Summary of Key Points
Recently, the news that Meta plans to sell its idle computing power has sparked panic in the market about an oversupply of AI computing power, causing the stock prices of major storage chip companies in Japan and South Korea (such as Samsung and SK Hynix) to plummet. However, subsequent announcements by South Korean companies about large-scale investments in AI, along with expert interpretations suggesting that the issue is one of structural mismatch rather than a general surplus, have led to a rapid rebound in stock prices. At the same time, the investment logic in the AI industry is changing: there is a shift from a previous rush for chips without regard for costs to a focus on return on investment. In the future, there will still be a shortage of high-end computing power, while there may be a surplus of mid- and low-end computing power, marking the beginning of a more refined operational phase in the industry.
The Stock Price “Carnival Ride”: The Logic Behind the Panic and Recovery
On July 1st, when news broke that Meta intended to sell its computing power, South Korean stock markets tumbled—The KOSPI index fell by over 8% during trading, with Samsung’s stock price dropping more than 10% (a decrease of 21% from its June high), and SK Hynix’s stock price plummeting by 14.57%, eroding a market value of $160 billion in just one day. Why was there such panic? The market feared that AI investment was overheating, leading to an oversupply of computing power and a subsequent decline in chip demand.
But on the 3rd, stock prices rebounded: Samsung’s stock rose by 8.04%, and SK Hynix’ by 9.92%. There were two main reasons for this turnaround: first, South Korean companies took immediate action by announcing massive investments—billions of South Korean won to build new chip factories, which provided confidence to investors with concrete financial commitments; second, experts clarified that the situation was not one of overall surplus, easing market fears. In short, the market was initially shocked by the bad news but later realized the situation wasn’t as dire, leading to a recovery in prices.
South Korean Giants “Betting Big” on AI: Why Invest Despite Volatility?
Samsung and SK Hynix have not only stayed put but increased their investments: Samsung plans to invest 140 trillion South Korean won (approximately 730 billion RMB) in building HBM chip factories and AI server packaging lines, while SK Hynix is investing 100 trillion South Korean won in NAND flash memory and advanced packaging technologies. Together with their previous total investments of over 4755 trillion South Korean won, this represents a full commitment to the AI industry.
Why are they so bold? First, the South Korean government is providing support, working together with businesses to develop the Chungcheong region as an “AI innovation center” with both policy and financial backing. Second, they believe in the long-term viability of the AI trend; even if there are short-term fluctuations, the AI industry is expected to grow rapidly. By investing in high-end technologies (such as HBM, essential for training large models), they aim to secure a competitive position in the future market. In the words of experts, this strategy is a way to “stabilize investors” and gain a strategic advantage.
Is the Claim of Excess Computing Power a Myth? — Structural Mismatch Is the Reality
Many people wonder if Meta’s decision to sell computing power indicates an actual surplus. Experts argue that this is not the case, and the problem lies in a structural mismatch:
- **Meta is selling “old computing power”: It is not the high-end power used for training large models like ChatGPT (such as the latest GB200 chips), but rather older models like H100/H200, or idle “inference power” (used for tasks like responding to user queries and generating images). Using this old power is costly, so it makes more sense to replace it with newer technologies.
- There is still a severe shortage of high-end computing power: Meta continues to purchase new high-end GPUs and expand its data centers, indicating ongoing demand. Additionally, companies are willing to pay a premium for Meta’s computing power, suggesting that the demand has not diminished. It’s similar to how an old computer at home might be idle, but there is still a high demand for new computers capable of running demanding games—high-end resources are always in short supply, while lower-end ones may experience surplus.
The Shift in AI Investment Logic: From Rush for Chips to Focus on Return on Investment
During the initial boom in AI, companies disregarded costs and bought as many chips as possible. Now, things have changed:
- Cloud providers are more cautious with their purchases: Leading companies (like Meta) are considering the return on investment, questioning whether the money spent will generate profits. Therefore, they are more selective when buying upstream products like storage chips, prioritizing those with good cost-performance ratios.
- Refined operations have become the norm: Companies are no longer just focusing on the chips themselves but also on the overall efficiency of systems—considering factors such as energy-efficient power supply, cost-effective cooling solutions, and optimal rack configurations to maximize computing power output. It’s like running a restaurant; previously, they only needed good ingredients, but now they also need to optimize water, electricity, gas, and seating capacity to achieve maximum efficiency.
- Investors are more pragmatic: They no longer invest based on hype alone but look for tangible profits. For example, if you run an AI company, you must demonstrate how much money can be earned using the computing power, otherwise, investors won’t invest.
Future Trends: High-End Computing Power in Short Supply, Mid- and Low-End Power in Surplus
In the long term, the AI industry is still in its early stages. Applications such as intelligent industrial robots and multi-modal AI (combining text, images, and video) will require large amounts of computing power for inference tasks. However:
- High-end computing power (for training large models): Will remain in short supply as these models become more complex and demanding, with only a few companies (like NVIDIA and Samsung’s HBM) capable of producing it.
- Mid- and low-end computing power: May experience local surpluses, especially for older inference technologies and standard chips. Companies will need to optimize their resources by selling excess power or upgrading to more efficient systems to avoid waste.
In summary, the AI industry is not on the decline but is transitioning from a period of rapid growth to a phase of more refined operations. Only those who can manage costs effectively and understand the technology will thrive in this new environment.
(The entire text is explained in plain language, avoiding technical jargon, making it accessible to non-financial professionals.)