Meta selling computing power triggers a chain reaction: How does the "computing power surplus" panic impact AI investment logic?

In July 2026, global capital markets are being pulled in two directions at the same time by the actions of two tech giants.

On one side, Samsung Electronics turned in a quarterly performance fit to be written into history—operating profit surged 1,810% year on year to 89.4 trillion won (about $58.4 billion), setting a new record for the third consecutive quarter. On the other, it was reported that social media giant Meta plans to sell off idle AI compute capacity to build a cloud computing business system called “Meta Compute.”

At first glance, the two developments seem unrelated, but they point to the same direction: expectations for investment in AI hardware are undergoing a collective repricing. Companies with the best earnings are being sold off, while companies that hold the most compute capacity are starting to sell “surplus” capacity—market pricing logic is changing in a subtle yet profound way.

How a Storage Supercycle Powered Samsung’s Historic Performance

Samsung Electronics’ surge in performance this time came almost entirely from its memory chip business. According to preliminary financials, in the second quarter of 2026 the company’s revenue reached 171 trillion won, up 129% year on year; operating profit was 89.4 trillion won, up more than 19 times from 4.7 trillion won in the same period last year. This figure even exceeded Samsung Electronics’ total profit over the three years from 2023 to 2025.

The core logic behind the performance is the continued tight supply of memory chips. Strong demand from AI data centers for high-bandwidth memory (HBM) has prompted manufacturers to tilt capacity toward higher-end products, resulting in a shortage of conventional DRAM and NAND memory chips and pushing prices upward across the board. According to HSBC data, in the three months from April to June this year, the average selling price of DRAM rose by more than 40% quarter on quarter, while NAND prices rose by more than 50%.

Samsung’s progress in HBM is especially crucial. Since the sixth-generation high-bandwidth memory, HBM4, began mass production in February 2026, it has significantly improved the profitability of the memory business. Analysts expect the memory chip shortage to last at least until 2027.

However, another aspect of this performance is also worth attention. While the memory business contributed about 112 trillion won in profit, Samsung’s system semiconductor and foundry businesses recorded losses of more than 2 trillion won, and the smartphone division is expected to lose about 1 trillion won. Nearly all of the roughly 90 trillion won in quarterly profit came from a single business line—this kind of “walking on one leg” structural imbalance is a key backdrop for understanding how the market will react next.

The Better the Results, the Harder the Fall: A Classic “Buy the Expectation, Sell the Facts” Tale

The market’s response sharply contrasted with the underlying fundamentals.

After Samsung Electronics released its earnings report on July 7, its stock price dropped sharply right at the open that day. During the session, it even fell below the 300,000 won level at one point, with a decline close to 10%. By the close, the stock was 296,000 won, down 6.92%. Dragged down by it, South Korea’s KOSPI index fell by more than 8% at one point during the session, triggering the sixth circuit breaker of the year.

The root of this unusual pattern—“the better the performance, the harder the fall”—lies in the market’s “anticipation pricing.” Samsung Electronics’ stock price had already risen significantly before the earnings were released—this year alone, the stock has more than doubled. On July 3, Samsung Electronics’ single-day gain reached 8.22%, closing at 309.5 thousand won. The market had already fully priced in the “better-than-expected” results.

A deeper worry is whether AI demand is sustainable. Albert Yong, Managing Partner at Petra Capital Management, noted: “What investors are truly worried about is the sustainability of this AI frenzy and whether U.S. tech giants will slow down their capital expenditures on AI infrastructure.” eToro market analyst Zavier Wong also echoed a similar view: “The stock price has been priced with the ‘historic quarter’ for months. Once the data confirms expectations but does not significantly exceed them, there’s nothing left to reward new investors.”

When “perfection” has already been priced in ahead of time, any “confirmation” that fails to beat expectations can become a reason to sell.

Why Meta’s Compute Rental Plan Triggered a Chain Reaction

Just a week before Samsung Electronics’ stock price crash, another piece of news had already set off a chain reaction among global tech stocks.

On July 1, media reported that Meta is working on a plan for its cloud infrastructure business, intending to sell AI compute and model access permissions to external customers. After the news broke, Meta’s stock surged 8.8%, but the AI hardware supply chain faced collective pressure—CoreWeave plunged 13.92%, while memory chip stocks such as Micron Technology and SanDisk fell by more than 10%. Sentiment then spread to Asian markets: on July 2, Samsung Electronics and SK Hynix each fell 9.06% and 14.57%, respectively.

The logic behind the market panic is not complicated: if even Meta—one of the world’s largest GPU purchasers—starts seeking to sell “surplus” compute capacity, does that mean AI infrastructure buildout has entered a turning point? Over the past two years, the “compute power scarcity” narrative that supported the continuous rise of AI hardware stocks is now facing fundamental doubt.

Meta’s situation itself also provides context. The company expects capital expenditures in 2026 to reach between $125 billion and $145 billion, nearly double 2025. With an investment of this magnitude, there must be a path to returns—and apart from the advertising business, the cloud business is almost the only way to convert idle capacity into revenue. More notably, according to media reports, during an internal employee meeting, Meta CEO Mark Zuckerberg said that in the past 4 months the development of AI agents “has not accelerated the way we expected.” When deployment at the AI application layer moves more slowly than expected, overbuilding on the hardware side becomes even more conspicuous.

Is “Compute Power Oversupply” the Real Question, or a False One?

Market concerns about “compute power oversupply” are being met with rebuttals from different angles.

In its latest report, a Nomura Securities analyst said that market worries about “compute power oversupply” may be overblown. Chip investment in Korea cannot be quickly converted into production capacity, and AI demand is also causing memory chip shortages. A Citigroup research note likewise believes the concern about “compute power oversupply” is exaggerated, as AI demand is still outstripping supply.

CITIC Securities pointed out that the core of Meta’s move is to activate and monetize existing stock of older compute assets; renting out compute capacity and continuing to increase compute investment are not contradictory. TF Securities’ overseas technology team also said the market should not simply understand this as “AI compute demand peaking.” More precisely, Meta is trying to transform AI infrastructure from a cost center into an asset that can be rented out, charged for, and platformized.

A judgment closer to reality is “structural mismatch,” not “comprehensive oversupply.” Industry insiders, citing industry data, say that at present the average effective compute utilization rate in intelligent computing clusters is below 20%, but the shortage of high-end compute needed to support large-model training is about 40%. Low-end compute is idle, while high-end compute is still scarce—this kind of structural divergence is precisely the detail that the market is likely to overlook in panic.

Data from research firm SemiAnalysis provides another perspective as well: in early 2026, among the various GPU on-demand rental capacities it tracked in the market, the available supply at one point was nearly sold out. If compute power truly were comprehensively oversupplied, the rental market would not show such a supply-demand pattern.

When “Expectations” Stay Ahead of “Reality”: What AI Hardware Valuations Are Going Through

Samsung Electronics’ stock plunge and Meta’s compute sale plan may appear to be two independent events, but they share the same underlying logic: the valuation of AI hardware assets has already priced in overly optimistic expectations.

Samsung’s case is especially typical. Operating profit of 89.4 trillion won is a record high, but the stock price had already priced in that expectation before the earnings were released. When the actual figures only “meet” rather than “significantly exceed” the market’s most optimistic estimates, “good news is already priced in” and “sell on the facts” becomes a rational choice for capital. Statistics show that among the eight past preliminary earnings releases from Samsung, the stock price fell on the day of or the day after the release in four cases—“buy the expectation, sell the facts” is not accidental; it is a repeating pattern.

Meta’s compute sale logic can also be understood from a valuation perspective. When a company invests more than $120 billion every year in AI infrastructure buildout, but the market doubts its return prospects, monetizing idle compute becomes a natural choice. This is not the end of AI investment, but a signal that capital expenditures are shifting from “no matter the cost” to a stage of “paying close attention to returns.”

From a more macro viewpoint, the combined investment by Alphabet, Amazon, Microsoft, and Meta in 2026 to expand AI infrastructure is expected to be about $650 billion—an increase of nearly 60% from roughly $410 billion in 2025. With such massive capital expenditures, investors will inevitably scrutinize the return cycle more strictly. When investment reaches this scale, the market becomes even more sensitive to the efficiency of every dollar spent.

From “Compute Power Is King” to “Efficiency Is King”: The Shift in AI Infrastructure Investment Logic

The shared lesson from the Samsung and Meta developments is that AI hardware investment is moving from the “scale race” phase into the “efficiency verification” phase.

Over the past two years, tech giants’ capital expenditure logic has been simple and direct: whoever has the most compute power has the strongest AI competitiveness. GPUs, HBM, data centers—everything aimed at “bigger, faster, more.” This logic helped drive valuation surges for hardware makers such as Nvidia, Samsung, and SK Hynix.

But when capital expenditures reach the scale of hundreds of billions of dollars, and when a major buyer like Meta begins considering renting out idle compute, the market’s focus naturally shifts from “how much was invested” to “what was produced.” The pace of deployment at the AI application layer, the capital expenditure return cycle, and the ability of hardware capacity to be absorbed—issues that were previously overlooked are becoming new focal points for pricing.

Samsung Electronics’ performance itself also provides supporting evidence. The rise in memory chip prices is not driven solely by demand; the shift in production capacity from conventional DRAM to HBM has also played a key role. Price increases driven by “supply discipline” are essentially a fragile balance—if the demand side shows any signs of slowdown, the pricing system could face rapid restructuring.

For investors, this means the analytical framework for AI hardware investment needs to be updated. The simple narrative of “compute power scarcity” is no longer sufficient to support valuations. What replaces it is ongoing interrogation of deeper factors such as demand sustainability, capacity absorption capability, and the monetization pace at the application layer.

Summary

Samsung Electronics’ historic second-quarter 2026 operating profit of 89.4 trillion won, up 1,810% year on year, paired with the capital market reaction of a 6.92% one-day stock plunge, together form 2026’s most representative “earnings paradox.” At the same time, news that Meta plans to sell idle AI compute capacity triggered collective panic among global tech stocks about “compute power oversupply.”

These two developments do not point to the end of AI hardware; they point to a restructuring of expectation management. When the market prices in “perfection” months in advance, and when capital expenditures at the scale of hundreds of billions of dollars start to face return scrutiny, AI infrastructure investment is moving from the “scale race” into the “efficiency verification” stage. Structural shortages in memory chips coexist with structural mismatches in compute capacity—high-end compute power is still in short supply, but market requirements for returns on each investment dollar are becoming stricter than ever.

For investors, this is both a warning of risks and an opportunity to reassess the logic behind AI hardware investment—when the tide shifts from “no matter the cost” to “efficiency matters,” assets with truly sustainable competitive strength will surface amid the volatility.

FAQ

Q: What was Samsung Electronics’ operating profit for Q2 2026 exactly?

According to Samsung Electronics’ preliminary performance released on July 7, operating profit for the second quarter of 2026 was 89.4 trillion won (about $58.4 billion), up 1,810% year on year. This figure is a preliminary estimate; the final earnings report will be released on July 30.

Q: If Samsung Electronics’ results are so strong, why did the stock price fall instead?

The core reason is “buy the expectation, sell the facts.” The market had already fully priced in strong performance before the earnings report, and Samsung Electronics’ stock price has already more than doubled within the year. When actual performance meets expectations but does not significantly exceed them, investors choose to take profits. In addition, concerns about the sustainability of AI demand and a slowdown in capital expenditures are also weighing on the stock price.

Q: Does Meta’s sale of compute mean AI compute is already oversupplied?

Most institutions believe this is a misreading. Meta’s sale of compute capacity is more of a business move to revitalize existing assets and optimize capital expenditure returns rather than a signal that AI demand has peaked. The current issue is more of a “structural mismatch”—low-end compute capacity is idle, but high-end compute capacity needed to support large-model training is still in short supply.

Q: How long is the trend of rising memory chip prices expected to last?

Analysts expect the memory chip shortage to last at least until 2027. Factors such as AI data center demand for HBM and the shift of traditional DRAM and NAND production capacity toward higher-end products are still pushing prices higher. But disagreements in the market over the future price trajectory are growing.

Q: What do these events imply for AI hardware investment?

AI hardware investment is moving from the “scale race” phase into the “efficiency verification” phase. The market’s focus is shifting from “how much compute power was invested” to “how much actual value the compute power generates.” For investors, this means paying closer attention to deeper factors such as demand sustainability, capacity absorption capability, and the pace of monetization at the application layer.