Sinolink 2026 US stock outlook: The internal melting point and external turning point of the AI bubble.

In 2025, the US stock market experienced a historic year with the impact of tariffs, fiscal shift, and the intertwining of industrial waves. The "Deepseek Moment" and the April "Independence Day Tariffs" each triggered market earthquakes, but the resilience of the US stock market continued to show after the impact. Since the third quarter, the OBBBA Act and the dovish turn of the Federal Reserve have brought positives in the fiscal and monetary aspects, while OpenAI announced a series of major investment agreements with companies like NVIDIA and Oracle, and the artificial intelligence frenzy pushed market sentiment to new highs. However, recently, the AI narrative has once again faced scrutiny (see "The Third AI Narrative Challenge in the US Stock Market"), as tech giants are "at all costs" to trigger a capital expenditure frenzy - cash flows are shrinking continuously while external financing is increasing. At the same time, the complex relationships of mutual investments, related transactions, circular financing, and intertwined feedback mechanisms have intensified the market's irrational prosperity. Is the AI investment bubble really there? How can its extent be quantified? Will there be a point in 2026 where the market's vulnerability is significantly exacerbated? I. The Bubble's Existence is Rational Many opinions believe that there is no bubble in the current AI investment field, citing the high revenue, healthy cash flow, and acceptable leverage of tech giants compared to the dot-com bubble in 2000 with many companies having no revenue. However, this comparison overlooks the fundamental differences in the objects and players involved. Today, the scale and concentration in AI investments far exceed that of 2000, with the investments by AI giants in the economy and the positive externalities they bring being incomparable to 2000. This means that if these AI giants encounter problems, the impact on the entire financial and technological ecosystem will be disastrous, a measure that cannot be assessed by simple revenue or valuation metrics. The formation mechanism of every bubble is similar, but the manifestations and carriers of systemic risk are different. From an industrial perspective, the enhancement of AI value on overall social productivity will be a lengthy process. While AI has benefits in coding and certain aspects within tech companies, its contribution to productivity improvement for most industries in the short term will be limited, primarily because organizational and process changes lag behind the technology itself. Just like at the beginning of the electricity revolution in the late 19th century, electric motors simply replaced steam engines on factory spindles but did not lead to efficiency improvements, if AI is not deeply restructured with respect to organizations, incentive mechanisms, and decision-making processes within companies, its value will be wasted. AI cannot complete decision loops today, most of it still is "assisting human predictions" rather than "replacing human decisions", causing senior decision-making to remain a bottleneck. Additionally, AI's dividend from outsourcing low-skilled jobs is limited, and to demonstrate AI's value, it must replace high-skilled jobs, requiring a long time to bridge the "AI gap". But compared to the distant industrial prospects, investing in AI has become one of the market's consensuses. The speed of the bubble's expansion and the landfall of industrial productivity are two different things. Just because the industry needs time doesn't negate the rationality of investing in AI. Currently, various parties have the motivation to inflate the bubble: tech companies "all in AI" to avoid obsolescence, financial institutions take advantage of loose liquidity to increase profits, media accelerates the bubble by increasing traffic, and American residents (middle class and affluent) bind AI stock bubbles to retirement funds, among other ways. Even if the bubble bursts, it is not necessarily a bad thing, because a new organizational revolution usually needs a low-cost environment to germinate. After the burst of the internet bubble in 2000, new things truly began to grow. The excessive infrastructure brought by the bubble (such as fiber optics, servers) became cheap after the collapse, providing fertile ground and extremely low operating costs for the rise of internet giants. After the AI bubble bursts, the cheap computing power, electricity, and infrastructure left behind will also become fertile ground for future new business models and small company innovation. When costs significantly decrease and industry standards are unified, the organizational revolution reconstructing the operation logic around AI can truly unleash productivity. For the US, carrying the AI bubble to the end is not just an economic act, but also crucial for the national destiny. On one hand, a large portion of American household wealth (especially those of the top 1% and top 10%) is concentrated in the US stock market, and the prosperity of the stock market is also the foundation of the US dollar's credibility. To support the dollar and the massive fiscal deficit, the US has no reason to actively burst the bubble and must allow the stock market to prosper continuously. On the other hand, the AI narrative is also a hidden form of debt. By financializing AI, pushing up stock prices (such as the Mag7) and asset values (such as credit backed by overvalued collateral assets), the US is attracting international investors and allied governments to foot the bill. The US is "selling cards, stocks, and dreams," transferring debt and risk to allies and international capital, making them bear the cost of building the computing infrastructure. In geopolitical games, the US, through monopolizing the ecosystem and consolidating itself, launches a "land grab movement" which shifts the cost to allies through debt transfer mechanisms. Once the bubble bursts, allies as the main payers and infrastructure builders, the US will gain a large amount of electricity, computing power, and new infrastructure at a lesser cost, laying the foundation for future innovation. Having foresaw the outcome of "friends die rich, not poor", the US is bound to carry the AI bubble to the end. For the US stock market, the impact of AI is evident, with the core support for valuations being the market's belief that AI technology can shape a bright future comparable to the Industrial Revolution and the Information Revolution. However, after excluding the Mag7, the "S&P 493 Index" has shown zero growth for the past two years, and high interest rates continue to suppress traditional sectors. Whether the value of AI will ultimately benefit the entire society remains unknown and can only be judged in hindsight. II. The Vulnerability of the "Iron Chain": Where is the Weakest Link? The artificial intelligence industry can be divided into three levels - chip manufacturers, cloud service providers, and model developers. Chip manufacturers provide AI hardware, benefiting first in terms of revenue and having the most robust cash flows. Cloud service providers offer computing facilities and services for model development, with costs mainly centered on hardware procurement and energy consumption, and revenues coming from cloud computing leases. Model developers focus on AI model development and training, with overhead mainly associated with computing resource procurement and revenue coming from API service subscriptions. In the past year, there have been numerous high-profile, cross-level investment transactions between enterprises within the three aforementioned levels. The marketization integration of the industry chain, from chip manufacturing to cloud computing to AI application, although beneficial in integrating industry chain resources, improving chip supply, computing support, and application scenarios, has temporarily boosted performance and valuations, improving financing capabilities, but has also blurred traditional industry boundaries, potentially creating demand illusions and nurturing the fragility of the industry chain. If AI giants cannot generate enough profit expectations in their core businesses, while the liquidity environment deteriorates marginally, the entire chain may face significant risks due to a loss of faith. Currently, there is severe lack of disclosure in the information relating to cyclical investments in sectors such as related transactions, customer concentration, etc. Titans in the complex network of capital and business relationships (such as cross-shareholdings, co-investments, strategic cooperation, etc.) should be considered as acting collectively. In cyclical investments, insufficient disclosure of related relationships may make it difficult for investors to see the true risks, with some revenues being double-counted and potentially exaggerating the monetization scale of the AI ecosystem, temporarily masked by loose liquidity. Moreover, the titans should more clearly disclose the extent of their dependence on key large customers; for example, Oracle should explicitly state in its financial reports that the majority of the increase in its Remaining Performance Obligations (RPO) mainly comes from a single contract with OpenAI. Under the ambiguous cyclical investment model, negative performance-related public opinion tends to amplify market sensitivity. For instance, an internal document from Oracle on October 7 showed that the gross profit margin of its NVIDIA-related cloud business was only 14% (overall gross margin of around 70%), sparking concerns in the market about its serious reliance on a few large clients in the AI cloud business with weak bargaining power. As a result, Oracle's stock price fell sharply by 7.1% intraday on that day. In addition, Microsoft's third quarter report showed a $31 billion loss on its investment in OpenAI, a 490% increase over the same period last year. Calculating based on Microsoft's 32.5% stake in OpenAI, this means that OpenAI's losses for one quarter exceeded $12 billion. Furthermore, looking at the AI industry chain, there is a significant difference in the profitability of upstream and downstream players. Represented by NVIDIA, upstream chip manufacturers are the first to enjoy high profits, benefiting from the surge in demand for AI chips, their pricing power and order visibility being strong. Midstream cloud service providers also have clear business models. Amazon, Google, Microsoft, and others have built resilient business models by deeply integrating AI into their core businesses, forming strong moats, with the share of revenue from cloud services by the three giants gradually increasing in the past two years. Oracle has captured the huge demand for computing power needed for AI training and reasoning, locking in significant revenue for the next few years with its cloud infrastructure and high-value contracts with top AI companies such as OpenAI, Meta, and xAI. However, there is intense competition among downstream model developers. The profitability has shown significant differentiation. Large-scale model providers like OpenAI, facing the huge costs of R&D and computing power, while enterprise application vendors like Salesforce and Adobe can overlay AI on their mature SaaS products, with lower marginal costs. From the contribution of AI giants to their respective stock prices this year, it can be seen that the profitability contribution of chip manufacturers is the highest, followed by cloud service providers, with model developers being the weakest. Meta, on the other hand, falls under the most special category. Unlike Microsoft, Google, Amazon that have cloud services businesses and use AI capabilities as tools and services to generate profits, Meta has the largest economic exposure, with 99% revenue dependent on digital advertising (related to the US economy). While it has invested heavily in building a robust AI social engine, its commercial returns depend more on the demand for digital advertising in the real economy and the future prosperity of the commercial ecosystem. The US is currently experiencing a typical "stagflation" environment, with a widening wealth gap and consumer polarization. The affluent who own more assets and stocks are getting richer in the AI bull market, while the middle and lower-class individuals burdened with student loans, car loans, mortgages are facing increased life pressures. Looking at the US stock market's third-quarter reports, high-end consumption (such as luxury-related consumption and airline first-class sales) remains strong, while low-end consumption continues to degrade, with more people turning to value meals at McDonald's, Walmart, or even cheaper supermarkets. When the US economy's weakness reaches a level where Meta's advertising clients can cut budgets, it may be a more vulnerable moment for the AI chain. III. The Fragility of Trillion-Dollar Capital Expenditure Starting in 2025, US tech companies' capital spending is showing competitive growth, with sustainability beginning to be questioned. In the third quarter of 2025, the capital spending scale of the top five AI-intensive companies (the "AI Big Five": Microsoft, Meta, Amazon, Google, Oracle) reached $105.73 billion, an increase of 72.9% year-over-year. The enormous capital expenditures have brought challenges in cash flows; by the end of the third quarter of 2025, the average Capex (capital expenditure) to CFO (operating cash flow) ratio for the AI Big Five was 75.2%, a rise of 29.7 percentage points from a year ago; the average Capex to revenue ratio was 28.1%, an increase of 12.3 percentage points from a year ago. Looking at the free cash flow (CFO - Capex - net debt repay) perspective, by the end of 2025 Q3, among the five AI giants with heavy investments, Oracle's free cash flow was already in the negative, unable to support the enormous capital expenditures in the same period, only sustaining itself by consuming existing cash reserves and increasing external financing. In terms of the ratio of average cash reserves at the end of the period to necessary expenses (Capex + net debt repayment + dividend payments + buyback expenses), by the end of 2025 Q3, the average value for the five companies was 94.4%, down by 39 percentage points from a year ago, with Meta being only 37.3%, indicating that the safety cushion of existing cash reserves might become thinner in the future. Building on this, we make the following calculations: Assumption 1 Extrapolating the capital spending, operating cash flow, and revenue of the AI Big Five at the average growth rate over the past year: by the second quarter of 2027, the average Capex/CFO is expected to reach 95.9%, nearing the peak of Intel post the burst of the bubble for the "four giants of the Internet (Microsoft, Intel, Cisco, IBM)"; by the third quarter of 2026, the average Capex/revenue for the AI Big Five is expected to reach 39.5%, surpassing the peak of Intel post the burst of the bubble. Assumption 2 Calculating the capital spending of the AI Big Five based on the median compounding growth rate (CAGR) expected from 2025 to 2028, with revenue and net profit based on Bloomberg consensus expectations, and maintaining the proportional trends of operating cash flow to net profit unchanged: by the third quarter of 2026, the AI Big Five's average Capex/CFO is expected to reach 96.9%, equivalent to the peak of Intel post the bubble burst; by the fourth quarter of 2026, the AI Big Five's average Capex/revenue is expected to reach 38.7%, nearing the peak of Intel post the bubble burst. Overall, the fragility of capital expenditures may gradually intensify in the second half of next year. However, considering that tech companies will continue to go "all in AI" to avoid obsolescence, capital expenditures have rigidity, and if companies save on other expenses (such as dividends, buybacks, and equity incentives), it may serve as a turning point for the narrative. Viewed from the perspective of free cash flow, turning negative free cash flow could be a moment when the fragility deepens. Assuming the proportional trends of operating cash flow to net profit remain unchanged and net profit is based on Bloomberg consensus expectations, with Capex based on market expectations for the median compounding growth rate, net borrowing based on the average of the past five years, by the end of the fourth quarter of 2026, Meta may face a crisis in its free cash flow. When Meta, with a already fragile fundamental position, falls into a deeper crisis, doubts on the narrative may be pushed to new heights. Moreover, as tech giants have significantly increased capital expenditures in the past year for data center construction, but depreciation will not be accounted for until officially put into use, the impact on the profit and loss statement is yet to be seen. If it is assumed that depreciation will be gradually brought in for fixed assets beginning from the fourth quarter of 2024, and a linear depreciation is accounted for over six years, as of the third quarter of 2025, the potential Capex depreciation/net profit ratio for the AI Big Five has already reached 11.8%, and will exponentially rise in the future. Based on calculation under Assumption 1, the implied depreciation of Capex will grow from $14.9 billion in the third quarter of 2025 to $114.5 billion by the end of 2028, approximately a 7.7-fold increase. By the end of 2026, 2027, and 2028, the implied Capex depreciation/expected net profit ratios will reach 37.6%, 60.2%, and 82.0%, respectively. Based on calculation under Assumption 2, the implied depreciation of Capex will grow from $14.9 billion in the third quarter of 2025 to $123.9 billion by the end of 2028, approximately an 8.3-fold increase. By the end of 2026, 2027, and 2028, the implied Capex depreciation/expected net profit ratios will reach 37.0%, 60.5%, and 87.7%, respectively. IV. High Leverage and Off-Balance Sheet Financing Risks In the first 11 months of this year, the total issuance of US hyperscaler corporate bonds reached $103.8 billion, over five times the annual issuance volume in 2024 ($20.1 billion), with the scaled-weighted interest rate rising from 4.75% to 4.91%. The surge in supply has raised bond spreads, with the 5-year CDS of companies like Oracle and Coreweave increasing by 49 basis points and 304 basis points, respectively, while the OAS spreads for investment-grade (IG) and speculative-grade (SG) corporate tech bonds have also trended upwards. It is widely believed that relying solely on public debt issuances may be difficult to fill the gap in funding faced by tech giants. The forecast indicates that global data center construction between 2025-2028 will require $2.9 trillion in Capex, with $1.5 trillion coming from external financing ($200 billion from corporate bonds, $150 billion from ABS and CMBS products, $350 billion from PEVC and sovereign capital, and $800-$1.2 trillion relying on the private credit market). The opacity of credit ratings and holders of private credit products poses a significant risk. Taking Meta as an example, with a design for a complex off-balance sheet financing scheme for the $27 billion Hyperion data center project - creating a joint venture called Beignet Investor controlled 80% by investment management firm Blue Owl Capital, which issued $27.3 billion in bonds. Meta held only a 20% stake, without consolidation in financial statements, so the huge debt doesn't directly appear on its balance sheet, but Meta provided substantial implicit guarantees to the joint venture, resulting in potential liabilities. Meta is not an exception; companies like xAI, Anthropic have also adopted similar SPV financing models. This reflects the common dilemma that tech giants face in the AI arms race, where they need astronomical amounts of funds while maintaining attractive financial statements and credit ratings. However, there are significant potential financial risks associated with these off-balance sheet financing operations, as if operated at the trillion-dollar level, systemic risks become hard to ignore. If AI chip and data center technology iterations occur at a pace beyond expectations, it implies that assets held by SPVs may significantly depreciate before generating sufficient returns, ultimately transferring risks to bond investors. Historically, off-balance sheet financing tools have been linked to major crises such as the bankruptcy of Enron in 2001 and the subprime mortgage crisis in 2007. With the massive capital demands of AI investments today, if a large number of companies rely on such concealed leverage, a single default event could trigger systemic risks through a highly intertwined