A wave of unprecedented AI investment is sweeping through global capital markets. Tech giants are competing to stockpile chips, expand data centers, and secure power resources, driving capital expenditure to record highs. Investors are flocking to AI-related sectors, pushing valuations to new peaks.

But as more money flows into data centers and computing infrastructure that have yet to generate meaningful revenue, and as companies turn to debt issuance to fund their AI ambitions, the central question has shifted: Even if AI holds long-term potential to reshape industries, when will the massive capital deployed today translate into actual corporate profits?

This contest is not just about technology—it's a test of capital patience. In the past, markets rewarded those who dared to invest. Today, investors are asking a harder question: after investing, how much will be earned? From American tech behemoths to Korean semiconductor leaders, recent sharp swings in AI-related stock prices reflect a fundamental shift in the underlying logic of global AI investment. Capital is no longer content to pay upfront for a grand future narrative. Instead, it is now meticulously calculating the costs, returns, and payback periods of this "computing arms race." As AI investment transitions from "piling up computing power" to "calculating returns," the second half of the AI investment cycle is just beginning.

Where Is All the AI "Cash Burning" Going?

Imagine this scenario: Managers in the high-end restaurant industry are privately discussing that next year, the wealthy will surely be clamoring for top-grade Wagyu beef. So, restaurants begin stockpiling heavily, using all available cash, loans, and even mortgaging property to order frozen Wagyu, hoping to catch a windfall. However, while the beef sits in cold storage, the restaurants can no longer pay employee wages, utility bills, or supplier invoices. Suppliers come demanding payment, and shareholders and banks, worried about asset impairment, also arrive to reclaim their equity and loans...

In this story, the managers' decision to hoard beef wasn't necessarily wrong, but the beef won't be sold until next year. No matter how much profit it might bring, they must survive this year before they can see any "hope."

The current AI industry is facing a similar predicament: the world's leading tech companies are intensifying computing infrastructure build-out, not only pouring in their own operating cash flow but also continuously "siphoning" financing from financial markets, transferring increasing risk to the capital markets. If commercialization falls short of expectations or corporate operations falter, the once-cherished AI concept can easily devolve into a speculative bubble.

Currently, the scale of AI infrastructure investment has reached historically rare highs for tech industry development. Led by the "Magnificent Seven" U.S. tech stocks, major technology companies are continuously raising capital expenditure, seizing core resources such as data centers, power supply, and high-end chips. According to Reuters analysis, following current trends, the combined capital expenditure of five "hyperscale cloud service providers" could exceed their total free cash flow by 2027. In the second quarter alone, Google's parent company Alphabet burned through $5.9 billion in cash.

In this frenzied AI race, massive infrastructure investment acts like a monster devouring cash flow. Corporate cash reserves deplete rapidly, while debt related to infrastructure that doesn't immediately generate returns continues to pile up. Data shows that Amazon, Google, Meta, and Oracle collectively issued approximately $194 billion in new bonds as of July 7, 2026, a 79% increase year-over-year. Goldman Sachs estimates that the five major hyperscale cloud providers could issue up to $250 billion in bonds this year. A July 2026 report from a UN independent research panel indicates that capital expenditure by major hyperscalers is projected to rise from roughly $150 billion in 2023 to $770 billion in 2026—a fourfold increase in three years. Meanwhile, the annualized revenue of major AI companies has also grown from about $2 billion in 2023 to over $70 billion in 2026.

Reuters had already warned in 2025 that the massive entry of tech giants into the bond market could put enormous pressure on the U.S. corporate bond market. In 2026, this trend has not only continued but intensified, with overall market risk exposure expanding further. As investment becomes increasingly reliant on debt-financed expansion, markets are no longer focused solely on technological prospects but also on whether future cash flows can cover today's debt. Investors are beginning to ask: How much revenue can these data centers actually generate? What are the GPU utilization rates? Can AI services cover the substantial costs of depreciation, electricity, and financing? How much of the massive investment will eventually convert into free cash flow?

The valuation logic of the capital markets for the AI industry has completely shifted. In the first half, capital only cared about the scale of corporate investment. In the second half, the market must precisely calculate input-output returns—this is the core reason behind the recent sharp volatility in AI stocks in both the U.S. and South Korea. Investors are not denying the long-term value of AI, but they are worried about the over-reliance on leveraged expansion and refuse to grant high valuations based solely on management's future blueprints for the AI industry.

From "Stacking Computing Power" to "Calculating Returns"

Recently, the South Korean stock market has experienced extreme volatility in the AI sector, directly reflecting the shift in global capital valuation logic. In June, driven by surging global demand for AI chips and high-bandwidth memory, SK Hynix's stock price surged over 340% year-to-date, briefly overtaking Samsung Electronics in market capitalization. By July, amid deepening doubts about investment returns, SK Hynix and Samsung Electronics faced severe sell-offs. SK Hynix's stock fell over 35% in a single month, while Samsung Electronics dropped over 21%. In just one month, Korea's AI leaders experienced a dramatic reversal from being darlings to being dumped.

Over the past two years, major tech companies have massively invested in data centers, GPUs, network equipment, and memory chips, driving order surges and valuation increases across the entire supply chain. But as industry capital expenditure enters the hundreds of billions of dollars, the inherent cyclical risks of the memory industry have returned to the spotlight: demand booms drive aggressive capacity expansion, but if downstream AI applications fall short of expectations, the currently scarce computing and storage capacity could turn into excess assets, putting downward pressure on product prices. Consequently, investors are no longer simply equating "expanding AI investment" with corporate value growth. They are now cautiously scrutinizing the medium-to-long-term sustainability of each company's expansion plans.

In the U.S., large tech companies are still building data centers at scale, but the market's key monitoring metric has shifted to free cash flow. Continuous capital expenditure expansion leads to significant cash outflows. Once capital spending exceeds a company's internal cash generation capability, the firm must resort to debt, equity financing, or even external capital to sustain expansion. The risk of AI investment has thus transmitted from the industrial end to corporate balance sheets and the entire capital market.

A June study this year pointed out that the current AI market exhibits both genuine industrial fundamentals and localized bubbles. While corporate revenue growth and AI application scale are indeed expanding, in some areas capital expenditure is growing faster than commercialization. Investors have prematurely factored unrealized productivity gains into valuations, leading to overvaluation. This also explains why the recent AI stock correction has been so severe. Previously, markets judged future profitability merely based on AI demand growth projections. Now, capital requires multiple indicators to be cross-validated: data center utilization rates, revenue generation per GPU, customer willingness to pay, and whether efficiency improvements translate into book profits—all must be incorporated into valuation models.

The world is entering the "second half" of AI investment. After this round of screening, companies that can consistently secure orders, generate stable cash flow, and effectively enhance efficiency and reduce costs across the real economy will enjoy valuation premiums. Conversely, companies that rely solely on industry concepts, distant expectations, or external financing to survive will gradually be weeded out by the market.

Who Can Emerge Victorious in the New Race?

Under this new valuation logic, the competitive challenges facing AI industries across different global economies are distinctly different.

In the U.S., the core challenge for companies is how to turn massive capital expenditure into tangible returns. The U.S. boasts world-leading AI model, chip, and cloud computing companies, with mature capital markets and smooth financing channels. However, the more concentrated the advantages, the higher the market's demands for investment returns. For a long time, U.S. tech companies benefited from a relatively loose interest rate environment, where low-cost capital supported large-scale computing expansion. Currently, the Federal Reserve has held the federal funds rate target range steady at 3.5%-3.75%, but factors such as energy shocks from Middle East conflicts have caused energy price volatility and persistently high inflation, intensifying market expectations for rate hikes. This means AI companies can no longer rely on long-term cheap capital to sustain expansion.

From a valuation perspective, rising market interest rates compress the present value of future cash flows. The larger the capital expenditure, the more companies need to present a clear, verifiable profit model to prove that investments will generate sufficient returns. The "second half" of U.S. AI investment is likely to see significant divergence: leading companies with proprietary foundation models, public cloud platforms, and stable government/corporate customer access will retain the capital strength to continue investing. In contrast, companies lacking stable customers, heavily reliant on external financing, or with immature business models will face stricter capital constraints and significantly reduced expansion pace.

In South Korea, the test is somewhat different. As a crucial node in the global AI hardware supply chain, AI-driven demand presents enormous opportunities for Samsung Electronics and SK Hynix, but it may also reawaken the inherent cyclical risks of the Korean semiconductor industry—demand upswings drive simultaneous industry-wide capacity expansion, which can lead to oversupply. Therefore, in the "second half" of Korean AI investment, the market's focus is not on further expanding capacity but on finding a dynamic balance between technology iteration speed, global customer structure, and capacity deployment cycles to avoid cyclical industry losses.

China presents a different industrial landscape. While China's AI supply chain still faces external constraints in certain segments like high-end chips and advanced computing, its complete manufacturing system, massive consumer market, and diverse vertical application scenarios offer an alternative path for AI commercialization. China's core competitive advantage in AI may not lie in replicating the U.S. "mega data center" model, but rather in deeply embedding AI into the entire chains of manufacturing, finance, logistics, healthcare, and consumer industries.

If the U.S. model is analogous to building a nationwide unified AI computing infrastructure highway network, and South Korea is like equipping that highway with key hardware, then China's core opportunity lies in driving various industries across the real economy to fully connect to the computing network, implementing intelligent upgrades in vertical scenarios. This also means the next phase of global AI investment may no longer have a single main storyline: the U.S. focuses more on foundational models and platform construction, South Korea emphasizes chips and memory, while China holds broader prospects in applications and industrial integration.

From a longer-term perspective, AI may be replicating the development paths of general-purpose technologies like the internet and mobile communications: capital rushes in before business models, valuations expand ahead of profits, and eventually bubbles are cleared out. What ultimately survives are companies that genuinely help entire industries improve productivity, reduce costs, and create new demand.

Therefore, instead of asking "Is the AI rally over?", a more meaningful question is: When the capital market no longer simply pays a premium for the "AI" concept, which companies can continuously demonstrate their long-term investment value? Perhaps this is the true core challenge of the "second half" of AI investment.