Author: Lim Kim Thye

Translated by: Wu Says Blockchain

Core Summary

·As the unit capacity cost continues to decline, the growth in Token usage benefits large-scale cloud service providers, with infrastructure business profit margins reaching 33% to 38%.

·The proportion of capital expenditure to operating cash flow has risen from 41% in 2023 to about 105% in 2026. Related companies' total free cash flow has turned negative, and funding gaps are beginning to be filled by debt financing.

·The market's pricing logic has changed: capital expenditures that cannot be converted into real growth will be penalized, while investments that can drive AI revenue growth will receive positive feedback.

·The market concentration is decreasing. Binance investors are also reducing concentration in the semiconductor industry and reallocating funds to software and capital markets.

The trend of Token maximization ends in the second quarter, and the two curves begin to diverge

The second quarter marked the end of the assumption that increased Token consumption would directly translate into revenue at the model layer. Enterprises have realized that Token consumption is not linearly correlated with productivity, and thus no longer pursue Token usage maximization blindly. Meanwhile, AI Agent has rapidly gained popularity, and the market is more focused on the efficiency of Tokens required to complete individual tasks. The development focus of cutting-edge models has also shifted from inference benchmarking to Agent programming, memory architecture, and cost-effective lightweight inference.

As of early September 2026, the weekly Token usage routed through OpenRouter reached 137 trillion, about 20 times that of early in the year. Open weight models accounted for more than half of the platform's production environment inference Tokens, up from one-third in the last study, and all five models with the highest Token usage adopted open weights. The capabilities of these models are about 90% of those of closed-source models, while the cost per call is only about one-sixth of the latter. Stripe has maintained daily calls of 50 million while reducing its GPU cluster to one-third of its original size, leading to a 73% decrease in inference costs.

Consequently, the benefits brought by AI growth have begun to diversify. Enterprises choose models based on specific tasks, and an AI Agent typically requires thousands of Tokens to complete a task. Low-cost open weight models have thus become key to achieving economic viability for long-running Agents. As unit capacity costs decline, increased Token usage benefits large-scale cloud infrastructure providers with profit margins reaching 33% to 38%, but companies selling model capabilities directly find it more challenging to convert this into revenue.

Figure 1: Token usage approaches 150 trillion; Token price index peaked at $2.62 in July

Capital expenditure exceeds operating cash flow

The five major cloud computing operators in the U.S. have provided a capital expenditure guidance of $725 billion to $800 billion for 2026, depending on whether leasing and prepayments are included.

More notable than absolute figures is the ratio of capital expenditure to operating cash flow. The proportion of capital expenditure to operating cash flow for the five companies has risen from 41% in 2023 to about 105% in 2026. This indicates that the funds used for expanding infrastructure capacity have exceeded the cash generated by their core businesses.

This change is quickly reflected in free cash flow. Alphabet reported a quarterly free cash flow of negative $5.9 billion, its first negative result since going public. Meta's free cash flow dropped from $8.5 billion in the same period last year to $784 million. Over the past 12 months, Amazon's free cash flow was negative $7.6 billion, while Oracle consumed $23.7 billion in cash in FY 2026. The combined free cash flow of the five companies changed from a positive $246 billion in 2024 to about negative $37 billion in 2026, marking the first negative turn in this investment cycle.

As internal cash can no longer cover investment needs, external financing has begun to be a major source of funds. The proportion of debt financing in capital expenditures for large-scale cloud service providers has risen from 9% in FY 2024 to 32% over the past 12 months ending mid-2026, with the five companies' total debt amounting to about $700 billion.

The Bank for International Settlements (BIS) has made a noteworthy observation: the loan spread for AI private credit is about 6.2 percentage points, close to the financing spread of non-AI borrowers. In other words, the risks reflected in the equity market have not yet been fully priced in the credit market.

Figure 2: Capital expenditure exceeds operating cash flow in 2026; free cash flow turns negative

Demand signals are substantial, but the realization cycle is long

Three companies have shown the same trend.

·Google Cloud's revenue grew by 82% to $24.8 billion, with an operating profit margin increasing from 20.7% to 35.6%. The amount of unfulfilled contracts reached $514 billion. Currently, Google Cloud processes 22 billion Tokens per minute, up from 16 billion in the previous quarter.

·Microsoft's commercial remaining performance obligations (RPO) reached $678 billion, an increase of 84% year-on-year. Azure's revenue rose by 43%, and paid seats for Microsoft 365 Copilot exceeded 30 million.

·Oracle's remaining performance obligations reached $638 billion, a staggering 363% year-on-year increase. However, the company's disclosed data indicate that only about 12% will convert into revenue in the next 12 months, while approximately 20% will not be realized until five years later.

In terms of model companies, Anthropic's annualized revenue is about $47 billion, and OpenAI's is around $25 billion. Both companies are currently in a loss position and are preparing for an IPO. Goldman Sachs predicts that by 2030, Token consumption will grow to 24 times the current amount, but only 12% of knowledge workers will use Agent AI at that time.

The scale of these numbers is indeed substantial, but very few can be converted into actual revenue in the short term, and this time lag is key to the analysis in this report. Unfulfilled orders do exist, and capital expenditures have already occurred, but large-scale cloud service providers and cutting-edge model companies must convert contract demands into confirmed revenue quickly enough to justify these investments. Meanwhile, the underlying assets supporting these businesses typically need to complete depreciation within four to six years.

Figure 3: Unfulfilled orders far exceed current revenues; Google Cloud reaches $514 billion, Oracle reaches $638 billion

The market's pricing logic has reversed

Investors have started to incorporate these factors into pricing, and the latest earnings season clearly reflects this change.

The overall data is not sufficient to fully represent this shift; individual stock performance is more representative. Among the index constituents, companies exceeding earnings expectations averaged a 0.6% increase in share price, lower than the past five-year average of 1.0%; companies with disappointing performance averaged a 2.5% decline, also smaller than the past five-year average of 3.0%. The market's overall reaction to positive and negative news has weakened, but there has been severe differentiation within the AI sector.

Calculating from the first complete trading day after earnings reports, the differentiation is very clear. Alphabet's revenue grew by 24%, cloud business by 82%, yet its stock price fell by 7.13%, solely due to the company raising its capital expenditure guidance. Meta's stock fell by 7.95% due to disappointing earnings per share, once again raising capital expenditure and free cash flow dropping to $784 million.

The other side's performance is equally stark. Microsoft's Azure revenue grew by 43%, remaining performance obligations increased by 84%, while slightly lowering the capital expenditure for the 2026 calendar year, driving its stock price up by 15.51%, marking the largest single-day increase since 2008.

Amazon raised capital expenditure to $220 billion, but due to AWS's growth accelerating to 37%, its stock still rose by 15.32%, pushing its market value past $3 trillion. Nvidia rose by 8.74%. Palantir, which needs almost no capital to expand its infrastructure, surged by 29.45%.

The judgment criteria adopted by the market have become very clear: unless it can be proven that investments are translating into growth, capital expenditure is a burden. For this reason, even with an increase in capital expenditure, Alphabet faced market punishment, while Amazon received positive feedback.

Figure 4: Both companies raised capital expenditure; results were completely opposite, with Alphabet down 7.1%, and Amazon up 15.3%

The market's main line in 2026 is diffusion, not further concentration

Market concentration has peaked and begun to decline. The "Seven Tech Giants" weight in the S&P 500 index has fallen from about 35.3% in October 2025 to 33.2% in September 2026. As of August, the S&P 500 index has risen by 12.28% this year, the Nasdaq composite index has risen by 13.46%, and the Russell 2000 small-cap index has risen by 19.12%.

Among these, the performance of small-cap stocks is particularly noteworthy. This indicates that AI trading is spreading to industries such as industrial and power, rather than concentrating further on large-cap software companies. The equal-weight index has outperformed the market-cap weighted index by about 3.6 percentage points, indicating that after the easiest gains have been realized, funds are flowing into higher volatility second-order beneficiary targets.

However, current valuations are limiting the space for this rotation to continue developing. The expected price-to-earnings ratio of the S&P 500 index is about 19.6 times, slightly below the past five-year average of 19.9 times, but above the past ten-year average of 19.0 times. The interest rate environment has also provided no support, with the current effective federal funds rate at 3.63% and the 10-year U.S. treasury yield at 4.77%.

Therefore, the current conclusion has reversed from last year: the market uplift in 2026 is spreading to more industries and companies of different market capitalizations, and the passive investment portfolio's excess allocation to AI leading stocks is shrinking rather than continuing to increase.

Figure 5: The weight of the "Seven Tech Giants" has fallen from its peak, with Russell 2000 leading with a 19.12% rise this year

Binance investors are also diversifying their allocations, but the starting point is more concentrated

Comparing data from the end of June 2026 to September 4, it can be found that the adjustment direction of the S&P 500 index and Binance users’ stock holdings is largely consistent. The semiconductor sector's weight in the S&P 500 has fallen from 18.8% to 14.8%, and technology hardware industry weight has decreased from 6.8% to 6.2%; Binance investors' adjustment has been more significant, with their technology hardware holding percentage dropping from 15.92% to 7.94%.

During the same period, the weights of almost all other sectors in the S&P 500 have slightly increased, indicating that funds have not transitioned to another single theme but have diversified from previously highly concentrated areas to multiple industries. Binance investors continued to reduce holdings in aerospace and defense, interactive media and services, and industrial stocks, while increasing their allocations to capital markets, software, and general retail industries.

There are three main differences between Binance investors and the benchmark index: the semiconductor allocation ratio is as high as 42.08%, significantly higher than the S&P 500; the capital market industry's allocation reflects their crypto-related preference; and the concentration of holdings is also higher, with the top ten industries accounting for about 92% of their stock allocations, while the combined weight of the top ten industries in the S&P 500 is approximately 53%.

Figure 6: From June to September, the weights of semiconductors in the S&P 500 index and Binance users' holdings both declined

Monthly capital flows reflect real-time industry rotation

Monthly net capital flows further confirm the trend of market diffusion and clearly present the process of monthly changes in investor positions.

July: Investors were still positioning around the AI theme and remained optimistic before earnings season. Most investors bought in during the market pullback in late July, with the semiconductor sector absorbing most net inflows, and the capital markets sector also receiving significant funds.

August: Investors became more cautious. They began to take profits in capital market stocks while continuing to increase their holdings in semiconductors, but the pace of inflow slowed significantly. Meanwhile, a large amount of funds shifted to the software industry, with the overall net inflow in August dropping to about half of July's.

Since September: The semiconductor sector experienced its first monthly net outflow, as rising long-term interest rates put pressure on risk assets. However, it should be noted that current data only covers the first week of September, and the upcoming Federal Reserve rate decision may quickly reverse this trend.

Comprehensive observation of the data over these three months reveals that the portfolio remains highly concentrated in AI hardware at the beginning of the quarter, gradually spreading to software, capital markets, and other second-order beneficiary sectors. This is consistent with the direction of industry rotation observed in the benchmark index, but comes from different investor groups, and the adjustment speed is faster.

Figure 7: From July to September, funds flowed out of the semiconductor industry and shifted to the software industry

Trading volume aligns closely with investor holdings

Trading activity closely matches the structure of investors' asset allocation. The top ten industries by trading volume are essentially the same as those mainly held by Binance investors. This indicates that investors are more likely to be continuously building and adjusting existing positions rather than making short-term rotations between unrelated themes.

In September, the semiconductor industry led with a trading volume share of 33.84%, followed by the capital markets and software industries with shares of 16.71% and 13.67%, respectively. Together, they accounted for approximately 64% of the trading volume in the top ten industries. The technology hardware sector ranked next with 10.42%, consistent with the aforementioned trend of investors reducing their holdings in that industry.

The concentration of holdings and trading in the same industries is an important signal when interpreting fund flows. When trading volume focuses on the industries that investors are increasing their holdings in, the associated fund flows are more likely to represent clear allocation intentions rather than being purely driven by short-term turnovers.

Pre-IPO perpetual contracts can directly price counterparty risks

Binance's Pre-IPO perpetual contracts allow investors to establish positions on valuations of private companies that other major platforms have not yet provided. After Anthropic reported quarterly revenue growth exceeding double and achieved slight operating profits, its contract price accumulated to an increase of about 35% in August; after OpenAI launched its latest cutting-edge model, Astra, its contract price rose by about 23% in early September, reflecting the market's immediate judgment on the model release.

These two model companies contribute a large portion of AI revenue for large-scale cloud service providers. Wells Fargo estimates that more than 70% of Microsoft's AI revenue comes from the two; Barclays estimates that the two contribute about 73% of Amazon's AI revenue; UBS expects their share of Google Cloud's total revenue to rise from 28% in 2026 to over 48% in 2027. The $300 billion contract between OpenAI and Oracle also accounts for about half of Oracle's unfulfilled orders worth $638 billion.

Therefore, the actual customer concentration is higher than indicated by the overall capital expenditure data, and related revenues are highly concentrated in the two non-public trading counterparts. If either company's growth slows significantly, the impact will transmit from the AI revenue of large-scale cloud service providers to Oracle's unfulfilled orders, further impacting the securities products funded by suppliers that support the relevant infrastructure. Currently, the Pre-IPO perpetual contract is one of the few tools that can directly hedge such specific risk exposures.

Figure 8: After Anthropic released earnings and OpenAI launched Astra, Pre-IPO contract prices were re-priced

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。