantee for customers' future expenditures.

Anthropic disclosed in its prospectus that in the third quarter of 2026, more than 80% of its revenue came from versions of the Claude model released within the past six months. Claude Opus 5.5, released in September, reduced the cost for customers to complete similar tasks by about 40%. Anthropic stated that efficiency improvements allow the company to reduce customer costs while maintaining profit margins, but continuous model R&D and optimization are still essential to maintain this competitiveness.

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

Compliance costs also cannot be ignored. In 2025, Anthropic accrued 1.5 billion U.S. dollars in settlement expenses due to the Bartz copyright lawsuit, equivalent to about 33% of the revenue of that year; OpenAI disclosed on October 23 that after internal research models broke through network isolation, the company once suspended the training and evaluation of its most advanced models, and the related training did not resume until 11 days later.

As model capabilities improve, ensuring the security of the R&D environment has also become a task that the company must continuously invest resources in.

From computing power procurement to model R&D, from customer competition to security compliance, leading AI companies are continuing to invest. They need to continuously develop more powerful models, win more customers, and rely on future revenue growth to cover the costs they have already incurred today.

In fact, the benefits created by technology do not necessarily stay in the hands of technology providers.

In 2004, economist William Nordhaus studied technological innovations in the United States from 1948 to 2001, and estimated that the excess profits obtained by innovators only accounted for about 2.2% of the present value of social benefits created by technological innovations. Most of the benefits flow to consumers through ways such as price reduction.

This study does not directly reflect the income distribution of today's AI industry, but it illustrates that the economic value created by a technology may far exceed the profits that technology providers can finally obtain.

For AI companies that continue to invest funds to develop cutting-edge models, the improvement of technical capabilities does not necessarily bring a proportionate increase in profits.

Whether future revenue growth can support these continuously expanding investments remains to be verified. But batches of computing power procurement orders have been continuously flowing to the upstream of the AI industrial chain.

Upstream players that got orders are also facing obvious pressure

So, have these upstream enterprises that received orders really reaped the AI dividends?

Nvidia's revenue in the second quarter of fiscal 2027 was 96.2 billion U.S. dollars, up 106% year-on-year, and its net profit was 59.7 billion U.S. dollars, up 126% year-on-year; TSMC's third-quarter revenue was 46.7 billion U.S. dollars, up 41.2% year-on-year, and its net profit was 763.1 billion New Taiwan dollars, up 68.7% year-on-year, refreshing the profit record for the sixth consecutive quarter; on October 28, SK Hynix released its third-quarter earnings report, with revenue of 104.62 trillion won, up 328% year-on-year, operating profit of 83.33 trillion won, up 632% year-on-year, and an operating profit margin of 80%.

The AI upstream sector has received real money, and even earned profits that were unimaginable in the past.

However, these profits have not made the capital market completely let go of its doubts.

Since October, many leading semiconductor stocks have experienced obvious corrections. TSMC's U.S. stock price fell from $485.80 on October 5 to $412.35 on the 22nd, a drop of about 15%; on October 22 alone, Nvidia fell 4.4%, Micron fell 6.26%, and ASML fell 6.93%. SK Hynix once fell by more than 30% from its phased high in late September.

To maintain growth, upstream AI enterprises need to continuously invest funds.

TSMC's capital expenditure budget this year has reached 60 billion to 64 billion U.S. dollars. The management expects that the cumulative capital expenditure from 2026 to 2028 will significantly exceed that of the previous three years. Although the company still maintains strong cash generation capabilities, the construction of new fabs and the expansion of advanced process and packaging capacities require investments years in advance. Whether future customer demand can continuously support these expansions determines the final return on investment.

The risks undertaken by Nvidia have even extended to the customer side. Financial documents disclosed in August this year show that the company's supply chain and capacity commitments increased from 119 billion U.S. dollars in the previous quarter to 279 billion U.S. dollars. At the same time, it also promised to purchase up to about 36 billion U.S. dollars of cloud services from some AI cloud service providers that buy Nvidia GPUs in the next few years, and provide a conditional credit guarantee of up to 105 billion U.S. dollars for the data center project built by SB Energy and leased by OpenAI.

Although these arrangements do not mean that Nvidia has paid the same amount of cash, it means that it is supporting downstream customers to continue building AI computing power through future procurement commitments and its own credit.

Apart from chip suppliers, large cloud vendors such as Microsoft, Google, and Amazon, although supported by mature businesses, their continuously expanding AI capital expenditures have also begun to put pressure on cash flow and return on investment.

On October 29, Amazon released its third-quarter earnings report, with AWS revenue reaching 46 billion U.S. dollars and operating profit reaching 17.8 billion U.S. dollars. However, the group's free cash flow over the past 12 months turned negative at 42.4 billion U.S. dollars, compared with positive 14.8 billion U.S. dollars in the same period last year. The company maintains its cash capital expenditure plan of about 220 billion U.S. dollars for 2026, most of which is used for AI infrastructure.

Google is also facing cash flow pressure brought by huge investments. The earnings report released on October 28 shows that although Google Cloud has achieved considerable profits, Alphabet's group capital expenditure in the quarter reached 51.4 billion U.S. dollars, and its free cash flow was negative 9.07 billion U.S. dollars. The company maintains its capital expenditure guidance of 195 billion to 205 billion U.S. dollars for 2026, and expects it to continue to increase in 2027.

Although Microsoft's quarterly free cash flow is still 18.6 billion U.S. dollars, its capital expenditure has also reached 45.2 billion U.S. dollars, and it is expected to exceed 50 billion U.S. dollars in the next quarter.

As a representative of enterprises that are aggressively expanding AI data centers, as of the end of August this year, Oracle's cumulative capital expenditure over the past 12 months was about 75.66 billion U.S. dollars, and its free cash flow was negative 28.72 billion U.S. dollars.

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

To complete this transition, Oracle promises to raise no more than 20 billion U.S. dollars through the capital market additionally by fiscal 2030, planning to raise 10 billion U.S. dollars each through equity financing and new debt, excluding the 20 billion U.S. dollars equity financing completed previously and existing debt refinancing.

For Neocloud enterprises that specialize in renting out AI computing power such as CoreWeave and Nebius, the capital pressure is more direct: taking CoreWeave as an example, in the second quarter of 2026, the company's revenue was 2.575 billion U.S. dollars, up 112% year-on-year, while its net loss in the same period reached 626 million U.S. dollars. Among them, net interest expenses alone reached 640 million U.S. dollars, which is equivalent to about a quarter of the quarterly revenue