Korea’s Record AI-Chip Rebound: What Amazon’s $220 Billion Bet Really Means

South Korea’s stock market delivered one of the most dramatic reversals in modern market history on July 31, 2026. The benchmark Kospi closed about 18% higher, its largest daily gain on record, as Samsung Electronics and SK hynix surged after Amazon and Microsoft offered fresh evidence that the artificial-intelligence infrastructure boom was still generating extraordinary demand for cloud capacity, servers and advanced memory. The rebound was spectacular, but the more important question for investors is not whether the AI trade came back to life for one session. It is whether the cash flows, supply conditions and customer commitments behind the rally are durable enough to justify the enormous valuations and capital spending now attached to the theme.

The immediate answer is mixed. Amazon Web Services reported 37% year-over-year revenue growth in the second quarter, its fastest expansion in 18 quarters, and Amazon increased its 2026 capital-spending plan to approximately $220 billion. Samsung reported record quarterly revenue and operating profit, while SK hynix said major customers were still requesting more memory and that it had concluded roughly 10 long-term supply agreements. Those are not the fingerprints of an industry in sudden demand collapse. Yet the same cycle is consuming cash at an extraordinary rate, exposing hyperscalers to execution risk and creating severe market fragility when leveraged investors crowd into the same suppliers.

For U.S. readers, the Korean rebound matters because the AI economy is global even when the most visible stocks trade in New York. Nvidia and other accelerator designers depend on foundries, packaging companies and memory suppliers across Asia. Amazon, Microsoft, Alphabet, Meta Platforms and Oracle cannot expand AI services without high-bandwidth memory, server DRAM, networking equipment, power infrastructure and data-center construction. Korea is therefore not a distant side story. It is one of the most important physical bottlenecks in the American AI investment cycle.

Key Takeaways

• Record rebound: The Kospi closed approximately 18% higher on July 31, 2026, its biggest one-day percentage gain on record, after several days of severe declines.
• Amazon’s demand signal: Amazon reported second-quarter AWS sales of $42.2 billion, up 37% year over year, and raised its 2026 capital-spending plan to about $220 billion.
• Memory remains tight: Samsung said AI-driven server demand and limited capacity kept the memory market undersupplied, while SK hynix reported continued requests for additional supply and pursued multi-year customer agreements.
• Cash-flow pressure is real: Amazon’s trailing-12-month free cash flow fell to an outflow of $7.6 billion as purchases of property and equipment rose sharply, mostly because of AI investment.
• The market structure is fragile: Leverage, concentrated index weights, forced selling and the unwinding of a major AI-focused hedge fund amplified the preceding collapse and probably magnified the rebound.
• What comes next: The central test is whether cloud and AI revenue can keep accelerating quickly enough to support hundreds of billions of dollars in annual infrastructure spending without permanently weakening returns on capital.

What Happened in Korea’s Stock Market

The July 31 surge did not occur in an ordinary market. It followed a brutal three-day retreat in which Korean equities repeatedly triggered trading safeguards, leveraged investors faced forced liquidation and the shares of the country’s two largest chipmakers swung by amounts normally associated with distressed companies rather than highly profitable global manufacturers. According to The Wall Street Journal’s market report, five of the 10 largest daily moves in Kospi history had occurred since March by the end of July.

The scale of the rebound reflected both fundamentals and mechanics. Fundamentally, Amazon and Microsoft reassured markets that the largest cloud buyers were still expanding AI capacity. Mechanically, a market that had been aggressively sold, hedged and deleveraged was vulnerable to a sharp reversal once the feared collapse in hyperscaler demand failed to appear in the latest results. Short sellers covered positions, investors who had moved to cash re-entered, and institutions were able to acquire assets after forced sellers had reduced exposure.

Samsung Electronics and SK hynix dominate the Korean index to an unusual degree. Reuters reported during the selloff that the two companies together represented more than half of the Kospi’s market value. When both stocks rise or fall by double digits, the index can behave less like a diversified national benchmark and more like a concentrated semiconductor vehicle. That concentration helps explain why Korean equities became a global proxy for expectations about AI memory, even though the index also includes automakers, banks, internet companies, industrial groups and consumer businesses.

The record daily gain therefore should not be interpreted as a clean referendum on Korea’s entire economy. It was primarily a revaluation of the probability that advanced memory demand would remain exceptionally strong. The market moved from pricing an abrupt deterioration in the AI capital cycle toward pricing renewed confidence that the largest customers were still constrained by insufficient computing capacity. The distinction matters because a one-day change in probability is not the same as proof that the risk has disappeared.

There was also an important calendar effect. The rally arrived at the end of a month in which Korean equities had suffered an extraordinary decline. Portfolio managers who had reduced exposure before major U.S. technology earnings faced the risk of underperforming benchmarks if they remained underweight after Amazon’s and Microsoft’s updates. Month-end rebalancing, benchmark sensitivity and the need to reduce short exposure could all intensify buying even when long-term conviction remained incomplete.

The result was a relief rally with fundamental support, but also a rally shaped by a market that had become disorderly. That combination is why the session was so large. Real demand evidence met an unstable positioning structure.

Why Amazon’s Earnings Changed the Mood

Amazon’s second-quarter report provided exactly the kind of evidence that nervous semiconductor investors needed. The company said consolidated net sales rose 20% year over year to $200.6 billion, while operating income increased 43% to $27.5 billion. AWS revenue reached $42.2 billion, up 37%, and AWS operating income climbed to $16.6 billion from $10.2 billion in the comparable quarter. The cloud unit’s operating margin was approximately 39.4%, indicating that rapid growth was not coming solely from low-margin expansion.

That performance mattered for memory suppliers because cloud revenue is one of the clearest ways to test whether AI infrastructure is beginning to produce commercial returns. Capital expenditure can be justified temporarily by strategic ambition, but the long-term economics require customers to pay for computing, storage, databases, model training, inference and related services. AWS’s acceleration suggested that at least one of the industry’s largest infrastructure platforms was converting capacity into revenue at a faster rate.

Amazon also increased its expected 2026 cash capital spending from roughly $200 billion to about $220 billion. Management attributed the increase partly to higher memory costs and said the company still expected demand to exceed available capacity. That statement has two implications. First, Amazon was not responding to weak demand by slowing construction; it was spending more because customers wanted more computing resources than the installed base could provide. Second, the cost of the buildout itself was rising, which benefits suppliers but raises the hurdle Amazon must clear to earn an acceptable return.

The market focused on the combination of faster AWS growth and higher spending. Either element by itself would have been less reassuring. Higher spending without accelerating cloud revenue could imply an increasingly speculative buildout. Faster cloud revenue without investment might suggest a short-lived capacity squeeze that would later constrain growth. Together, they supported a narrative in which demand was strong enough to justify continued expansion.

Amazon’s report also contained an accounting complication that should not be overlooked. Net income rose to $62.6 billion, but the company said the quarter included $53.4 billion in non-operating pre-tax income, primarily related to its investment in Anthropic. That gain boosted reported earnings but did not represent ordinary operating profit generated by retail, advertising, logistics or AWS. Investors evaluating the sustainability of Amazon’s economics should therefore focus more heavily on segment operating income, cash flow and the relationship between infrastructure spending and AWS growth than on headline net income.

That distinction illustrates a broader principle in AI finance: rising private-company valuations can create large paper gains for strategic investors, but those gains do not fund data centers unless the investment is monetized or otherwise converted into cash. The operating business must still generate cash, debt capacity or external financing to support physical investment.

AWS Is Becoming More Important to Amazon’s Earnings Quality

AWS represented about 21% of Amazon’s quarterly sales but generated roughly 61% of its segment operating income. That imbalance demonstrates why cloud performance has an outsized influence on Amazon’s valuation and capital allocation. The retail operations remain enormous and strategically valuable, but the cloud segment supplies a disproportionate share of operating profit and gives Amazon the financial capacity to invest across AI chips, model services, logistics, robotics and satellite communications.

The operating leverage in AWS can work in both directions. When revenue growth accelerates while existing infrastructure is used more efficiently, margins can rise and cash generation can expand quickly. When the company builds capacity ahead of demand, depreciation, energy costs, networking expenses and personnel costs arrive before the associated revenue. The second-quarter result suggested favorable utilization, but the negative free-cash-flow figure showed how aggressively Amazon was investing for future capacity.

Amazon’s third-quarter guidance called for net sales of $197 billion to $202 billion and operating income of $22.5 billion to $26.5 billion. The company noted that comparisons were affected by the timing of Prime Day and foreign-exchange movements. Those factors are important for the consolidated business, but the strategic question for AI suppliers is simpler: can AWS sustain growth near the second quarter’s pace while capital spending remains elevated?

If AWS growth holds above 30% and margins remain strong, investors may tolerate negative free cash flow for a period because the expenditure appears linked to visible customer demand. If growth decelerates sharply while spending remains fixed, the same projects could be viewed as overbuilding. The market’s interpretation can change quickly because data centers are long-lived assets, whereas quarterly revenue expectations are revised continuously.

The Most Important Number Was Negative Free Cash Flow

Amazon’s trailing-12-month operating cash flow rose 33% to $161.4 billion. Ordinarily, that would be an extraordinary sign of financial strength. Yet the company’s free cash flow, defined in its reporting as operating cash flow minus purchases of property and equipment net of proceeds and incentives, fell to an outflow of $7.6 billion. A year earlier, Amazon had generated a positive $18.2 billion under the same measure.

The reason was a $66.1 billion year-over-year increase in net purchases of property and equipment, primarily reflecting AI investment. Net property and equipment purchases reached about $169 billion over the trailing 12 months. In other words, Amazon’s operating engine generated more cash, but infrastructure investment grew even faster.

This is the central financial tension of the AI boom. The largest technology platforms were once valued partly because software and online services could scale with relatively modest incremental physical capital. AI changes that model. Advanced computing requires enormous clusters of accelerators, memory, networking, cooling systems, electrical equipment and real estate. The hyperscalers are becoming hybrids: software and service companies built on increasingly capital-intensive industrial infrastructure.

Negative free cash flow does not automatically mean the investment is unwise. A company can create substantial value by accepting near-term cash outflows to build assets that earn attractive returns over many years. Amazon itself has a history of investing ahead of demand in fulfillment and cloud infrastructure. The relevant question is the expected return on each additional dollar of capital, not whether spending is high in isolation.

However, the burden of proof rises with the scale of the program. At $220 billion, Amazon’s annual plan is larger than the market capitalization of many major corporations and exceeds the annual economic output of some countries. Small forecasting errors become expensive. A 10% overbuild would represent tens of billions of dollars. Delays in obtaining power, customers, chips or network equipment could reduce utilization. Rapid improvements in hardware efficiency could make older installations less competitive sooner than expected.

For memory suppliers, the spending is revenue. For Amazon shareholders, it is a claim on future cash flow. The same transaction can therefore be bullish for one part of the supply chain and financially demanding for another. That is why the AI trade cannot be assessed as a single homogeneous investment theme.

Why High-Bandwidth Memory Became the AI Bottleneck

Traditional computing discussions often focus on processors. In AI systems, memory has become nearly as important because model performance depends on moving enormous volumes of data between memory and accelerators at very high speed. A powerful graphics processor that cannot access data quickly enough may sit partially idle, reducing the economic value of the entire system.

High-bandwidth memory, commonly called HBM, addresses that problem by stacking multiple DRAM dies vertically and connecting them through advanced interconnects. The architecture provides much greater bandwidth than conventional memory modules while reducing the physical distance data must travel. HBM is expensive, technically demanding and dependent on advanced packaging. Production yields and qualification timelines can therefore limit supply even when manufacturers are willing to invest aggressively.

The AI infrastructure chain requires several difficult technologies to work together. Accelerator designers must deliver chips with competitive performance. Foundries must manufacture the processors and specialized base dies. Memory companies must produce high-quality stacks. Packaging providers must integrate the components. Data-center operators must secure power, cooling and networking. A shortage in any one area can constrain the output of the complete system.

Korea’s strategic importance comes from its leadership in memory. SK hynix has been a major supplier of HBM to the AI accelerator market, while Samsung has invested heavily to expand advanced memory and recover share in leading products. Micron Technology provides additional competition, but the market remains concentrated among a small group of companies with the engineering expertise and fabrication capacity required for high-volume advanced DRAM.

HBM also changes the economics of memory manufacturing. Conventional DRAM has historically been highly cyclical because suppliers add capacity during strong pricing periods, eventually creating oversupply and price declines. HBM requires more wafer capacity per unit of usable output, more sophisticated packaging and closer customer collaboration. Those characteristics can support stronger pricing and longer contracts, but they do not abolish the cycle. If capacity expands faster than AI demand, or if customers redesign systems to use memory more efficiently, pricing can still weaken.

The distinction between temporary scarcity and structural scarcity is therefore essential. Temporary scarcity rewards suppliers until capacity arrives. Structural scarcity can support elevated returns for longer because the technical barriers, qualification requirements and customer relationships are difficult to replicate. The current evidence suggests that HBM remains structurally challenging, but investors should not assume every memory product will share the same economics.

HBM Demand Is Expanding Beyond Training

The first wave of generative AI infrastructure was dominated by training large models. Training consumes vast quantities of computing power but occurs intermittently for each model generation. Inference—the process of using trained models to answer questions, generate content, analyze data or operate software agents—can create a more persistent workload because it occurs every time a customer uses an AI service.

If AI adoption shifts from experimentation to continuous production use, inference demand could broaden the market for memory. Enterprises may deploy models inside customer-service systems, coding tools, financial applications, advertising platforms, industrial operations and consumer devices. Each use case may require different balances of latency, cost, accuracy and privacy, creating demand for multiple memory configurations rather than one standardized product.

Samsung’s second-quarter statement explicitly connected its outlook to agentic AI and expected stronger demand for server DRAM, enterprise solid-state drives and HBM. SK hynix has described a “full-stack AI memory” strategy that includes HBM, DDR5, SOCAMM2, GDDR7, enterprise SSDs and emerging technologies such as compute express link products. The companies are positioning for an ecosystem in which memory demand spreads across the entire data path.

That broadening could make the cycle more durable, but it also complicates forecasting. Training demand can be estimated from announced clusters and model-development plans. Inference demand depends on user adoption, pricing, efficiency, software design and competition. A dramatic reduction in the cost of inference could stimulate usage, but improved algorithms could also reduce the memory required per task. The likely outcome is continued growth accompanied by periods of rapid repricing as supply and efficiency change.

Samsung’s Record Quarter Was Both Bullish and Cautionary

Samsung Electronics reported second-quarter revenue of 171.5 trillion won and operating profit of 89.5 trillion won, both record quarterly figures. Its Device Solutions division, which includes semiconductors, produced 127.5 trillion won in revenue and 89.2 trillion won in operating profit. The company said its memory business achieved record revenue and profit as it focused limited capacity on server products and benefited from higher industry pricing.

Those numbers confirm that the memory shortage was economically significant, not merely a narrative used to support equity valuations. Pricing power and product mix translated into extraordinary profit. Samsung also said it had expanded HBM4 sales, shipped HBM4E samples and expected the market to remain undersupplied in the second half despite some moderation in mobile and personal-computer demand.

Yet the results also reveal how a semiconductor boom can create stress elsewhere inside a diversified technology company. Samsung’s mobile and network businesses reported an operating loss of 0.7 trillion won as component costs rose. The same memory pricing that rewarded the semiconductor division increased costs for devices. This internal transfer illustrates the broader economy: one company’s scarcity rent becomes another company’s margin pressure.

For the AI thesis, Samsung’s product progress matters because competition between Samsung and SK hynix can increase supply and reduce the risk that one producer becomes a permanent bottleneck. For supplier profitability, however, greater competition can eventually weaken pricing. Investors must distinguish between an expanding market and a permanently favorable margin structure. The first can occur without the second.

Samsung also has exposure to foundry manufacturing, displays, smartphones, appliances and automotive electronics. That diversification can provide strategic flexibility, but it means the stock is not a pure HBM investment. Strong memory earnings may be offset by weaker consumer demand or rising component costs elsewhere. Conversely, an improvement in device demand could support the company even if memory pricing normalizes.

The record quarter therefore supports the view that AI infrastructure demand is real, while also warning that scarcity can redistribute profit rather than create it uniformly across the technology sector.

SK hynix: Record Profit, an Earnings Miss and a Lesson in Expectations

SK hynix delivered a record operating profit of 60.5 trillion won for the second quarter, more than six times the prior-year level. Revenue rose 257% to 79.3 trillion won. Those figures would normally be celebrated. Instead, the shares fell sharply because operating profit and revenue were below elevated analyst expectations and because slower HBM4 shipments delayed some revenue recognition.

This is a classic example of the difference between business performance and stock performance. A company can report spectacular growth and still disappoint if the valuation assumes even better results. During a momentum-driven rally, estimates often rise quickly, and the market begins to price not merely strong demand but flawless execution. Any shipment delay, pricing concession or cautious comment can then produce a disproportionate reaction.

SK hynix said major customers continued to request additional memory supply and that it had concluded discussions on around 10 long-term agreements, typically lasting five years and including safeguards such as deposits. Such agreements can improve demand visibility, support investment planning and reduce exposure to spot-market volatility. They may also limit the upside if market prices rise far above contracted terms.

The company planned to increase 2026 capital spending into the high-40-trillion-won range, compared with 30.2 trillion won in 2025. That expansion indicates confidence, but it also raises the risk that the industry adds capacity near the peak of pricing. Management said investment would be adjusted according to demand, yet semiconductor projects often have long lead times and large sunk costs. Once factories and packaging facilities are committed, the ability to respond to a sudden slowdown may be limited.

SK hynix also reported net cash of 88 trillion won and maintained a longer-term goal of more than 100 trillion won. Investors were increasingly focused on how the company would balance expansion, financial resilience and shareholder returns. This debate is healthy. A memory company can use boom-period cash to build capacity, strengthen its balance sheet, repurchase shares or increase dividends. Each choice has different implications for future returns and cyclicality.

The company’s long-term agreements deserve particular attention. If customers are willing to commit capital and deposits for future supply, that suggests demand is not based solely on vague aspirations. But the quality of those commitments depends on contract terms, cancellation protections, pricing formulas and the creditworthiness of counterparties. Public summaries cannot reveal every detail. Long contracts reduce one form of uncertainty while potentially introducing others.

The Crash Before the Rebound: Leverage Turned a Correction Into a Market Event

To understand the record gain, it is necessary to understand the collapse that preceded it. The Kospi fell nearly 11% on July 28 and dropped as much as 12.6% intraday on July 29 before closing down 6%. Trading was temporarily halted. Reuters estimated that as much as $2.18 trillion in market value had been erased during the rout, while Korean officials met to discuss additional stabilization measures.

The selloff was not driven solely by a sudden change in the expected quantity of AI servers. It was amplified by leverage. Retail investors had used borrowed money and leveraged exchange-traded products to increase exposure to a concentrated group of semiconductor stocks. When prices fell, brokers required additional collateral or closed positions. Those sales pushed prices lower, causing more losses and more forced selling.

This mechanism can produce market moves that appear disconnected from the pace of change in underlying business value. A memory factory does not become 20% less productive because its owner’s stock falls 20% in a day. However, the market-clearing price reflects the urgency of buyers and sellers, not only a calm estimate of future cash flows. If the marginal seller must raise cash immediately, the price can fall far below the level a long-term investor might consider reasonable.

Leverage also explains why a rebound can be equally violent. Once prices stabilize and new information reduces the perceived probability of a demand collapse, short sellers buy shares to close positions and underweight investors rebuild exposure. The market can rise faster than long-term earnings estimates because positioning changes more quickly than fundamental forecasts.

Korean policymakers had already attempted to curb the use of single-stock leveraged products. During the rout, the finance minister apologized for the introduction of such products and regulators discussed investment caps, higher trading costs and simulated-trading requirements. The concern was not that leverage is always inappropriate, but that concentrated leverage in a volatile index can create feedback loops with broader financial and political consequences.

The policy challenge is difficult. Restrictions introduced after a collapse may reduce future risk, but they can also force additional deleveraging at unfavorable prices. Broad bans can push activity into less transparent products or offshore markets. Effective regulation usually focuses on disclosure, suitability, collateral, liquidity and the ability of investors to understand path-dependent returns rather than relying only on product labels.

Why Leveraged ETFs Can Diverge From a Long-Term Thesis

A daily leveraged ETF is designed to deliver a multiple of a security’s or index’s daily return, before fees and other effects. It is not guaranteed to deliver the same multiple over weeks or months. In a volatile market, daily rebalancing can cause performance to erode even if the underlying asset eventually returns to its starting level.

Consider a simplified example. A stock falls 20% from 100 to 80 and then rises 25% to return to 100. The unleveraged investor is back at the starting point. A two-times daily product falls 40% to 60 and then rises 50% to 90. It remains 10% below its starting value. The mathematics become more damaging when volatility is repeated and financing costs are included.

This effect can encourage short-term trading and create mechanical flows near the market close as funds rebalance exposure. When a large proportion of activity concentrates in the same names, the products can amplify volatility without changing the underlying companies’ operating outlook.

The Korean episode is therefore relevant to U.S. investors even if they never buy a Korean security. Single-stock leveraged ETFs and options have expanded in the United States as well. The lesson is that a correct long-term thesis can still produce severe losses if the instrument, leverage or holding period is mismatched with the thesis.

The Situational Awareness Unwind: A Case Study in Being Right Too Early, Too Large or Too Leveraged

The market turmoil also intersected with the forced reduction of a prominent AI-focused hedge fund. Reuters reported that Situational Awareness, run by former OpenAI researcher Leopold Aschenbrenner, sold most of its public-equity portfolio to Ken Griffin’s Citadel after heavy losses. The fund had generated a reported 439% return from the beginning of 2026 through June but was down about 67% in July, according to a fund letter cited by The Wall Street Journal and summarized by Reuters.

The portfolio included major AI-related companies, and part of it had been financed with leverage from prime brokers. Reuters reported that Goldman Sachs, JPMorgan Chase, Bank of America and Citigroup helped facilitate the transaction. Citadel acquired the leveraged portion of the public portfolio, while Situational Awareness retained a smaller book that included private investments such as Anthropic.

The episode is important not because one fund determines the value of the AI industry, but because it demonstrates how market structure can overwhelm a thematic forecast. A manager may correctly anticipate that AI infrastructure demand will grow for years and still suffer catastrophic losses if the portfolio is concentrated, financed with unstable leverage or dependent on continuous liquidity.

Leverage compresses time. An unleveraged investor can wait through a drawdown if the thesis remains intact and no cash is needed. A leveraged investor may be forced to sell before the expected business outcome arrives. The lender controls the survival horizon. That can transform a long-duration technological forecast into a short-duration funding problem.

The sale also shows why large, liquid firms can benefit during disorder. A buyer with available capital and risk capacity can acquire assets from a seller whose priority is immediate balance-sheet relief. The buyer does not need certainty that the assets are undervalued; it needs favorable terms, diversification and the capacity to absorb volatility.

It would be incorrect to conclude that Citadel’s purchase proved the AI stocks were cheap. The transaction was privately negotiated, and public reporting did not disclose every price, hedge or financing term. It is more accurate to say that the deal removed a source of forced supply and transferred risk from a pressured holder to a better-capitalized one. That change could improve market liquidity without resolving the long-term valuation debate.

The event also weakens any claim that the July decline was purely a fundamental rejection of AI. Forced liquidation, crowded positioning and prime-broker risk management were part of the move. At the same time, leverage alone cannot explain why prices had become vulnerable. Investors were willing to reduce exposure because valuations and expectations had risen to levels where even record profits could disappoint.

The Global AI Capital-Spending Cycle Is Larger Than Any One Company

Amazon’s $220 billion plan sits inside a broader investment wave. Microsoft, Alphabet, Meta Platforms and Oracle have also committed enormous sums to data centers, servers, chips, networking and power. Reuters reported on July 22 that current-year consensus capital-expenditure estimates for five major hyperscalers had risen from roughly $485 billion in January to about $730 billion in July. A separate Reuters report cited UBS estimates that hyperscaler capital spending could rise 76% in 2026 to $673 billion, then slow to 25% growth in 2027 and 6% in 2028.

The estimates differ because definitions, fiscal years and company sets vary. Some include finance leases, some emphasize cash capital expenditure, and companies do not consistently separate AI spending from ordinary cloud, logistics or office investment. The precise aggregate should therefore be treated as an estimate rather than an audited industry total. The direction is unmistakable: technology infrastructure spending has entered a scale previously associated with national industrial programs.

For suppliers, the size of the cycle provides revenue visibility. Data-center projects require long-lead electrical equipment, cooling systems, land, construction, fiber, servers and memory. Orders can persist even if quarterly equity sentiment changes. For hyperscalers, the scale increases operating and financial risk because assets may take years to produce returns.

The capital cycle also extends beyond listed technology companies. Utilities must build generation and transmission. Private data-center developers raise debt and equity. Governments offer tax incentives and accelerate permitting. Semiconductor manufacturers invest in fabrication plants and packaging capacity. Equipment makers expand production. The AI economy is therefore becoming integrated with industrial policy, energy markets, municipal planning and credit markets.

That integration can make the boom more durable because many projects are supported by long-term contracts and strategic priorities. It can also make a downturn more consequential. If expected AI returns disappoint, the effects could spread through corporate debt, utility forecasts, construction employment, semiconductor pricing and regional tax bases.

Capital Expenditure Is Not the Same as AI Revenue

Investors often compare one company’s capital-spending plan with another company’s AI revenue, but the figures are not directly comparable. Capital expenditure purchases long-lived assets that support multiple services over several years. Revenue is recognized as customers consume services or buy products. One data center may support cloud computing, databases, conventional enterprise workloads, advertising systems and AI models simultaneously.

Amazon’s spending also includes areas beyond AI, although management said AI was the primary driver of the increase. Microsoft’s reported capital expenditure can include finance leases. Alphabet’s spending supports search, YouTube, cloud and internal systems. Meta invests in recommendation engines, advertising and future model capacity. Oracle’s infrastructure is tied to both cloud contracts and financing arrangements.

The economic analysis therefore requires matching asset lives, utilization and gross profit rather than simply dividing annual AI revenue by annual capital expenditure. A facility built in 2026 may generate revenue through the 2030s. The risk is that the hardware inside it can become obsolete much faster than the building or power connection.

Accelerators and memory may have useful economic lives shorter than the supporting real estate and electrical systems. New generations can provide better performance per watt, reducing the competitiveness of older clusters. Companies can mitigate this risk by using older hardware for less demanding inference or conventional workloads, but residual values remain uncertain.

The most informative disclosures would include utilization, contracted capacity, customer commitments, depreciation assumptions, returns by generation of hardware and the proportion of spending tied to secured revenue. Public companies generally provide only part of that information. Investors are therefore forced to infer returns from cloud growth, margins, cash flow and management commentary.

The Bull Case: Why the AI Memory Boom Could Last Longer Than Skeptics Expect

The strongest bullish argument begins with actual demand rather than speculative possibility. AWS grew 37%. Samsung said the market remained undersupplied. SK hynix said major customers wanted more memory and entered long-term agreements. Apple warned that component shortages could affect multiple products. These developments indicate that scarcity was affecting both producers and buyers across different parts of the technology industry.

Second, AI usage may still be in an early adoption phase. Many enterprises are testing tools but have not yet redesigned core workflows around them. Consumer usage is growing, but autonomous agents, multimodal services, industrial applications and real-time personalized systems remain underdeveloped. If these workloads move into production, inference demand could expand for years.

Third, lower unit costs may increase total demand rather than reduce it. This is the economic phenomenon often associated with Jevons’ paradox: efficiency improvements can make a resource cheaper to use, encouraging enough additional consumption that total usage rises. If the cost per AI query falls, developers may embed models in more applications and users may interact with them more frequently.

Fourth, high-bandwidth memory is difficult to manufacture. The industry cannot add perfectly substitutable supply overnight. Qualification with major accelerator customers takes time, advanced packaging remains constrained and yields can limit output. Those barriers may allow leading suppliers to maintain pricing and margins longer than in a conventional commodity-memory upswing.

Fifth, the largest buyers have strong balance sheets and strategic incentives to continue investing. Cloud providers fear losing enterprise customers and developer ecosystems if they lack capacity. Even if near-term returns are uncertain, underinvestment could carry a larger strategic cost. This competitive dynamic can sustain spending beyond the level a purely financial model would recommend.

Sixth, cloud growth suggests monetization is becoming visible. Amazon’s second-quarter AWS performance and strong operating margin show that the infrastructure is not merely sitting idle. Microsoft and other providers have also disclosed rising AI-related revenue run rates. The revenue base may still be small relative to aggregate spending, but the direction supports continued deployment.

Seventh, memory demand is diversifying. HBM receives the most attention, but server DRAM, enterprise SSDs, low-power server memory and high-capacity storage are also benefiting. A broader product cycle reduces dependence on one chip specification or one customer program.

Finally, government policy supports domestic semiconductor capacity. South Korea views memory leadership as a strategic asset. The United States and other countries are funding semiconductor manufacturing and infrastructure. Strategic capital can extend investment cycles even when private-market returns are temporarily volatile.

The Bear Case: Why Record Earnings Can Still Mark a Dangerous Stage

The skeptical argument starts with expectations. Samsung and SK hynix produced extraordinary profits, yet SK hynix shares fell because the results were not extraordinary enough. When a market requires perpetual upside surprises, valuation risk is high. Future earnings can grow while the stock declines if the multiple contracts.

Second, capital spending may be approaching a peak growth rate. UBS’s estimates suggested that hyperscaler spending growth could slow sharply after 2026. Suppliers can continue growing when spending rises, but their valuation often depends on acceleration. A move from 76% growth to 25% growth is still substantial investment, yet it can feel like a recession to stocks priced for ever-faster expansion.

Third, free cash flow is deteriorating. Reuters calculated that five major hyperscalers could spend more on capital expenditure than they generate in free cash flow by 2027 under consensus estimates. Amazon had already moved into negative trailing free cash flow. If debt issuance increases and investor demand weakens, financing costs could become a constraint.

Fourth, technology efficiency is unpredictable. Better chips, improved model architectures, quantization, caching, smaller specialized models and software optimization may reduce the computing and memory required for each task. Demand could still rise, but suppliers may overestimate how much hardware is needed per unit of economic output.

Fifth, customer concentration is high. A small group of hyperscalers accounts for a large portion of advanced AI infrastructure purchases. Long-term contracts improve visibility but also expose suppliers to the strategic decisions of a few counterparties. If one customer delays a data-center project or redesigns a system, the impact can be material.

Sixth, supply will respond. Samsung, SK hynix, Micron, foundries, packaging companies and equipment suppliers are investing heavily. Scarcity creates the incentive that eventually reduces scarcity. The timing may be uncertain, but semiconductor history repeatedly shows that high margins attract capacity.

Seventh, geopolitical and trade risks can disrupt both demand and supply. Export controls, restrictions on advanced equipment, tensions around Taiwan, U.S.-China technology competition and energy insecurity can alter production plans. These risks can support prices in the short term by limiting supply, but they can also reduce the addressable market or force inefficient duplication.

Eighth, market leverage can convert ordinary disappointments into systemic-looking events. The Korean rout and the Situational Awareness unwind showed that crowded positions can create forced selling. Even if the industry remains profitable, shareholders may endure extreme volatility.

Ninth, reported profits may contain non-recurring elements. Amazon’s net income was boosted by an Anthropic-related gain, and SK hynix’s net profit included investment gains associated with its Kioxia stake. Operating results were still strong, but headline earnings can overstate recurring cash generation.

Tenth, AI monetization is uneven. Cloud providers can charge for infrastructure, while consumer-platform companies may rely on advertising improvements or future products. The ability to earn direct revenue differs across the buyer group. A generalized “AI spending” number can conceal major differences in business quality.

How to Reconcile the Bull and Bear Cases

The evidence supports neither a simple bubble declaration nor an uncomplicated supercycle narrative. AI infrastructure demand is real, and the largest suppliers are producing record earnings. At the same time, expectations, leverage and capital intensity have created conditions in which strong businesses can experience violent share-price declines.

A useful framework separates four questions:

1. Is end demand growing? AWS growth, customer commitments and memory shortages suggest yes.
2. Are suppliers earning attractive profits? Samsung and SK hynix clearly are, although profit distribution varies by product and company.
3. Are buyers earning adequate returns on capital? Evidence is improving, but the answer remains incomplete because spending is rising faster than free cash flow.
4. Are stock prices reasonable relative to those outcomes? The July volatility shows that market expectations may move far faster than underlying operations.

An investor can be bullish on AI adoption and cautious on particular AI stocks. A company can benefit from strong revenue growth while its shareholders receive poor returns if the initial valuation was too high. Conversely, a temporary market panic can create attractive prices even when the industry’s long-term risks remain substantial.

The correct analytical unit is therefore the individual business and its position in the supply chain. Memory suppliers benefit from scarcity but face cycles. Hyperscalers control customer relationships but fund the infrastructure. Accelerator designers earn high margins but depend on foundries and memory. Utilities gain demand but must finance generation and transmission. Software companies may benefit from lower computing costs without carrying the same capital burden.

Apple’s Supply Warning Confirmed That the Shortage Was Broader Than Data Centers

Apple’s fiscal third-quarter results added another piece of evidence. The company reported quarterly revenue of $109.4 billion and diluted earnings per share of $2.02, but its outlook was affected by supply constraints. According to Investor’s Business Daily’s earnings report, Chief Executive Tim Cook described an extreme memory shortage and warned that the pressure could affect more products in the coming months.

Apple’s experience matters because it competes for components with data-center buyers. Semiconductor manufacturers allocate capacity toward products with the highest strategic value and profitability. When AI servers absorb more advanced memory and fabrication resources, consumer-electronics companies may face higher costs or limited availability.

The effect is not necessarily a direct transfer of identical chips. An iPhone does not use the same HBM stack as an AI accelerator. However, semiconductor supply chains share fabrication equipment, engineering talent, packaging capacity, substrates, chemicals and capital budgets. Manufacturers decide how to allocate investment across product categories. Strong server margins can reduce the incentive to expand lower-margin consumer supply as quickly.

For Apple, supply limitations can delay revenue rather than destroy it if demand remains strong and customers wait for products. Yet prolonged shortages can push buyers toward competitors, reduce promotional flexibility and pressure gross margins. The company’s guidance therefore provided a demand signal for memory suppliers and a risk signal for device makers.

The contrast with Samsung is particularly instructive. Samsung’s semiconductor division benefited from higher memory prices while its mobile division suffered from component inflation. Apple, which does not own a comparable merchant memory business, experiences the cost pressure without the offsetting supplier profit. Vertical integration can therefore change how a company absorbs an industry shortage.

The Bank of Japan and the Yen Added a Macro Layer to the Technology Rally

The July 31 market environment was not shaped by technology earnings alone. The Bank of Japan kept its policy rate at 1% after raising it in June and upgraded its economic outlook, while Japanese authorities intervened to support the yen. Reuters reported that the currency had experienced its largest intraday rise in more than two years during the suspected intervention before giving back part of the gain.

Currency movements matter to Asian technology companies because revenue, costs and investor returns are often denominated in different currencies. A weaker yen can support Japanese exporters’ reported earnings, while a stronger won can reduce the local-currency value of dollar revenue for Korean exporters. Currency hedging can reduce the immediate impact, but it rarely eliminates long-term economic exposure.

The yen also affects global funding. For years, low Japanese interest rates encouraged investors to borrow in yen and purchase higher-yielding assets elsewhere. As Japanese rates rise and currency volatility increases, carry trades can unwind. That process can reduce liquidity across global equities, including technology stocks, even when company fundamentals have not changed.

The Bank of Japan’s decision therefore contributed to a broader environment of uncertain discount rates. The U.S. Federal Reserve had also faced credibility questions and rising long-term Treasury yields. Higher bond yields reduce the present value of distant cash flows and increase the financing cost of data centers. AI companies with profits expected far in the future are particularly sensitive to changes in discount rates.

At the same time, inflationary pressure from energy and semiconductor shortages can keep central banks cautious. The AI buildout is not occurring in a zero-rate world. Companies must earn returns above a higher cost of capital while competing for power, labor and equipment.

The record Korean rebound was therefore a technology event, a positioning event and a macro event. Amazon’s earnings improved the demand outlook, but interest rates and currencies continued to shape how much investors were willing to pay for that demand.

What the Korean Rebound Means for U.S. Investors

U.S. investors can be exposed to the Korean AI-memory cycle even without owning Samsung or SK hynix directly. The connection runs through chip designers, cloud platforms, semiconductor equipment makers, exchange-traded funds, suppliers of power and cooling systems, and multinational companies whose products depend on memory availability. A change in Korean pricing can therefore appear in U.S. earnings through several different channels.

The first channel is cost. Accelerator companies and server manufacturers purchase advanced memory as part of complete computing systems. If HBM prices rise faster than system prices, gross margins can narrow unless the cost is passed to customers. Cloud providers may absorb the increase initially to secure capacity, then raise service prices or slow price reductions. Enterprise customers eventually bear some portion of the cost.

The second channel is volume. Tight memory supply can cap the number of accelerators shipped, limiting the revenue of companies that would otherwise sell more systems. A supplier can report strong orders and a large backlog while recognizing revenue more slowly because one component is unavailable. This makes supply-chain analysis as important as demand analysis.

The third channel is capital intensity. Higher equipment and component prices increase the cost of each data-center project. A fixed annual budget then buys less computing capacity. Amazon’s decision to increase spending partly because of memory costs suggests that the company preferred to protect deployment volume rather than accept a smaller buildout. Not every buyer can do the same.

The fourth channel is valuation. Korean shares can act as a real-time signal of expectations for global AI infrastructure. A sharp move in SK hynix may influence sentiment toward Micron, Nvidia, Broadcom, data-center equipment companies and U.S.-listed semiconductor funds. The relationship is not perfect, but it can become powerful during crowded market periods.

The fifth channel is currency. U.S. investors in foreign securities receive returns affected by both the local share price and exchange rate. A Korean stock can rise in won while producing a smaller dollar return if the won weakens. Currency-hedged and unhedged funds can therefore perform differently even when they own similar companies.

The sixth channel is policy. U.S. export controls, subsidies and trade rules influence where advanced chips can be sold and where new factories are built. Korean companies must navigate relationships with customers and governments in the United States, China, Europe and other markets. Policy can change the economics of capacity long before it changes reported revenue.

Direct Ownership Is Not the Only Way to Express a View

Investors often treat a theme as a choice between buying or avoiding the most obvious beneficiaries. In reality, the AI infrastructure chain contains businesses with different risk profiles. A supplier of electrical equipment may benefit from data-center construction without depending on one generation of HBM. A cloud platform may gain from AI adoption but suffer from capital spending. A software company may benefit from cheaper models while avoiding direct infrastructure ownership.

This does not mean adjacent companies are automatically safer. Utilities face regulatory limits and construction risk. Equipment companies can experience order cycles. Software firms face competition and uncertain pricing. The analytical advantage is simply that the same long-term theme can be approached through businesses whose economics are not identical.

For diversified investors, the Korean episode reinforces the value of looking through fund labels. A broad “technology” or “Asia” fund may be heavily concentrated in a small number of semiconductor companies. An equal-weighted index can behave differently from a market-cap-weighted benchmark. A leveraged product may introduce path dependency that overwhelms the underlying thesis.

None of these observations determines what any individual should buy or sell. They identify the questions required to understand the exposure already present in a portfolio.

China Is the Largest Strategic Uncertainty for Korea’s Memory Leaders

China affects the Korean memory industry as a customer, manufacturing location, competitor and policy risk. Samsung and SK hynix have significant business relationships and production exposure connected to China, while the United States has tightened controls on advanced semiconductor technology. The companies must preserve access to major markets while complying with evolving restrictions.

Chinese memory manufacturers have improved rapidly, particularly in conventional DRAM and NAND. If they expand output, they can pressure pricing in products below the most advanced HBM tier. That may push Korean suppliers to concentrate even more heavily on premium products, where technical barriers and customer qualification provide greater protection.

Competition can arrive indirectly. A Chinese company does not need to lead HBM immediately to affect Korean profitability. By increasing supply in conventional memory, it can reduce prices and force established producers to allocate more capital toward advanced products. That changes the return profile of the entire portfolio.

China is also a major source of demand for smartphones, computers, industrial electronics and cloud services. Weak consumer demand can reduce conventional memory consumption even while AI-server demand remains strong. The industry could therefore experience a two-speed market: scarcity in advanced server products and softer conditions in lower-end devices.

Export controls introduce additional complexity. Restrictions can reduce the addressable market for the most advanced chips, but they can also encourage Chinese customers to buy available products earlier or invest in domestic alternatives. Companies may need separate product designs for different markets. Duplicated supply chains increase cost and reduce efficiency.

For U.S. policymakers, dependence on Korean memory creates both a strategic partnership and a vulnerability. The United States can design leading accelerators and build domestic data centers, but it still relies on Asian manufacturing for critical components. Industrial policy may gradually diversify production, yet advanced memory ecosystems require years of investment and accumulated engineering knowledge.

The long-term competitive outcome will depend on technology, yield, customer trust, capital discipline and government policy rather than one quarter’s market share. Investors should be cautious with precise forecasts because the industry can change materially between product generations.

Power May Become a More Binding Constraint Than Chips

Even if semiconductor supply expands, AI infrastructure cannot grow without electricity. Large data centers require generation, transmission, substations, backup systems and cooling. Project timelines are often determined by grid interconnection rather than the delivery date of servers.

This matters for memory demand because a chip order is economically useful only if the complete system can be installed and powered. Hyperscalers may reserve components ahead of available electricity, creating inventory or delayed deployment. Alternatively, they may shift projects to regions with faster permitting and more reliable power.

Power scarcity can strengthen the position of the largest buyers. Amazon, Microsoft, Alphabet and Meta can sign long-term contracts, finance dedicated generation and negotiate with utilities. Smaller cloud providers may struggle to secure comparable capacity, increasing industry concentration.

The energy mix also affects operating cost and political acceptance. AI facilities can create jobs and tax revenue, but they may compete with households and existing businesses for electricity and water. Communities may demand infrastructure contributions or reject projects. Environmental rules and decarbonization targets can lengthen development schedules.

From a financial perspective, power constraints can produce both overbuilding and underutilization. A company may secure chips before a facility is ready, leaving expensive assets idle. It may also build electrical capacity that becomes unnecessary if hardware efficiency improves. The asset lives of buildings, transformers and accelerators are different, making capital planning unusually difficult.

The Korean memory rally therefore should not be interpreted as proof that every announced AI project will become productive capacity. It indicates that buyers are willing to pay for scarce components. The conversion of those components into profitable services still depends on energy, construction, software and customer demand.

Accounting Questions Investors Should Watch

The AI capital cycle creates accounting issues that can materially affect reported profit. The first is depreciation. Companies estimate how long servers and networking equipment will remain useful. Extending the estimated life reduces annual depreciation expense and increases near-term operating profit, even though cash spending does not change. Shortening the life has the opposite effect.

There is nothing inherently improper about changing useful-life estimates if technology and reuse patterns justify the decision. The analytical concern is comparability. Investors should examine whether earnings growth comes from higher revenue and better utilization or partly from revised depreciation assumptions.

The second issue is finance leases. Some infrastructure is acquired through leases rather than immediate cash purchases. Depending on the metric, reported capital expenditure may include or exclude those commitments. Comparing companies requires consistent treatment.

The third issue is stock-based compensation and strategic investment gains. Amazon’s Anthropic-related gain increased net income without increasing operating cash flow. Similar valuation changes can create large quarterly swings. Analysts should separate operating earnings from investment remeasurement.

The fourth issue is customer deposits and long-term contracts. Deposits can improve supplier cash flow before revenue is recognized. They also signal commitment, but they may create refund or performance obligations. The cash-flow statement, balance sheet and contract terms must be considered together.

The fifth issue is inventory. During shortages, buyers may accumulate components to protect production. Rising inventory can support supplier revenue temporarily but may reduce future orders if customers later draw down stock. Inventory growth should be compared with revenue, backlog and deployment.

The sixth issue is construction in progress. A data center under construction consumes cash but does not yet generate full revenue. Large balances can signal future capacity or project delays. Once the asset enters service, depreciation begins and margins may change.

The seventh issue is impairment. If a facility or hardware generation becomes uneconomic, a company may write down the asset. Impairments are non-cash at the time of recognition, but they confirm that earlier cash investment did not earn the expected return.

A Practical Dashboard for Following the AI-Memory Cycle

No single indicator can determine whether the cycle remains healthy. A useful dashboard combines operational, financial and market measures.

1. Cloud Revenue Growth

AWS, Microsoft Azure, Google Cloud and Oracle cloud growth provide evidence of customer consumption. Accelerating revenue supports the spending thesis, especially when margins remain stable. Slowing growth is not automatically bearish if capacity is constrained, but it requires explanation.

2. Capital Expenditure and Finance Leases

Track both absolute spending and the rate of growth. A high level with slowing growth can still support supplier revenue, but equity valuations may depend on acceleration. Compare spending with operating cash flow and debt issuance.

3. Free Cash Flow

Free cash flow shows whether internal cash generation is keeping pace with investment. Temporary weakness may be acceptable during expansion. Persistent deterioration increases dependence on external financing and raises the importance of returns on capital.

4. Memory Pricing and Product Mix

HBM, server DRAM, conventional DRAM and NAND can move differently. Supplier margins depend on mix as well as volume. A shortage in one product does not guarantee broad pricing power.

5. Capacity, Yield and Qualification

Announcements of factory capacity should be separated from qualified output. New products require acceptable yields and customer approval. Delays can support prices but hurt the company experiencing the delay.

6. Long-Term Supply Agreements

Monitor contract duration, deposits, pricing mechanisms and customer concentration when disclosed. Agreements improve visibility but may cap upside or create obligations.

7. Data-Center Power Availability

Grid connections, energy contracts and construction schedules determine how quickly chips can become revenue-producing systems. Power delays may shift demand between regions or years.

8. Inventory and Backlog

Rising backlog can indicate strong demand, but investors should assess whether it is firm, cancellable or supported by deposits. Inventory growth can reflect preparation for demand or unsold output.

9. Credit Markets

Bond spreads, issuance volume and order-book coverage reveal whether lenders remain willing to finance the buildout. A sharp increase in financing costs could slow projects even when strategic demand remains high.

10. Positioning and Leverage

Fund flows, options activity, margin debt and leveraged-product exposure can explain price volatility that fundamentals alone cannot. Crowded positioning increases the risk of forced moves in both directions.

Three Scenarios for the Next Phase of the AI Infrastructure Cycle

Scenario One: Demand Continues to Outrun Supply

In the strongest scenario, enterprise adoption accelerates, consumer AI usage expands and agentic applications create continuous inference workloads. Hyperscalers maintain high capital spending because new capacity is quickly absorbed. HBM and server-memory supply remain constrained by packaging and qualification. Cloud revenue grows rapidly enough to offset depreciation and financing costs.

Under this outcome, Korean memory suppliers could maintain elevated margins longer than in past cycles. The main risks would be execution, customer concentration and valuation rather than an immediate demand shortage. Device makers could continue facing higher component costs.

Scenario Two: Growth Remains Strong but Supply Catches Up

This may be the most balanced scenario. AI usage grows, but supplier investment gradually reduces shortages. HBM prices normalize, cloud capacity becomes more available and capital-spending growth slows from extreme rates. Revenue continues expanding, but the largest upside surprises become less frequent.

In this environment, stock selection matters more than broad thematic exposure. Suppliers with cost advantages, differentiated products and disciplined capital allocation may outperform. Companies valued solely on scarcity could struggle even while their absolute earnings remain high.

Scenario Three: Monetization Disappoints and Financing Tightens

In the bearish scenario, enterprises fail to generate adequate returns from AI projects, consumer usage grows more slowly than expected or efficiency improvements reduce hardware demand. Cloud revenue decelerates while committed construction continues. Free cash flow deteriorates, credit spreads widen and companies delay projects.

Memory orders would then weaken after a lag because long lead times and contracts initially support shipments. Inventory could rise, pricing could fall and newly built capacity could produce a conventional semiconductor downturn. Leveraged investors would amplify the equity decline.

The July 31 rebound reduced the market’s estimate of the third scenario but did not eliminate it. Amazon’s results provided one quarter of encouraging evidence. The capital cycle will require years of profitable utilization to prove itself.

Timeline: How the Market Reached the July 31 Reversal

• 2024–2025: Generative AI adoption accelerated demand for accelerators, advanced packaging and HBM. Korean memory suppliers invested in new generations and benefited from rising prices.
• Early 2026: Hyperscalers increased capital-spending plans, while AI-linked semiconductor shares became some of the market’s most crowded positions.
• May 2026: The Kospi moved above 7,000 for the first time as Samsung and SK hynix rallied on memory scarcity and AI demand.
• June 2026: The Bank of Japan raised its policy rate to 1%. SK hynix shipped HBM4E samples and prepared for a U.S. ADR listing.
• July 10, 2026: SK hynix listed American depositary receipts on Nasdaq, increasing access for U.S. investors.
• Mid-July 2026: Concerns about valuations, leverage and the sustainability of hyperscaler spending intensified. Korean regulators tightened rules around leveraged products.
• July 28–29, 2026: The Kospi suffered record declines. SK hynix reported record profit but missed elevated estimates. Forced selling and leveraged-product unwinds amplified losses.
• July 30, 2026: Samsung reported record results. Amazon disclosed strong AWS growth and increased its spending plan. Reuters reported that Situational Awareness sold much of its public portfolio to Citadel.
• July 31, 2026: The Kospi closed about 18% higher, led by Samsung and SK hynix, as investors returned to the AI infrastructure trade.

Frequently Asked Questions

Why did the Kospi rise so much on July 31, 2026?

The index rebounded after Amazon and Microsoft reported results that supported continued AI infrastructure demand. Samsung and SK hynix rose sharply because memory is central to that buildout. Short covering, month-end positioning and the removal of forced sellers likely magnified the gain.

Was the rally based only on speculation?

No. Amazon reported 37% AWS growth and increased capital spending, while Samsung and SK hynix reported record or near-record operating results and continued supply constraints. However, the size of the daily move was also influenced by leverage and market mechanics.

What is high-bandwidth memory?

High-bandwidth memory is a form of stacked DRAM designed to move large amounts of data rapidly between memory and processors. It is widely used with advanced AI accelerators because model training and inference require very high data throughput.

Why are Samsung and SK hynix so important to AI?

They are among the world’s leading memory manufacturers and produce HBM, server DRAM and storage products used in AI systems. Their manufacturing scale, engineering expertise and customer qualifications make them critical suppliers.

Does Amazon’s $220 billion plan consist entirely of AI spending?

No. Amazon’s capital spending covers multiple businesses and asset categories. Management said AI was the primary driver of the increase, and higher memory costs contributed, but the total should not be treated as a pure AI figure.

Why did Amazon have negative free cash flow despite strong earnings?

Operating cash flow increased, but net purchases of property and equipment rose even faster. Amazon reported a trailing-12-month free-cash-flow outflow of $7.6 billion, primarily because of AI infrastructure investment.

Was Amazon’s $62.6 billion net income all generated by operations?

No. Amazon said second-quarter net income included $53.4 billion of non-operating pre-tax income, primarily from its Anthropic investment. Segment operating profit and cash flow provide a clearer view of recurring operations.

Why did SK hynix shares fall after reporting record profit?

The result fell below exceptionally high analyst expectations, and some HBM4 shipments were delayed. Stock prices respond to the difference between results and expectations, not only to whether profit is high in absolute terms.

Do long-term memory contracts eliminate the cycle?

No. They can improve visibility and reduce spot-market exposure, but pricing terms, customer concentration, technology changes and future capacity still matter. Contracts may also limit upside when market prices rise.

Could increased supply end the shortage?

Yes, eventually. Samsung, SK hynix, Micron and other suppliers are investing. The timing depends on construction, equipment, yields, packaging and customer qualification. Supply can remain tight longer than expected, but high profits encourage expansion.

How did leverage worsen the Korean selloff?

Borrowed money and leveraged products created forced sales when prices fell. Margin calls and daily fund rebalancing pushed additional shares into the market, causing declines that triggered further liquidation.

What did the Situational Awareness transaction show?

It showed that a long-term AI thesis does not protect a leveraged investor from short-term funding pressure. The transfer of much of the fund’s public portfolio to Citadel removed forced supply but did not settle the valuation debate.

Is the AI infrastructure boom a bubble?

The term is too broad to answer with a single label. Demand, revenue and profits are real, but some valuations and spending assumptions may be excessive. Different companies can have very different outcomes within the same technological cycle.

What is the biggest risk to Korean memory companies?

The largest risks include a slowdown in hyperscaler spending, rapid capacity expansion, customer concentration, technology transitions, Chinese competition, export controls and extreme equity-market positioning.

What is the strongest evidence that demand remains healthy?

Amazon’s AWS acceleration, the increase in its spending plan, Samsung’s undersupply outlook, SK hynix customer requests and long-term agreements, and supply warnings from device makers collectively support continued demand.

What would signal that the cycle is weakening?

Important warning signs would include slowing cloud revenue, project delays, weaker memory pricing, rising customer inventory, reduced capital-spending guidance, deterioration in credit conditions and lower utilization of new data centers.

What Past Semiconductor Cycles Can and Cannot Teach Us

Memory has a long history of alternating between shortage and oversupply. The industry requires enormous fixed investment, products become standardized over time and manufacturers have an incentive to expand when prices are high. Because new fabrication capacity takes years to build, supply often arrives after the original shortage has already begun to ease. The result can be a sharp transition from exceptional profitability to falling prices.

That history is relevant, but HBM is not identical to conventional commodity DRAM. Advanced stacks require close coordination with accelerator designers, specialized packaging and demanding thermal and reliability performance. Customer qualification can be lengthy. A producer cannot always redirect ordinary DRAM capacity into qualified HBM output immediately. These constraints may lengthen the profitable part of the cycle.

The current boom also differs because demand is being led by a small group of exceptionally well-capitalized customers. In earlier personal-computer and smartphone cycles, millions of end buyers indirectly determined demand through device purchases. Today, a handful of hyperscalers make enormous centralized procurement decisions. That can improve visibility when contracts are firm, but it increases the impact of any one customer changing strategy.

Another difference is that AI infrastructure combines technology competition with national strategy. Governments view advanced semiconductors as essential to economic and security policy. Subsidies, export controls and domestic-production incentives can sustain investment that would not occur under purely private return requirements. Such support may reduce the risk of underinvestment but increase the risk of duplicated or uneconomic capacity.

The lesson from history is therefore not that a downturn must begin immediately after record profit. It is that high margins create a supply response and that equity markets usually anticipate the turn before reported earnings decline. Investors who wait for weak financial statements may react late, while investors who assume the peak too early can miss years of strong cash generation.

A disciplined approach watches lead indicators. These include equipment orders, construction schedules, customer deposits, contract pricing, inventory, wafer allocation and changes in capital-spending guidance. No indicator is decisive alone, but together they can reveal whether scarcity is becoming more or less severe.

Valuation Matters Even When the Technology Thesis Is Correct

The July collapse demonstrated that a company’s operational success and its stock return are separate questions. A share price represents the present value of expected future cash flows, adjusted for risk and discounted by interest rates. When expectations become extreme, a company can grow rapidly and still deliver a poor return because the starting price assumed even faster growth.

Semiconductor valuations are especially difficult near cyclical peaks. Current earnings may be unusually high because supply is constrained and customers are willing to pay premium prices. Applying a high multiple to peak earnings can double-count optimism: the investor assumes both that profits remain elevated and that the market should value those profits as if they were stable.

The opposite mistake occurs during downturns. Depressed earnings may cause conventional price-to-earnings ratios to look high even when the stock is near a cyclical low. Analysts often use normalized earnings, replacement cost, free cash flow through a cycle or enterprise value relative to midcycle profit. Each method depends on assumptions that can be wrong.

AI adds another layer because the market may assign strategic value to technology leadership beyond near-term cash flow. A company with superior HBM yields or customer relationships may capture future platforms not reflected in current earnings. Yet strategic value is not unlimited. Competitors can improve, customers can redesign systems and new products can reduce the importance of a current advantage.

For hyperscalers, valuation requires estimating whether capital spending will produce incremental revenue or merely preserve competitive position. Defensive investment can be rational even if the direct return is modest, because failing to invest could weaken the core business. But from a shareholder perspective, necessary spending still consumes cash. A competitive arms race can create enormous industry revenue while reducing returns for the participants financing it.

Interest rates also matter. When long-term Treasury yields rise, investors can earn more from lower-risk assets and must discount distant technology cash flows at a higher rate. A business may report unchanged long-term prospects while its fair value falls because the required return has increased. This is one reason the Federal Reserve and Bank of Japan discussions were relevant to AI stocks during the same week.

Valuation analysis should therefore use ranges rather than one precise target. A realistic model can vary cloud growth, memory pricing, capital intensity, depreciation, tax rates and discount rates. The width of the range is itself informative. When reasonable assumptions produce dramatically different values, position sizing and financial resilience become more important than confidence in a single forecast.

Long-Term Supply Agreements Improve Visibility but Transfer Risk

SK hynix’s disclosure that it had concluded approximately 10 long-term agreements attracted attention because memory has historically been sold with substantial exposure to shorter-term pricing. Multi-year commitments can reduce volatility by giving suppliers confidence to invest and giving buyers assurance that critical components will be available.

However, every contract divides risk rather than eliminating it. A fixed-price agreement protects the buyer if spot prices rise but benefits the supplier if prices fall. An indexed agreement shares market movements but may create disputes over benchmarks. Take-or-pay provisions protect supplier volume, while cancellation clauses protect buyers from technological change. Deposits strengthen commitment but may be refundable under certain conditions.

Technology transitions create another challenge. A customer may commit to a quantity of one HBM generation and later prefer a newer product with better performance per watt. Contracts need qualification, substitution and upgrade provisions. A supplier that guarantees delivery may face penalties if yields are insufficient. A customer that guarantees purchases may receive products that are less competitive than alternatives available later.

Credit quality matters as well. The largest hyperscalers have substantial resources, but parts of the AI ecosystem rely on newer cloud companies and private developers. A contract is only as strong as the counterparty’s ability and willingness to perform. Deposits and parent guarantees can improve protection, yet they do not remove legal and operational risk.

For investors, the existence of long-term agreements is encouraging evidence of demand commitment. The terms determine how much economic value they create. Because detailed contracts are rarely public, analysts should avoid treating headline contract counts as equivalent to guaranteed high-margin revenue.

Industrial Policy Can Support Capacity and Distort Economics

South Korea has strong reasons to protect its semiconductor leadership. Memory exports support employment, research, tax revenue and the trade balance. The government has discussed strategic funds and patient capital for industries including artificial intelligence, defense, aerospace and biotechnology. Such programs can help companies finance infrastructure whose benefits extend beyond private shareholders.

The United States, Japan and European countries are pursuing related policies. Governments want domestic or allied capacity for advanced chips, packaging and data centers. Incentives can accelerate construction, reduce financing costs and diversify supply chains. They can also create political conditions concerning employment, location, technology transfer and national security.

Industrial policy changes the competitive landscape because companies do not face identical capital costs. A subsidized plant can operate profitably at prices that would be unattractive for an unsubsidized competitor. Over time, widespread subsidies can produce excess capacity. Governments may then protect domestic producers through tariffs, procurement rules or export restrictions, fragmenting the global market.

For Korea, alliance relationships are both an opportunity and a constraint. Access to U.S. customers and technology supports advanced products, while operations connected to China require careful compliance with export rules. Strategic alignment can provide subsidies and market access but reduce flexibility.

The economic value of a policy-supported project should therefore be assessed at two levels. The national return may include supply security, employment and technological capability. The shareholder return depends on cash flow after capital cost, operating expense and policy obligations. A project can be strategically successful while producing mediocre private returns.

Risk Management Lessons From the July Volatility

The most immediate lesson is that concentration can hide inside diversification. An investor may own a national index, an Asia fund and a semiconductor fund and believe the positions are separate, while all three depend heavily on the same memory companies. Looking through to underlying holdings is essential.

The second lesson is that leverage changes the definition of risk. Without leverage, volatility is uncomfortable but survivable if the asset remains valuable and the investor has time. With leverage, volatility can create mandatory action. Margin requirements, option expiration and fund rebalancing can force decisions at the worst possible moment.

The third lesson is that liquidity is conditional. A large company may trade actively in normal conditions, but order books can thin during a panic. The available price for a large position may be far below the last quoted price. Funds holding similar assets may become sellers simultaneously.

The fourth lesson is to distinguish thesis risk from financing risk. Thesis risk is the possibility that AI demand or company competitiveness is weaker than expected. Financing risk is the possibility that the investor cannot hold the position long enough to discover whether the thesis is correct. The Situational Awareness episode showed how financing risk can dominate.

The fifth lesson is that hedges can fail to behave as expected. Options may become expensive, correlations can rise and foreign-exchange movements can offset local gains. A hedge should be evaluated under stress scenarios rather than assumed to work because it did historically.

The sixth lesson is that rebalancing rules matter. A portfolio that trims positions after large gains can reduce concentration automatically. A momentum strategy may do the opposite and add exposure as prices rise. Neither approach is universally superior, but the behavior should be understood before volatility arrives.

The seventh lesson is to preserve decision-making capacity. Cash, unencumbered assets and modest position sizes allow an investor to respond to new information. A fully leveraged portfolio delegates decisions to lenders and market prices.

These principles are not predictions that Korean shares will rise or fall. They are general safeguards for markets in which technology optimism, concentrated benchmarks and leverage interact.

Final Assessment: A Real Boom Inside a Fragile Market

Korea’s record rebound was not an illusion. Amazon’s cloud growth, Samsung’s record semiconductor profit and SK hynix’s customer commitments provided concrete evidence that AI infrastructure demand remained powerful at the end of July 2026. The physical economy behind artificial intelligence—memory, servers, data centers and power—continued to expand.

But the rally was also not a final verdict. It occurred after leverage and concentration turned a correction into a historic collapse. Amazon’s free cash flow had moved negative despite strong operating cash generation. Hyperscalers were committing capital at a scale that will require years of successful monetization. Memory suppliers were adding capacity because current prices were attractive, creating the conditions for eventual normalization.

The most defensible conclusion is that the AI infrastructure cycle is both fundamentally real and financially demanding. The Korean market had become too concentrated and leveraged to translate business news into orderly price changes. Record profits did not prevent a crash, and one day of record gains did not remove the long-term risks.

For readers evaluating the sector, the essential task is to separate technological adoption from investment returns. Artificial intelligence can transform the economy while some AI-linked securities disappoint. Capital spending can create valuable infrastructure while weakening near-term cash flow. Memory scarcity can support exceptional profits while encouraging the capacity that eventually reduces those profits.

Amazon’s $220 billion plan strengthened the case that the buildout still had momentum. The next stage will be judged less by announcements and more by utilization, cloud revenue, customer retention, free-cash-flow recovery and returns on invested capital. Korea’s chipmakers are positioned at the center of that test.

Sources

• Amazon.com Announces Second Quarter 2026 Results
• Samsung Electronics Announces Second Quarter 2026 Results
• Reuters: SK hynix record profit misses elevated forecasts
• Reuters: South Korea stock rout breaks records
• The Wall Street Journal: Korea’s Kospi logs record daily gain
• Reuters: Global stocks react to technology results and interest rates
• Reuters: Citadel buys most of Situational Awareness’s public holdings
• Reuters: AI investment boom pressures Big Tech free cash flow
• Reuters: Investors prepare for slower hyperscaler spending growth
• SK hynix 2026 strategy and shareholder meeting summary
• SK hynix HBM4E sample announcement
• Apple Reports Third Quarter 2026 Results
• Investor’s Business Daily: Apple guidance and supply constraints
• Reuters: Bank of Japan holds rates and updates its outlook
• Bank of Japan monetary policy meeting documents

Affiliate disclosure: Businessfinance.news may earn compensation from qualifying actions completed through selected links on this website, at no additional cost to the reader. Affiliate relationships do not influence our editorial reporting, analysis, or conclusions.