The global AI sector is facing questions about potential overinvestment in chips and data centers, yet a leading Chinese AI company has been forced to temporarily turn away new customers due to overwhelming demand.
On the evening of July 19th, Moonshot AI sent a notice to users of its Kimi service. The company stated that within 48 hours of launching the Kimi K3 model, user requests had surged far beyond expectations, pushing its existing computing cluster to its limits. To ensure service quality for existing paying users, Kimi decided to temporarily suspend new consumer subscription sign-ups, prioritizing its available computing power for current subscribers while accelerating capacity expansion.
This move came just three days after the Kimi K3 launch, which many had cited as fresh evidence of a looming AI compute surplus.
Released on July 16th, the Kimi K3 model boasts 2.8 trillion parameters, an open-source framework, multimodal capabilities, and a context window exceeding 1 million tokens. The market was particularly struck by how Moonshot AI, without the massive training resources of top US labs, managed to significantly narrow the gap with leading closed-source models through architectural innovation, synthetic data, and post-training techniques. This path closely mirrors the open-source strategy of DeepSeek, reigniting market fears of a "DeepSeek moment" replay.
For the global AI market undergoing a significant correction, this was quickly interpreted as a more pessimistic investment thesis: if stronger models can be trained with less computing power, have tech companies overestimated compute demand from the start, justifying their trillion-dollar investments in chips and data centers?
Kimi K3 thus became a new rationale for Wall Street to sell AI stocks. Rich Privorotsky, a Goldman Sachs partner and head of its global equities business, characterized the K3 launch as a deep industry signal, warning that the era of aggressive compute expansion may be ending, necessitating a re-evaluation of AI valuation logic.
However, Kimi's subsequent public notice quickly complicated this argument. More efficient model training does not mean running those models requires less compute. On the contrary, as models become more powerful and cheaper to use, users are likely to offload more and more complex tasks to AI.
Kimi's message to Wall Street might be: this is not our fault to bear.
Compute Glut or Underestimated Demand?
Kimi's public notice does not prove that every data center globally will avoid being underutilized.
Some companies may have overestimated customer demand, some chips may be replaced by new architectures before fully depreciating, and some data centers built with high-cost financing may never achieve sufficient utilization. As AI infrastructure investment reaches the trillion-dollar scale, Wall Street's focus is shifting from "are there enough chips?" to "will there be too many chips?".
This concern is not baseless. UBS forecasts that capital expenditure growth for hyperscale cloud providers could slow from 76% in 2026 to 25% in 2027, further decelerating to 6% in 2028. A July Bank of America fund manager survey showed 82% of respondents view semiconductors as the most crowded trade in the current market. As some tech firms increasingly rely on debt and external financing for data center expansion, investors are also demanding clearer evidence of revenue, cash flow, and capital returns.
Yet, a slowdown in capital expenditure growth is not equivalent to a decline in compute demand. An industry cannot sustain investment growth above 70% indefinitely. A deceleration from extreme highs may simply reflect a larger base or signal that supply is gradually catching up with previously pent-up demand. The recent sell-off in AI stocks partly reflects not a sudden loss of customers for infrastructure, but that prior valuations left little room for any moderation in capital spending.
Another market-derived conclusion is also debatable: that improved model training efficiency will reduce the entire industry's need for compute power.
Kimi's experience suggests the opposite. Its "pause on new user registrations due to insufficient compute" refocuses the abstract debate on capital expenditure back to the reality of service delivery. Training a stronger model may not require replicating the resource intensity of top US labs, but once a model gains real users, inference demand can grow even faster. Model efficiency lowers the cost per call, thereby lowering the barrier to use. The more capable and cheaper the model, the more work users delegate to AI, potentially increasing total compute consumption.
The compute needed to train a model and the compute needed to run it post-deployment are distinct questions. Training methods can become more efficient, but once a model attracts a large user base, the growing volume of calls still requires more inference resources.
Furthermore, compute is not a standardized commodity that can be instantly reallocated. Idle servers may not be directly suitable for handling inference tasks for a popular model. Bottlenecks in chips, GPU memory, networking, storage, power, liquid cooling, or software orchestration can all prevent theoretical capacity from translating into stable service. Therefore, it's entirely possible for some data centers to be underutilized while popular AI clusters face supply shortages.
Kimi reveals not a shortage of all compute, but a continued scarcity of "effective compute" that matches real demand. This also implies that the AI market needs to reassess not just compute demand, but also the distribution of industry profits. As AI development moves from purchasing GPUs to server delivery, enterprise deployment, and scaled inference, the growth dividends may spread from a few chip and memory companies to a broader range of players in infrastructure, cloud services, applications, and smart devices.
Wall Street May Simply Be Exiting the "Most Crowded Trades"
The recent sharp correction in the AI market does not necessarily signal the end of the AI theme. A more likely scenario is that capital is rotating out of the most concentrated, highly leveraged, and overstretched trades of the past year.
South Korea serves as the starkest example of this shift.
Driven by the AI chip and memory rally, the KOSPI index was once the world's best-performing major stock index this year. However, Samsung Electronics and SK Hynix alone account for over half of the index's weighting. Even after the Korean market retreated about 25% from its peak, entering bear market territory, it still held a year-to-date gain of nearly 60%. Reuters noted that this selling wave is not purely risk-aversion but more a case of funds taking profits and rebalancing benchmarks after massive gains to reduce portfolio concentration.
Citigroup recently downgraded South Korea from overweight to neutral while upgrading China to overweight within its emerging market allocation. This shift is noteworthy, as Citi did not alter its positive long-term view on the AI trend; it merely reduced its tactical exposure to the Korean AI chip trade. Feedback from Morgan Stanley's recent global roadshows also indicates a resurgence of international investor interest in Chinese stocks.
In this regional rebalancing, Chinese equities are becoming a key destination, with A-shares and Hong Kong stocks playing different roles. A-shares benefit more from policy support, domestic liquidity, and profit recovery in manufacturing and tech hardware. The Hong Kong market, home to internet platforms, AI applications, smart device makers, and globalized tech firms, and still trading at a valuation discount, serves as the primary gateway for international capital allocating to Chinese tech assets.
Following the recent sharp adjustment in the Korean market, the Hang Seng Tech Index has rebounded roughly 12% from its late June low, which analysts see as a sign of global capital rotating from crowded trades into undervalued assets.
This does not mean capital will mechanically "sell Korea, buy China." More accurately, the AI theme is broadening from a narrow trade focused on a few upstream chip companies to encompass a wider industrial chain, including servers, enterprise deployment, internet applications, and smart devices. Capital may be exiting not AI itself, but its most crowded expression.
Post-Rotation, Hong Kong Stocks Await a Set of AI Report Cards
Capital rotating from crowded Korean chip trades to Chinese assets can drive valuation repair, but it cannot substitute for actual profit delivery. TENCENT will release its Q2 results on August 12th, followed by LENOVO GROUP reporting its quarterly data for the period ending June 30th the next day. Other Hong Kong-listed tech leaders like Alibaba and Xiaomi will also enter the earnings spotlight.
This will constitute a more complete set of Chinese AI samples. TENCENT needs to demonstrate that its increasing AI investments are translating into growth for its advertising, gaming, and enterprise services. Alibaba must find a profit balance between accelerating its cloud business and investing in instant retail. LENOVO GROUP needs to prove whether AI terminals like AI PCs can form a new profitable mix, while also demonstrating whether demand for its infrastructure business—AI servers—can move from order books and pipeline opportunities into actual revenue and profit statements.
Lenovo's upcoming earnings report warrants particular scrutiny amid "high expectations." Last quarter, the company's revenue grew 27% year-over-year, with adjusted net profit doubling, significantly raising the market's benchmark. Considering factors like a low base, some pre-stocking, and memory cost pressures, investors should not mechanically extrapolate the 27% revenue growth and profit doubling as a new normal. The more important aspects to watch on August 13th are whether the group can maintain double-digit growth, if profit growth continues to outpace revenue, and if the scale expansion of its Infrastructure Solutions Group (ISG) can be consistently converted into operating profit, all while the PC business maintains its profit foundation amid rising component costs like memory.
Recent rating adjustments show the market has already priced in some optimism. Morgan Stanley upgraded Lenovo to "overweight" with a target price of HK$30; Citigroup and Nomura maintained buy ratings, raising their target prices to HK$31 and HK$35, respectively. The common basis for institutional optimism is growth in AI servers, ISG profit improvement, and potential market share and profit benefits from supply chain capabilities. These upgrades first express confidence in Lenovo's growth continuity and also imply expectations for profit elasticity. If the earnings report shows profit improvement keeping pace with revenue growth, Lenovo's transition from scale expansion to quality growth will be further validated. Should profit performance exceed market expectations, the recent valuation re-rating could gain stronger fundamental support. The market has seen the growth; August 13th may reveal the profit elasticity behind it.
Against the backdrop of the global AI sector's significant pullback, this earnings season is crucial. It must ultimately answer whether the reallocation of overseas capital to Chinese assets is merely a tactical rotation away from high-valuation, high-concentration markets, or a more sustained, profit-supported re-rating.
Kimi's user-driven "circuit breaker" has proven that the real world is not lacking in compute demand. Next, the profit statements of TENCENT, Alibaba, and LENOVO GROUP will collectively test whether this demand can translate into revenue, profit, and cash flow for Hong Kong-listed tech companies along the chain of cloud computing, servers, applications, and smart devices. Among them, Lenovo is not the only sample, but its report comes closest to addressing the question of how compute demand translates into infrastructure profits.
If this set of earnings can provide a positive answer, what Hong Kong stocks may catch is not just capital fleeing crowded trades, but potentially the re-pricing for the next phase of global AI allocation.
Comments