The AI industry is charging forward like an accelerating train, but by July, investors inside the carriage were feeling increasingly nauseous. The AI sector remained intensely hot in July, with companies like Taiwan Semiconductor Manufacturing, Google, and Intel delivering robust earnings, frontier models continuing to improve, and model companies' revenue and user base expanding steadily.
However, capital markets have become significantly more sensitive to negative news, reacting with greater severity. A single report about Meta potentially selling computing power, a breakthrough in an open-source model, or even a gross margin figure slightly below buyer expectations could trigger violent swings across the entire AI sector. Looking back at the whole of July, the Nasdaq index fell 3.20%, while the once red-hot Philadelphia Semiconductor Index dropped 20.61%, and the memory index declined 31.79%. As the industry's temperature continues to rise, yet stock prices begin to overreact to every whisper of wind, we need to reassess: is this the ebbing tide, or merely turbulence within the current?
On the industry front, the AI wave accelerated in July. First, the entire AI hardware supply chain is benefiting from AI development, with the AI dividend no longer being solely enjoyed by Nvidia. Taiwan Semiconductor Manufacturing comprehensively raised its full-year revenue guidance, increased its capital expenditure budget, and actively expanded production capacity. The CEO of Taiwan Semiconductor Manufacturing, typically conservative in his outlook, repeatedly emphasized his steadfast confidence in the long-term AI trend during the earnings call. When analysts questioned future AI growth expectations, he confidently responded: "Stronger, stronger, stronger." The shortage of advanced process capacity is the hardest "physical constraint" for this round of AI infrastructure. With nearly perfect earnings and aggressive expansion plans, Taiwan Semiconductor Manufacturing directly addressed market concerns about a slowdown in capital expenditure. Intel also delivered results that comprehensively beat expectations, with revenue growing 25% year-over-year, its fastest pace in 15 years. Even Intel, amidst its transformation, has hitched a ride on the revaluation wave for Agent computing power allocation. The CEO clearly stated that AI is simultaneously driving demand for CPUs, ASICs, advanced packaging, and wafer foundry services, emphasizing that "customers come first, then capital expenditure" while increasing spending.
Second, the accelerating commercialization of end-user applications is providing evidence for the rationale behind capital expenditure. Google Cloud revenue grew by 82%, with operating profit approximately 3.1 times that of the same period last year. AI infrastructure investments are rapidly converting into cloud revenue and profits. OpenAI, alongside releasing GPT-5.6, launched an integrated application combining ChatGPT and Codex. Bolstered by model updates, product upgrades, and refreshed usage quotas, the weekly active users of this integrated application surged from 2 million in March to 10 million in July. Anthropic's ARR growth momentum also continues, with third-party data suggesting Anthropic's ARR had surpassed $70 billion by July. Image 1: Weekly active users of the ChatGPT+Codex integrated application jumped from 2 million to 10 million in four months, demonstrating accelerating monetization of AI applications, Source: @thsottiaux.
Finally, models continue to iterate rapidly, simultaneously achieving performance improvements and cost reductions. GPT-5.6 Sol, Terra, and Luna, launched in July, cover flagship, balanced, and low-cost tiers. Luna approaches GPT-5.5 peak performance at less than half the estimated cost of its predecessor, while Terra surpasses the previous generation at an even lower cost. Opus 5.0 offers higher performance at the same price as version 4.8. Significant progress is also evident in the open-source camp. The Kimi K3 release demonstrates performance in high-value-added scenarios like coding and EDA that rivals Opus 4.8, even showcasing a demo of chip design using the open-source model. The Pareto frontier, representing the optimal balance between model performance and price, is continuously expanding outward. Image 2: Large model capability-price Pareto frontier as of July 31st: New generation models consistently offer stronger performance at lower costs, pushing the frontier curve outwards and upwards, Source: arena.ai.
In stark contrast to the bustling industry, the capital market was jumpy, with any negative news triggering a storm. Early in the month, Bloomberg reported that Meta was considering "selling computing power," leading the market to immediately extrapolate an "oversupply of AI computing power and cautious spending by big tech," causing a two-day decline in the hardware supply chain and a circuit breaker in South Korea. KIS cut earnings forecasts for SK Hynix, citing slower HBM long-term contract price growth compared to DRAM, causing SK Hynix's stock to fall 15% on the day. Google's Q2 earnings report revealed negative free cash flow, sparking investor concerns about future enterprise value. Momentum factor unwinding was the core driver of this correction. The SOX index retraced roughly 20% from its peak, with a single-week decline of about 10%. However, the broader S&P 500 remained relatively stable and the VIX stayed low during the same period, indicating the pressure was concentrated on AI and high-momentum assets rather than representing systemic risk. The significant decline in Goldman Sachs' momentum factor index (GSPRHIMO) corroborates this view. The market didn't suddenly start worrying about AI development; rather, everyone was previously standing on one side of the boat. When the wind shifted, everyone rushed to move to the other side, fearful of being left behind and becoming a bloody stepping stone. Image 3: Goldman Sachs High Volatility Momentum Hedge Portfolio (GSPRHIMO) experienced a significant drawdown recently, synchronizing with the deep decline in SOX and diverging from the broader market trend. This confirms that the core of this decline is momentum factor unwinding rather than fundamental weakening, Source: Goldman Sachs.
Short-term stock prices are determined by capital flows, but long-term stock prices are determined by company value. Therefore, only by better understanding the views and judgments of both bulls and bears regarding the future of AI can investors more rationally and objectively assess a company's long-term value, thereby making better investment decisions.
The first battleground for bulls and bears lies in capital expenditure. The bearish argument suggests that the intensity of capital expenditure has reached its peak. If capital expenditure truly reaches the market's anticipated ~$1.3 trillion in 2027, then the free cash flow of several major CSPs would all turn negative, leaving no room for further upward revisions in future spending. This concern intensified after Google's quarterly free cash flow turned negative. Bears also argue that if capital expenditure is revised upwards and CSP free cash flow turns negative, company valuations would be severely impacted; conversely, if capital expenditure is not revised upwards, the hardware supply chain would face immense pressure. The bullish counterargument is that free cash flow is not the sole source for capital expenditure. Firstly, current capital expenditure can be funded through various off-balance-sheet methods, such as SPVs and operating leases. Secondly, CSPs currently have stable core businesses, healthy balance sheets, and access to various financing methods to raise funds. Moreover, the bearish static consideration of free cash flow fails to account for the revenue and margin improvements generated by AI. Essentially, if an investment's ROIC is sufficiently high, it will attract capital from all sectors of the economy, much like the golden age of real estate in China's past.
The second battleground revolves around AI demand. Bears argue that the massive capital expenditure does not match current AI revenue. Even though large model companies are seeing rapid revenue growth in the coding domain, this single area cannot generate sufficient returns, and the next potential growth catalyst is not yet visible. Furthermore, with the rapid development of open-source models, large model companies may be forced to cut prices due to competition, leading to revenue growth without profit growth, ultimately failing to generate adequate returns. The bullish perspective counters that one cannot forcefully match current ARR with capital expenditure. Capital expenditure is a leading indicator, while ARR is a lagging indicator. Therefore, the focus should be on the lifecycle ROIC. Amazon's Q2 earnings report specifically mentioned that AI infrastructure construction can break even within just three years. Currently, server lifespans are at least 5 to 6 years, and data center lifespans can exceed 30 years, providing strong evidence that the ROIC can work. Beyond coding, AI is rapidly penetrating fields like healthcare, finance, law, and science. It remains unclear which area AI will first achieve mass adoption, but just as few predicted coding would be AI's first breakthrough application before it happened, bulls remain optimistic about AI's future application potential. Regarding open-source models, bulls believe there remains a significant capability gap between open-source and closed-source models, but many investors lack the ability to distinguish between them. On one hand, models are fine-tuned specifically for public leaderboards, making it difficult to discern capability gaps in benchmarks. On the other hand, as most investors execute relatively simple tasks, they cannot truly differentiate between the capabilities of open and closed-source models. Concurrently, current open-source models are also moving away from being fully open; for example, the latest Kimi K3 does not use the MIT license and imposes restrictions on commercial use. A likely future scenario is that 90% of tokens are processed by low-cost, less intelligent models, but these tokens generate only 10% of the value, while 10% of tokens are processed by higher-cost, more intelligent models, generating 90% of the value. Image 4: The capability gap between leading and catching-up models only becomes apparent when task difficulty enters the "deep end" — overly simple tasks and public leaderboards mask this true moat, Graphic: Gemini AI.
The constraints mentioned by the bears certainly exist, but they appear more to be pacing issues encountered during the development of a new technology rather than directional problems or core contradictions. The underlying driving force of AI development remains unchanged: smarter models, lower costs, more users, and broader applications are continuously evolving. Short-term stock prices can be swayed by capital flows, positioning, interest rates, expectations, and narratives, but technological development is not constrained by narratives, nor should investor opinions be dictated by stock prices. The long-term value of technology is ultimately determined by productivity and production relations, and enterprise value is not defined by short-term stock prices. Image 5: Market narratives swing back and forth amidst violent turbulence, while the curve of technology and productivity consistently trends upwards — short-term noise does not alter the long-term direction, Graphic: Gemini AI.
Markets can flip from euphoria to panic within days, and back again from panic to euphoria within a single day. Technology and productivity, however, do not halt their advance due to a single news headline, and companies do not undergo earth-shattering transformations in a few days. July was a highly instructive month: the industry side delivered nearly perfect report cards, yet the market oscillated repeatedly. This "short-term divergence between performance and stock prices" is precisely the window of opportunity long-term investors need. We do not predict whether the market will rise or fall on the next trading day; that is not within our circle of competence. What we see, however, is that the AI train has not slowed down; it is accelerating. Short-term market noise will persist, but every instance of mispricing caused by such noise offers a chance for long-term investors to examine and evaluate. Our view is not born of blind optimism, but from current industry logic and data. The louder the clamor, the greater the need for composure.
Disclaimer: Markets carry risk, and investment requires caution. Under no circumstances should the information in this article serve solely as a reference for readers. Companies mentioned are for illustrating industry logic only. All content is without investment advice. Readers should not rely on this article to replace their own independent judgment. East Harbor is not responsible for any investment losses, risks, or disputes arising from the use, citation, or reference of this content.
Comments