Follow the money into the machine

You do not commit US$29 billion to a share buyback to make a subtle point.

In August 2026, South Korean chipmaker SK Hynix committed to buying back and cancelling 24 million shares and lifted its shareholder-return target to more than half the free cash flow generated between 2025 and 2027. As investors questioned how long the artificial intelligence spending boom could last, the company placed an extraordinary amount of money behind its answer.

Its confidence makes more sense once you follow what happens every time somebody uses an AI model. Data must move rapidly between processors and memory. SK Hynix leads the market for the high-bandwidth memory performing that job, while TSMC manufactures and packages many of the advanced processors installed in data centres.

Models can be copied or overtaken. Manufacturing their components at enormous scale and with microscopic failure rates is near-impossible to imitate.

The falling price of intelligence

Asia is no longer competing underneath the bonnet.

Years of targeted and escalating restrictions which limited China’s access to advanced chips did not hamper local players like DeepSeek, Alibaba and Moonshot AI. These firms responded by extracting more intelligence from the computing capacity available to them, competing aggressively on price and by offering open weights that businesses can use.

More efficient models might initially sound like bad news for chipmakers. If each AI task requires less computing power, demand for hardware should fall. But greater efficiency also makes AI cheaper, opening it to more businesses, products and applications. Total demand for computing power can therefore rise even as the amount required for each task falls.

This helps explain why greater model efficiency has coincided with extraordinary physical investment. Combined capital expenditure by Amazon, Alphabet, Microsoft, Meta and Oracle is expected to approach US$800 billion in 2026, more than five times its 2022 level.

Chart 1: Capital expenditure – top five hyperscalers

That spending is flowing beyond advanced processors into memory, servers and networking equipment. But investment cannot immediately create the specialised manufacturing capacity each requires. As demand has accelerated, bottlenecks have emerged across the supply chain.

Where demand meets its limits

High-bandwidth memory is among the tightest bottlenecks. An advanced processor can perform calculations at speed only if memory can keep it supplied with data. Capacity for the latest generations is allocated well into 2027, giving leading producers unusual visibility over demand and pricing.

Producing processors and memory is only part of the job. They must also be assembled and connected so they can work together inside a data centre. TSMC’s capacity to perform this advanced assembly is fully booked, while orders for some essential materials extend into 2027. Linking thousands of chips also requires faster connections, increasing demand for equipment that can move enormous volumes of data between them.

The AI supply chain, as the Future Fund has worked out, is a series of interdependent constraints. Each one directs spending towards another part of the value chain, allowing earnings to broaden from AI accelerator designers into memory producers, foundries and other parts of this ecosystem.

Following the earnings

The broadening is already playing out. 12-month forward earnings for global semiconductor companies have increased by approximately 510% over the past five years, compared with about 150% for the Nasdaq 100.

Share prices have historically tended to follow earnings. Recent weakness has opened a gap between the two, suggesting investors have become more cautious about the durability of AI capital expenditure before analysts have made corresponding reductions to semiconductor earnings forecasts.

Chart 2: Semiconductor prices versus earnings growth (% change)

That caution is also visible in valuations. As at 28 August 2026, the MSCI World Semiconductors Index traded at 18.2 times 12-month forward earnings, compared with 21.5 times for the Nasdaq-100. That discount exists despite semiconductor earnings having grown considerably faster over the past five years.

Chart 3: Global semiconductors trade at a forward P/E discount to the Nasdaq 100

The combination of stronger earnings and discounted valuations creates a more compelling starting point for investors, although the benefits may not be shared evenly. Semiconductor cycles can turn quickly as new capacity comes online, while companies across the value chain have different margins and sensitivities to demand. Exposure to geopolitical tensions also varies across the supply chain.

All this means the opportunity is broadening, making careful selection more important.

Redrawing the investment map

The semiconductor supercycle has prompted a rethink among Australian investors who have traditionally approached emerging markets as a commodities or US dollar play.

South Korea and Taiwan control manufacturing capabilities that some of the world’s richest companies cannot readily reproduce. China is competing through the cost, efficiency and accessibility of its AI models. Important parts of the AI economy are not simply spilling into emerging markets. They are being developed and built there.

That shift is visible inside the VanEck MSCI Multifactor Emerging Markets Equity ETF (ASX: EMKT). As at 31 July 2026, information technology represented 44.5% of the portfolio, while TSMC, Samsung Electronics and SK Hynix were its three largest holdings and together accounted for about 30%.

EMKT is not a semiconductor ETF. Its index applies value, momentum, low size and quality factors across emerging market companies, capturing the changing composition of the asset class while remaining selective about an uneven opportunity.

The VanEck Semiconductor ETF (ASX: SMHG) begins from the opposite direction, investing purely in the global semiconductor value chain.

AI is becoming cheaper and easier to access while remaining dependent on scarce manufacturing capabilities. And while new models will continue to arrive with great fanfare, a less visible contest over manufacturing, packaging, connectivity and scale is well underway.

SK Hynix’s US$29 billion decision suggests the companies controlling those constraints increasingly have the earnings and cash flow to show for it.