Harbor View October 2026: Preserving Understanding in Times of Fear
In October, global AI-related assets underwent a notable correction, with technology and semiconductor stocks declining far more sharply than the broader market: the S&P 500 fell 0.93%, the Nasdaq 100 dropped 6.65%, and the Philadelphia Semiconductor Index plunged 15.36%.
Looking back over the past three years, skepticism about AI development is nothing new to us 鈥?only the focus of concern has shifted each time. In 2023, the market first worried about severe hallucinations in large language models, uncertain enterprise willingness to pay, and whether ChatGPT was truly a revolution reshaping the productivity foundation or merely a hard-to-monetize "tech toy." Under the dual pressure of high interest rates crushing valuations and commercialization doubts, the Nasdaq and the first batch of AI concept stocks experienced their first significant washout of over 10%鈥?5%.
By mid-2024, the market questioned why tech giants were spending hundreds of billions of dollars annually on GPUs while seeing no killer software revenue matching that scale. This concern resonated with rumors of delayed delivery of Blackwell architecture chips and the global unwinding storm in yen carry trades, causing the Philadelphia Semiconductor Index and the "Magnificent Seven" to plunge 20% to 30% within just three weeks.
In early 2025, China's open-source model DeepSeek demonstrated reasoning capabilities rivaling top closed-source models at extremely low training costs, instantly detonating a nuclear-level "compute deflation panic" on Wall Street. This fear directly caused computing power leader Nvidia to plummet nearly 17% in a single day, and combined with tariff friction concerns in Q1 2025, drove the Nasdaq into a deep阶段性 correction. In hindsight, AI faces vastly different concerns almost every year 鈥?from "can it be used" to "is there ROI," from "will algorithms eliminate compute" to today's "is cash flow sufficient, will compute become excessive."
The primary investment force behind global AI infrastructure is large cloud service providers, or CSPs. Based on current full-year 2026 estimates, the combined group capital expenditure of seven companies 鈥?Amazon, Google, Microsoft, Oracle, ByteDance, Tencent, and Alibaba 鈥?totals approximately $786 billion, approaching $800 billion. This does not yet include spending by Meta, some new cloud players, sovereign capital, and other traditional enterprises. It should be noted that group capital expenditure also includes non-AI businesses, so the entire investment of these seven companies cannot be directly treated as AI-exclusive spending.
Why are these companies investing so aggressively? Beyond competitive pressure to secure an "entry ticket" to the next-generation technology paradigm, the more fundamental business reason is that current AI investments genuinely deliver reasonable real returns financially. We can do a rough test: spread capital expenditure over a certain useful life, then compare with the combined incremental EBITDA from cloud and traditional businesses, to observe whether new earnings can cover the required return on investment. AI may directly contribute to cloud revenue or improve operating efficiency in advertising, software, and e-commerce, so observation at the group level is meaningful.
The chart below assumes a six-year useful life, zero residual value, and beginning-of-period commissioning, separately calculating the EBITDA needed to cover depreciation and achieve 10% and 20% annualized returns. Taking Q3 2026 as an example, the quarterly EBITDA thresholds corresponding to four rolling batches of investment are approximately $24.4 billion, $36 billion, and $49 billion, while group EBITDA increased about $42.7 billion year-over-year, falling between the 10% and 20% thresholds (bottom-right chart), close to a 20% return. This supports the judgment that new earnings can still support a considerable portion of investment, and one cannot conclude that all new investment has lost economic meaning just because free cash flow turned negative. At the same time, it should be noted that investment takes time to mature; immediately matching it to current-quarter profits underestimates the overall return rate.
Based on the company-disclosed figures aggregated in the chart below, the combined quarterly free cash flow of the four major CSPs began declining from early 2025, reaching approximately negative $30 billion by Q3 2026. This means current capital expenditure has begun to exceed what operating cash flow can cover during the same period, and funding pressure is emerging. Anticipating negative organic free cash flow, since 2026 major tech giants have broken their decade-plus tradition of being "net cash reservers," and begun raising funds on a large scale through equity issuance, corporate bonds, and off-balance-sheet financing leases to support capital expenditure. According to cash flow statements for the first nine months of 2026, Amazon obtained $115.1 billion in long-term debt financing cash inflows, including both bonds and loans; Alphabet raised approximately $85.2 billion net from debt issuance, and approximately $42.9 billion and $19.1 billion net from common stock and mandatory convertible preferred stock respectively. Oracle also completed approximately $20 billion in ATM equity financing during June-August this year. Meanwhile, the combined 10-year long-term procurement commitments of the four major CSPs plus Meta have reached $3.2 trillion as of this month, of which the total off-balance-sheet data center lease contracts signed but not yet commenced amount to $820 billion to $1.06 trillion.
The free cash flow on big tech balance sheets has indeed been spent, and high interest rates plus widening credit spreads have made various financing environment constraints more demanding. But the four major CSPs' own "self-generating capacity" is accelerating growth. In Q3 2026, the combined EBITDA of the four major CSPs was approximately $166.4 billion, up 34.4% year-over-year.
So why does the market remain so depressed? One important reason is that continued industry growth and growth meeting prior expectations are two different things. In our May and June observations, we cautioned that model progress, physical infrastructure construction, and enterprise organizational transformation each have different speeds. Capital expenditure returning from extremely high growth to more normal high growth is not necessarily bad in itself, yet it is enough to impact hardware earnings forecasts that linearly extrapolate high-speed growth. In the current 2027 expectation comparison table, buy-side estimates for six companies' capital expenditure total approximately $1.57 trillion, while sell-side estimates are about $1.45 trillion, representing growth of roughly 91% and 76% respectively relative to the 2026 base in the table.
Combined with Nvidia's guidance of up to 70% revenue growth for 2027, this means that for the four major CSPs alone, even assuming EBITDA continues to grow 40% next year, filling such aggressive buy-side expectation gaps would require an additional $280 billion to $400 billion in net financing through equity, debt, and leases. In the current financing environment, this appears unsustainable 鈥?credit spreads on the worst-rated Coreweave's credit bonds have even surged to 15%. If one expects 2027 growth to exceed the 126.1% year-over-year growth already seen in Q3 2026 (see Figure 6), expecting "trees to grow to the sky," that would violate financial and physical laws. Facing the 2027 growth slope, we must be more realistic.
If Capex does not shrink, does Anthropic's revenue growth signal insufficient AI demand and data center oversupply? Another major negative weighing on the market recently is that the quarter-over-quarter growth rate of annualized revenue run rate (ARR) at leading AI large model lab Anthropic has indeed slowed very rapidly. In the chart below, Anthropic's monthly ARR growth declined from a high of 58% at the start of the year to about 10% in August and about 2.1% in September. Data as of October 11 reflects signs of ARR re-acceleration, but the data is incomplete.
Why is the market so sensitive to Anthropic's growth slowdown? Because while AI Lab compute procurement accounts for a small share of CSPs' total existing revenue, it occupies an extremely large share of CSPs' incremental cloud revenue. Taking leader Anthropic as an example, through Q3 2026, its training and inference compute spending was approximately $9 billion, plus about $2.7 billion in cloud channel revenue-sharing fees, totaling approximately $12 billion directly contributed to CSPs 鈥?though this represents only 4% of the four major North American cloud providers' $300 billion in existing cloud revenue, it accounts for 12% of the four major cloud providers' $100 billion in incremental cloud revenue in the first three quarters of this year! If OpenAI and other leading model companies are all included, AI Labs' share of CSP incremental revenue is quite substantial; even in the revenue structure of emerging compute cloud representative SpaceX AI, the vast majority is "topped up" by AI Labs.
So is it true that CSPs' largest customers simply have insufficient demand? First, from a unit economics perspective, AI Labs' input-output efficiency has not deteriorated but is continuously and significantly improving. According to Anthropic's latest prospectus filing dated November 2, 2026 (see Figure 8), the revenue generated per unit of compute and infrastructure investment has soared from 0.40x in Q1 2025 to 1.72x in Q3 2026, achieving a 4.3x efficiency leap in just six quarters! This is gross margin and operating leverage genuinely "squeezed out" through model architecture optimization racing against unit compute costs, against a backdrop of continuously declining token unit prices. This positive operating leverage proves that AI Labs' underlying business model is fully viable, and their compute demand on CSPs has solid endogenous support. Meanwhile, although about half of Fortune 500 companies have now adopted Anthropic's products, related token spending accounts for only 0.1% of these giants' total expenses 鈥?just 0.1% penetration has already supported Anthropic's nearly $70 billion annualized revenue, and future enterprise-level penetration potential remains enormous.
Second, we must squarely face the potential risks implied by the ARR quarter-over-quarter growth decline. The most concentrated risk lies in the time gap between revenue growth and fixed commitments. According to Anthropic's filing materials, related commitments to be fulfilled by the end of 2030 total approximately $447.5 billion, of which approximately $69.8 billion falls in the twelve months after September 30, 2026. If Q3 revenue of approximately $17 billion no longer grows, simple annualization yields only about $68 billion, almost unable to cover minimum compute spending commitments, let alone margin declines. For Anthropic, the world's largest AI compute consumer and strongest demand source, once revenue stalls, compute does indeed face oversupply risk. This is precisely the core concern currently suppressing the US stock compute sector.
We must simultaneously see three industry truths clearly, otherwise we are "blind men touching an elephant." First, the competitive landscape has evolved from Anthropic's "sole dominance" in coding Agents at the start of the year to "a hundred flowers blooming" by year-end; one cannot judge total industry demand solely by Anthropic's ARR alone. For example, OpenAI's ARR in the chart below re-accelerated mid-year, with July-September growth notably higher than prior months and significantly above Anthropic. Meanwhile, Google's coding large model, Meta's Muse, xAI's Grok Bot coding and reasoning tools are all gaining momentum, and even open-source large models from China are catching up in coding and Agent performance. The current market landscape is one of fierce competition and diverse flourishing, with massive new compute consumption diverted to OpenAI, various cloud providers' self-developed models, and open-source ecosystems. Looking at just one company is biased.
Second, the previous "sole dominance" had caused token prices to rise rather than fall, while the recent "token price war" triggered by competition, although temporarily lowering nominal ARR growth, may be a necessary path toward application prosperity. Observing the LLM Token price index (see Figure 11), during the first half of 2026 when Anthropic monopolized the high-end coding market, the industry-wide average token price not only failed to decline but ran counter to the trend 鈥?this intelligence cost increase caused by localized monopoly is very unhealthy for the downstream application ecosystem. The recent slowdown in nominal revenue growth of leading model companies is largely due to unit prices plunging over 50%, rather than a shrinkage in underlying token call volumes. At near-zero intelligence costs, massive downstream applications and real demand can flourish, stimulating healthy industry development.
Third, large model lab revenue growth is never a smooth linear curve; it more often "climbs step by step" alongside leaps in application scenarios and product models. Looking back at history, whether Anthropic or OpenAI, the stages where revenue growth significantly amplified all originated from breakthrough new application scenarios. The AI coding application that drove Anthropic's explosive revenue growth this year had already shown early signs in 2025 or even earlier, building momentum for a long time before fully erupting in 2026. Similarly, on top of the current AI coding technology foundation, new Agent scenarios derived from AI office, AI gaming, AI marketing and advertising are experiencing astonishing leaps in model capabilities and user scale. According to the latest QuestMobile data (see Figure 12), as of September 2026, the monthly active user scale of typical domestic office Agent products has already erupted 鈥?for example, Tencent's WorkBuddy reached 32.353 million MAU (a quarterly quarter-over-quarter surge of 647.1%).
AI mainly helps people save time, especially on the consumer end 鈥?is "saving time" unlikely to form major demand? Looking back at human civilization history, one of humanity's most powerful underlying motivations is "laziness" (or the eternal pursuit of saving time and energy). From airplanes and high-speed rail for long-distance travel, to washing machines, robot vacuum cleaners, and electric kettles at home, to e-commerce, food delivery, and ride-hailing evolved from the mobile internet over the past decade-plus 鈥?they are all essentially helping people "save time." Admittedly, entertainment applications like gaming and short videos are studying how to consume your time; but the vast majority of human technological progress is essentially about studying how to help people save time or energy. The more critical economic difference is: industries studying "how to spend time" have a locked ceiling, because every human being only has 24 hours a day 鈥?that is the physical limit; while industries studying "how to save time" can infinitely compress the time length needed to satisfy people's various needs, theoretically having no upper limit, essentially "creating life (experiences)."
Therefore, although the essence of AI Agents is "saving time," this absolutely does not mean that "improving tedious work efficiency" is its entire application demand. We can cite several real scenarios that everyone has deeply experienced in daily life but long overlooked:
Scenario One: Full-cycle personal and family health management. For example, we want to check appointment availability for gastroenterology doctors at different hospitals in Shenzhen via WeChat, continuously monitoring newly released appointment slots; aggregate medical records, diagnoses, and examination reports scattered across different hospital mini-programs to form a chronological personal health file, then present it to doctors when needed. If an Agent can complete information retrieval, organization, and reminders after obtaining authorization, it has already solved many things people have long wanted to do but found too troublesome. Medical judgments should still be the responsibility of relevant professionals.
Scenario Two: Intelligent filtering and task triage of massive social information. For another example, there are hundreds of unread messages in WeChat. Which ones are making requests to me that need timely handling; which are irrelevant and can be ignored; which might interest me and are worth keeping? If an Agent can understand context, classify on our behalf, extract tasks, and organize reading order, what is saved is not just a few minutes, but the consumption of attention being constantly interrupted.
Scenario Three: Organizing and activating vast social relationship chains. For instance, many people have thousands of friends on WeChat, with many contacts poorly annotated at the time or never contacted since adding. An Agent can not only organize backgrounds and complete tags for these thousands of contacts, but also proactively provide plans in specific scenarios 鈥?for example, when you are going to Beijing next week for a business research trip, the Agent automatically filters out friends in Beijing suitable for meeting or regular gatherings and assists in scheduling.
Scenario Four: Precipitation of community intellectual collisions and multimodal creation. For example, in a high-quality industry group or book club group, everyone engaged in an inspiring heated discussion around a certain topic. An Agent can automatically organize and distill this fragmented group chat content, simultaneously brainstorm, automatically search and supplement external background materials, and ultimately directly convert it into a video or podcast with rich text and images, so we can watch, commemorate, or record it at any time.
Please carefully consider these four scenarios: none of them fall under traditional "clock-in work content," yet all are things we have long wanted to do but found too tedious and time-consuming, ultimately having to give up. This is precisely the core value leap of C-end AI Agents: Agents actually do all the things we "wanted to do but had no time for," and absolutely not just improving efficiency on "must-do drudgery." Philosophically, those things we genuinely want to experience, care about, and create but lack time for constitute our true "life." Therefore, once Agent generalization capabilities are adequate and token costs drop to a range acceptable to the masses, individual consumers paying for Agents is essentially "spending money to extend the breadth of their lives" and enjoy richer life experiences. From this dimension, C-end users' genuine willingness to pay will be very strong.
Of course, there is another side of the coin: whether Agents can do the job and whether users are willing to pay is one thing; whether independent Agent startups can actually collect the money is another 鈥?because that depends on the competitive landscape and ecosystem barriers. Recall the WeChat contact relationship chains, real-name medical information bound in hospital mini-programs, historical diagnosis records, and domestic and international instant messaging social networks or overseas personal email market shares we mentioned above 鈥?these ecosystems carrying user historical data and external connection permissions are the task "environments" where Agents can help you save time, and this "environment" itself is the deepest commercial barrier. If underlying large model and AI technology progress is not enough for users to switch their primary communication software or primary email for an independent Agent, then the original super-portal platforms still hold the largest moat and ultimate monetization pricing power.
From a total demand perspective, individual consumers can already hardly live without AI, and will only use it more and more deeply. Individuals spending money on tokens and Agent services is mainly to save time and enrich life experiences, unlike enterprises needing to precisely calculate financial ROI on every investment 鈥?this destinedly means the C-end will breed extremely massive terminal compute consumption.
We often ask: why do the vast majority of people miss one era after another? The answer lies in that even in a major bull market, the investment path is never smooth sailing. People used to say US stock investing is easy, as if simply choosing tech leaders and holding on would effortlessly harvest era dividends; only after truly experiencing sustained major declines does one realize the difficulty. US stocks have no price limit, and the intensity of volatility often far exceeds expectations. Era dividends are never gifts won by lying down 鈥?they always come wrapped in turbulence, fear, and repeated tests. Those who can truly traverse cycles need not only to see the long-term direction clearly, but also to have the composure to withstand the stormy waves along the way.
THE END
Disclaimer: Markets carry risks, and investment requires caution. Under no circumstances is the information in this article intended for readers' reference only; companies mentioned are solely for illustrating industry logic, and all content carries no investment advice. Readers should not rely solely on this article to replace their own independent judgment. Harbor Investment Management assumes no responsibility for investment losses, risks, and disputes arising from the use, citation, or reference of this article's content.
Comments