The AI semiconductor sector was dealt a significant blow overnight following a move by Meta, sending shockwaves through related markets in Asia today.
South Korea's KOSPI index closed down nearly 8%, with Samsung Electronics shares falling close to 10% and SK Hynix dropping over 14%.
Japan's Nikkei 225 index fell more than 2%, with shares of Kioxia plummeting over 14%.
The storage-related concept sector in Hong Kong and A-shares markets also saw declines, dragged down by the broader sentiment.
Concerns about a potential oversupply of computing power have once again cast a shadow over the market.
Understanding the Recent Sell-off
In recent years, capital expenditure by US tech giants in the AI field could only be described as frenzied.
Hundreds of billions of dollars have been poured into data centers, computing chips, and expensive large model development.
However, any industry's development is subject to the fundamental economic principle of diminishing marginal returns.
As model parameters have surged from hundreds of billions to trillions, performance improvements have begun to plateau, while the costs of electricity, depreciation, and maintenance to keep these "pipelines" running have risen exponentially.
Major companies have suddenly realized that revenue from consumer payments and cost-saving benefits for enterprise clients currently cannot fill such a massive funding gap.
Meta's organizational restructuring is essentially a sign of management yielding to Wall Street pressure, cutting inefficient operations, and compromising on ROI (Return on Investment).
When an industry transitions from a growth phase focused on "expanding the pie" to a phase focused on "improving internal efficiency," drastic organizational changes are often not a positive sign.
Meta's decisive actions signal a shift towards more refined research and development, moving away from blindly pursuing large, comprehensive model systems towards targeted vertical pipelines that can directly drive advertising conversion or commercialization.
It also indicates a focus on internal cost-cutting, squeezing out "bubble" within AI teams to concentrate valuable computing resources on the most certain core products.
This type of adjustment is not an isolated case.
The growing pains and diverging strategies resulting from AI investments are intensifying among major US tech firms, a change that is profoundly impacting the capital markets.
The initial investment thesis was: High CapEx → Frenzied GPU Purchases → Valuation Inflation for Computing Power Stocks.
Now, the logic has shifted to: Commercialization → Reducing/Optimizing Pipelines → Diverging Demand for Computing Power.
As the near-monopolist in global AI computing power, NVIDIA has captured the industry's most lucrative profit margins in recent years, while Microsoft, as a representative of combined computing power and application, has also seen its valuation soar.
However, as core clients like Meta begin to reassess and restructure their AI pipelines, the market's excess demand for GPUs and HBM memory is theoretically entering a phase of "inventory reduction" or "demand rationalization."
The sharp decline in the AI semiconductor sector last night is a manifestation of this economic reality.
Prior to this, cloud computing providers had already faced widespread skepticism over their high capital expenditures, leading to sustained valuation pressure.
Is a Demand Inflection Point Approaching?
The key question is whether Meta's actions truly signal an inflection point for the previously red-hot AI chip demand.
Objectively speaking, a single statement from Meta is insufficient to determine the future trend for the entire industry.
If other major computing power purchasers like Microsoft, Google, and Amazon follow suit, then caution would indeed be warranted.
It is difficult to predict their future statements, as external observers cannot see inside these companies or know management's current stance on computing power investments; staying updated on the latest news is crucial.
However, it is important to recognize that Meta's situation is somewhat unique, driven by several factors.
First, nearly 98% of Meta's revenue comes from advertising, lacking the stable cloud revenue streams of AWS, Azure, or GCP.
Yet, like cloud providers, Meta spends tens of billions annually on GPUs and data centers, leading to persistent investor skepticism about the return on these investments.
By renting out computing power and model APIs, it aims to transform a pure cost center (data centers) into a revenue-generating asset, amortizing hardware depreciation, creating incremental cash flow, improving its financial narrative, and supporting its stock price.
Previously, its Llama model was open-source and free, building ecosystem value without direct monetization.
Now, with model API fees and computing power rental as two revenue streams, it is creating a second growth curve to reduce the impact of advertising cycle fluctuations on its performance.
Second, it seeks to capture AI ecosystem influence and bind global developers by offering low-cost computing power and free/cheap Llama models to attract small and medium-sized AI companies and developers to build applications on its platform.
Massive external usage and fine-tuning of Llama provide Meta with vast amounts of real-world data and security feedback to iterate its foundational models.
This strategy also aims to prevent developers from flocking entirely to closed ecosystems like OpenAI and Anthropic, consolidating its position as the open-source large model leader.
Third, by becoming a two-way player in the computing power market, it hedges risk.
Previously, Meta was a pure buyer of computing power, reliant on external suppliers.
Now, with its own computing power available for external sale, it gains pricing and allocation influence.
Even if internal demand fluctuates in the future, the rental business can help cover hardware costs, allowing for more aggressive investment in AI infrastructure.
In practice, Meta is renting out idle GPU clusters in its data centers, primarily consisting of previous-generation cards like the H100, for tasks like fine-tuning, inference, and small-scale training, competing with specialized AI cloud providers like CoreWeave on a usage-based fee model.
Internal needs always take priority; all computing power and model resources are first allocated to Meta's own businesses like advertising recommendations, Threads/Instagram, Meta AI, the metaverse, and cutting-edge large model training.
Only idle time slots on older hardware are rented out externally; the latest generation of top-tier training GPUs are reserved entirely for internal use.
Furthermore, Meta's full-year 2026 capital expenditure guidance is $1250–1450 billion, with a significant portion allocated to new high-end GPUs for training next-generation ultra-large-scale models.
If there were a genuine overall computing power surplus, Meta would not need to continue committing hundreds of billions to infrastructure.
In a recent shareholder meeting, Mark Zuckerberg stated that companies approach Meta weekly, willing to rent computing power at rates above Meta's own cost, indicating that overall AI computing power remains tight and the industry is not saturated.
While the AI computing power sector, particularly in US, South Korean, and Taiwanese stocks, has suffered significant declines, more evidence is needed to conclude that an industry-wide demand inflection point has arrived.
For now, this can be viewed as a "panic sell-off" triggered by excessively crowded trading in the AI chip sector.
A Harsh Reality for the Computing Power Industry
Meta's actions have also revealed a harsh truth for the capital markets regarding the computing power industry.
In the past two years, Meta stockpiled a massive number of H100 GPUs for initial Llama model training.
Now, as cutting-edge training shifts to newer hardware, the H100 is no longer the core training workhorse, leading to significant idle capacity that cannot be fully utilized for internal inference tasks.
This directly exposes the capital-intensive, high-depreciation nature of the computing power business.
Just two years ago, NVIDIA's H100 was a hot commodity, commanding high prices with limited availability.
Now, in less than two years, it appears to be facing obsolescence, simply because the GB200 is here, with the Rubin chip on the horizon.
The performance of new products will undoubtedly surpass that of older ones.
With better-performing options available, major companies naturally prioritize them, as the AI race is akin to a car competition; having inferior technology makes it difficult to outpace competitors.
However, capital expenditure has been spent, and depreciation is a fixed cost.
If hardware becomes idle so quickly, before depreciation costs are fully amortized, and external rental demand is weak, the impact on corporate profits and cash flow can be severe.
It is important to remember that Wall Street primarily values major firms like Meta using Discounted Cash Flow (DCF) models.
Once cash flow shows problems, valuation pressure intensifies significantly.
Coupled with the prevalence of quantitative trading on Wall Street, mechanical selling by funds becomes almost inevitable when valuation pressure mounts.
Recently, some pessimistic market commentators have expressed views that Meta's heavy spending on computing power infrastructure could potentially turn its cash flow negative in the second half of the year.
Not only Meta, but other major companies heavily investing in computing power, such as Microsoft, Amazon, Google, Oracle, and pure-play computing power rental firms like CoreWeave, face similar challenges.
If large, established companies have thicker financial cushions to withstand this, pure-play computing power rental companies are in a much more precarious position.
Of course, every major shift creates both winners and losers.
If computing power costs genuinely decline, who stands to benefit?
The answer is straightforward: the midstream and downstream segments of the AI industry.
First, large model companies would benefit, as they would not need to burn as much cash on training and inference, alleviating funding pressure and potentially accelerating model update cycles.
Second, AI application companies would gain, as they could access lower-cost computing power for various development projects, speeding up the commercialization of AI applications.
Final Thoughts
The investment thesis for the AI industry has continuously evolved over years of development.
In the current landscape, the era of "simply buying AI concept stocks and winning" is largely over.
The narrative driving global capital markets regarding AI has long shifted from "grand vision" to "extremely pragmatic profit realization."
This is a core reason why memory chip stocks have performed strongly over the past six months, while the stocks of companies purchasing computing power have remained subdued.
Following Meta's recent moves, at least two critical questions now confront investors.
The first is whether sectors with excessively crowded trading, even if the underlying logic still seems sound but exhibits high volatility, are still worth heavy investment, and whether it is necessary to try to capture the very last dollar of profit.
The second is whether industries currently burdened by computing power costs and capital expenditure pressures are poised for a potential reversal.
Comments