THE STRUCTURAL CHANGE

Many investors believe they hold diversified global equity portfolios, yet that diversification may be providing less protection than expected. As artificial intelligence plays a larger role in market valuations, markets that once moved for different reasons are becoming exposed to the same underlying source of risk.

That shared exposure is most evident in chip-heavy markets. South Korea, Taiwan and the United States remain distinct economies, yet their largest technology companies increasingly respond to the same cycle of semiconductor demand, data-centre investment and confidence in future AI revenue, even while broader and equal-weighted indices behave differently.

The pattern became increasingly apparent by late July 2026, when the 60-day correlation between the Kospi and the Nasdaq 100 rose to roughly 0.50, its highest level since 2021. It did so against a backdrop of extreme market concentration. Samsung Electronics and SK Hynix together accounted for more than half of the Kospi, while the ten largest constituents of the S&P 500 represented close to 37% of the index, well above the dot-com-era peak of approximately 27%.

Neither rising correlation nor high market concentration proves that a lasting shift has taken place. Markets can move together temporarily during periods of stress, while a small group of companies can dominate an index without affecting other markets. The significance lies in the combination. When concentrated markets also begin moving more closely together, shocks can spread more easily across borders through the same AI-linked companies and supply chains.

The figures will inevitably change, but the pattern has endured long enough to suggest a shift in how these markets behave.

 

HOW THE SHOCK SPREAD

The pattern became visible in the final week of July 2026, when two developments in China unsettled semiconductor investors.

One was a report that a state-backed manufacturer had begun producing locally developed chipmaking equipment. The other was the Shanghai debut of ChangXin Memory Technologies, China’s largest memory-chip producer, whose shares rose 466% on their first day of trading. Together, these developments suggested that China was beginning to reduce its reliance on foreign suppliers across key parts of the semiconductor supply chain.

China’s new deep-ultraviolet machines still lagged behind ASML’s systems and remained well behind the extreme-ultraviolet equipment needed to produce the most advanced AI chips. Even so, the news arrived while semiconductor shares were already under pressure from concerns about the scale and financing of AI investment. Reports that Nvidia could support a large share of OpenAI’s future funding had added to questions about how closely investment, demand and financing had become linked across the industry.

The selling spread quickly across chip-heavy markets. On 28 July, the Kospi fell 8% during morning trading, triggering a market-wide circuit breaker, before closing down 10.8%. Taiwan’s TAIEX fell 4.65% and the Philadelphia Semiconductor Index lost 4.5%, yet the broader US market moved in the opposite direction. The Dow rose by more than 1%, the S&P 500 edged higher and the equal-weighted S&P 500 reached a record.

Yet this did not mean national markets had merged into one. Chip and AI-linked shares were repricing together across borders, while less exposed parts of those markets continued to trade on their own fundamentals.

The decline reversed on 31 July after strong earnings from Microsoft and Amazon had lifted US technology shares in the previous session. The Kospi gained as much as 17% at one stage and the TAIEX closed almost 8% higher. The speed of the recovery suggested that leverage and market positioning had amplified the selloff, and the same connections that carried the shock across markets also carried the rebound.

 

WHY THE CONNECTION IS SO TIGHT

Market concentration helps explain why a change in sentiment around a few companies can move an entire index. The deeper link between markets, however, comes from the way capital circulates within the AI industry.

Chipmakers invest in model developers, model developers purchase computing capacity from cloud providers, and cloud providers spend heavily on chips and data centres. A single AI project can therefore appear as an investment on one company’s balance sheet, contracted revenue on another’s, and future demand for a third.

South Korea and Taiwan sit at the manufacturing end of this cycle. Their markets open before Wall Street and often react first when confidence in AI spending changes. Because a small number of semiconductor companies carry so much weight in their indices, those reactions can be severe even when the broader US market barely moves.

There is a historical parallel. During the telecom build-out of the late 1990s, equipment makers such as Lucent and Nortel used vendor financing to help customers buy their products. This did not mean demand was false, but the financing sometimes made orders appear stronger and more independent than they were.

Independence is harder to establish today. Capital moves through multiple companies and balance sheets before reaching its destination. As a result, no single set of accounts tells the full story, making it difficult for investors to determine how much spending reflects real demand.

 

GROWTH AND WHAT IT COSTS

The financial links within the AI industry do not make demand artificial, but they make it harder to judge where that demand originates. The test is whether adoption is spreading beyond the companies financing one another, and whether the returns eventually justify the cost of building the infrastructure.

The results reported at the end of July gave a split answer. Microsoft 365 Copilot passed 30 million paid seats, while Azure, Google Cloud and AWS grew by 43%, 82% and 37% respectively. Microsoft also said that all the quarter’s backlog growth came from customers outside frontier-model companies. Google Cloud and AWS widened their margins as they expanded, suggesting that growth was supported by paying demand rather than bought through broad discounting.

The cost of meeting that demand continued to rise. Microsoft’s free cash flow fell 23% to $19.6 billion as capital spending reached $41 billion. Meta produced almost $32 billion from operations but spent just over $31 billion on new capacity, leaving less than $1 billion behind. Amazon’s headline profit was heavily boosted by gains on its investment in Anthropic, another example of how financial links within the industry can make reported numbers harder to interpret.

There is a fair defence of this spending. Data centres and chips are paid for when they are built, while the revenue they support may arrive over many years. Pressure on cash flow during construction does not mean the investment will fail, but it places greater weight on revenue that has not yet been earned.

Each reporting season can therefore be judged in the same way. Is adoption continuing to spread beyond the companies funding one another, and is the cash left after the spending growing or shrinking?

For now, adoption is broadening while the build-out is leaving less cash behind.

 

THE BET INSIDE THE INDEX

The question is not whether AI is a bubble. It is whether broad portfolios still provide meaningful diversification when one economic driver sits beneath a growing share of global market value.

The comparison with the late 1990s matters, but not because today’s leading companies lack revenue or profits. Many are among the strongest businesses in the market, and paid demand for AI services is expanding. The similarity lies in the scale of capacity being built before the returns on that capacity are known. The risk is that demand arrives too slowly, or generates too little cash, to justify the capital already committed.

The July selloff was only a partial test. It began with a supply-side threat and remained concentrated in chip-heavy markets, while the equal-weighted S&P 500 reached a record. A more serious test would come from the other direction, if investors began to doubt that customer adoption and cash returns could keep pace with spending. That shock would reach not only the suppliers of AI infrastructure, but also many of the companies carrying the greatest weight in the world’s largest indices.

The pattern is not permanent. It would weaken if paid adoption continued to broaden, cash returns improved, or correlations between chip-heavy markets returned towards their historical range. Any of those outcomes would suggest that AI was maturing into a wider and more independently funded source of economic growth rather than remaining a concentrated investment cycle.

Until then, broad indices carry far more exposure to one economic outcome than their number of holdings suggests. That exposure did not require an explicit decision. It accumulated as confidence in the AI build-out lifted the value of the companies leading it and increased their weight within the indices that hold them.

The result is a portfolio or index that appears diversified on paper, while remaining increasingly dependent on a single economic driver.

 

Disclaimer:

The article represents the personal views of the author and does not constitute investment advice.

This publication is provided for general information and market commentary purposes only. It does not constitute investment advice, nor does it constitute an offer, solicitation or recommendation to buy, sell or hold any financial product or to adopt any investment strategy.