Can Asia Tech Continue to Outperform?

Asia EM: Integral to the global AI infrastructure super-cycle

In order to meet the rise in AI demand, the AI capital expenditure continues to be revised upwards, especially for the large hyperscalers in the U.S. which are expected to spend $700bn this year to satisfy the insatiable compute demanded by users. One company’s capex is another’s revenue – 50% of the infrastructure cost of a modern data center today goes towards semiconductors, and the leaders in that space based in Asia.

Meanwhile, an emergent domestic AI ecosystem is taking shape in China, where semiconductor localization has become a national imperative backed by explicit government support. Governments (for good reason) have incentives to retain sovereignty over their AI ecosystems, which have important strategic implications especially with regard to national security.

This is where global fragmentation and inflation intersect. As supply chains regionalize and strategic industries are re-shored, the irreplaceable nodes of the semiconductor value chain command scarcity value and pricing power. Tight supply across leading-edge foundry capacity and memory is translating directly into pricing gains – and higher earnings for the companies. As a result, the Taiwanese and Korean stock markets are up 110% and 170% respectively over the past year, pushing the combined index weight of those two markets to over 50% of EM as a whole. The sector composition of EM has also materially shifted from “old economy” sectors over a decade ago to now being dominated by the tech sector at around 45%.

Taiwan: The Epicenter of Leading-Edge Semiconductor Manufacturing

Taiwan stands as the undisputed cornerstone of the global semiconductor industry—a position that underscores both its strategic significance and its remarkable economic achievement. At the heart of this dominance is the world's largest leading-edge foundry, which manufactures the vast majority of the world's most advanced chips, powering everything from smartphones and data centers to artificial intelligence systems and autonomous vehicles.

Taiwan's success in semiconductors is no accident. It reflects decades of disciplined investment in research and development, world-class engineering talent, and a deeply specialized supply chain ecosystem that is extraordinarily difficult to replicate. The world's ability to satisfy its growing demand for AI compute is largely determined by the capacity this leading foundry has available—and that capacity remains supply-constrained. While global fragmentation is driving incremental foundry capacity to be built in the US, Japan, and Europe, this leading operator remains the partner of choice for government grants and incentives.

For investors, Taiwan represents a compelling structural growth story: as global demand for computing power accelerates—driven by AI—Taiwan-based manufacturers are uniquely positioned to capture outsized value. With technology sector concentration close to 90% of the market, we expect earnings to grow at an impressive 43–47% in 2026 and 23–27% in 2027, keeping us positive.

South Korea: Memory's Growing Centrality and an Unprecedented Supply/Demand Imbalance

Global DRAM (Dynamic Random Access Memory) and NAND supply is dominated by two leading South Korean companies and is currently in significant shortage due to accelerating demand for AI compute. We believe this memory cycle could last longer than prior cycles—a key driver of our constructive view toward South Korean equities.

Memory manufacturers over-invested in 2020–2022, as elevated work-from-home demand during COVID-19 led to excessive supply expansion and a subsequent decline in prices in 2022. However, explosive growth in GPU demand toward the end of 2022 drove a surge in GPU-specific memory—high bandwidth memory (HBM)—a new and fast-growing form with attractive economics. To meet this demand, leading DRAM companies converted significant conventional excess capacity toward HBM. This conversion process is one-way, and it is estimated that the three leading DRAM manufacturers' collective conventional DRAM capacity is roughly 20% lower in 2025 than in 2022.

An unforeseen development is that rapid advances in AI have driven an exponential surge in AI server demand for conventional DRAM to address bottlenecks in training and inference. A key new form factor, SOCAMM (Small Outline Compression Attached Memory Module), offers higher bandwidth with much lower power requirements and is expected to accelerate demand for conventional DRAM. In addition, agentic AI is driving up CPU intensity, and CPUs typically utilize two to four times the memory of GPUs. Collectively, server demand is expected to reach approximately 73% of DRAM demand by 2028, up from 39% in 2025.

The combined effect of reduced conventional DRAM supply and an exponential surge in new forms of DRAM demand is causing the extreme tightness we are witnessing—and the surge in prices. Industry pricing usually declines 20–25% in most years, but these are not usual times. While new capacity is being added, it does not appear likely to alleviate the tightness until at least 2028.

Amid such supply tightness, a number of rumored long-term agreements (LTAs) are being negotiated between large buyers—such as the US hyperscalers—and the major memory manufacturers, in order to incentivize new capacity and reduce the risk of over-expansion. The terms remain confidential but are rumored to include minimum prices, pre-payments, and volume commitments that could reach as much as 60–70% of a hyperscaler's estimated needs. Investors don't typically respect the enforceability of LTAs, but the memory providers are communicating that these agreements carry stronger legal enforcement mechanisms than prior ones.

Similar to DRAM, NAND underwent a severe industry downturn in 2022/2023, with prices bottoming in mid-2025, and AI compute is now driving accelerated demand as inference workloads increase exponentially. With the number of tokens generated growing rapidly, GPU memory limitations are becoming a critical bottleneck—making KV cache (in NAND) offloading an essential method for scaling inference workloads. The KV cache plays a crucial role in the decoding process by storing previously computed data to shorten the time required to generate the next token. Solid-state memory (NAND) is therefore key to optimizing inference models, and demand is driving a significant pricing upcycle. The growth of inference workloads with agentic AI is expected to underpin accelerated NAND growth. With DRAM manufacturers prioritizing investment in areas such as HBM, NAND capacity additions could be slower to react, keeping prices higher for longer.

We now estimate earnings for MSCI Korea to grow by 280–300% in 2026 and 25–30% in 2027 with reasonable valuations near trough levels. While the contribution from memory cannot be understated, it is worth noting that Korean earnings outside of memory are also expected to grow at an impressive 44–45% in 2026 and 20–21% in 2027, driven by favorable exposure to one of our preferred themes: fragmentation. Specifically, Korean industrials benefit from national champions in strategic sectors such as defense and energy infrastructure, which are benefiting from accelerating capital spending budgets and driving above-average earnings growth. In addition, the strength and duration of this memory cycle is generating a significant wealth effect—among employees of these semiconductor companies and retail investors alike—that could further boost domestic cyclical growth in areas such as consumer discretionary and financials. 

China: Growing Self-Sufficiency and Capability in Domestic AI Infrastructure

Amid escalating US–China geopolitical tensions and export bans, China has made localization of domestic semiconductor equipment and chip supply a national imperative. The government is investing heavily in the sector, with large-scale funds supporting domestic AI chip designers and rapid IPO approvals for promising companies. With recent mandates requiring at least 50% domestic equipment in new fabs, the substitution of imported semiconductor equipment is also likely to accelerate. China's semiconductor localization rate is estimated to have reached 41%, up from 30% in 2024. While the US has recently approved the export of a competitive AI chip from a leading US semiconductor company—a meaningfully better alternative to current Chinese domestic supply—the trend toward localization is expected to continue. In particular, domestic chip capability for inference remains lower, and the drive for self-sufficiency is causing demand to outstrip supply, keeping the market for compute in China tight.

While domestic chip companies work to improve chip design and performance, China possesses other strong capabilities and advantages. First, its domestic networking capabilities are world-class, with a history of supplying advanced optical modules to both US and Chinese tech giants. These strengths help compensate for the lower per-chip performance of domestic AI chips by enabling large-scale, high-performance computing clusters. A leading domestic private tech giant, in particular, is innovating with rack-scale solutions, leveraging optical networking to scale up AI compute power—a development expected to drive networking revenues up 90–150% in 2026.

We also view power as a key competitive advantage for China. It is estimated that China added seven times the generation capacity the U.S. added in 2025 alone. Over the past decade, China has invested in electricity generation capacity at a 9.5% CAGR and has ample spare capacity to meet increased demand from AI data centers—even if no new capacity is added for the next two years. Alongside new generation capacity and grid upgrades, this leaves China with a surplus of generation capacity to meet the demands of AI compute. Power is not be a constraint for China's AI ambitions.

On the application front, Chinese AI models offer clear value for money for end users. Ever since a leading Chinese AI lab's model emerged in January 2025, an increasing number of Chinese models across text, image, and video have demonstrated impressive capabilities in competition with US models. Coding is next. Chinese models offer a strong cost-per-performance proposition and are driving accelerating global user adoption and compute demand within China. The same leading Chinese AI lab recently introduced a model optimized for a leading Chinese technology company's chipsets, creating a potentially new AI ecosystem centered in China.

While investor focus remains firmly on the suppliers of AI infrastructure, the same cannot be said for the companies investing in that infrastructure. Significant skepticism persists toward the return on investment for the internet giants and their potential to monetize AI—skepticism that has driven recent underperformance in Chinese offshore equities. We view investors as too pessimistic. With valuations inexpensive and earnings estimates having been reset, the risk/reward skews positively for Chinese equities. However, more patience is likely needed for Chinese companies to realize earnings gains from the AI journey, as the broader macro and fundamental backdrop remains uninspiring, capping the strength of earnings growth in this market in general.

Conclusions and Investment Implications

The AI infrastructure super-cycle has transformed EM Asia from a cyclical, trade-dependent complex into the indispensable supply base of the defining technology of this era. Artificial intelligence is driving demand that structurally outstrips supply across logic, memory, networking, and power. Global fragmentation is entrenching the scarcity value of the irreplaceable nodes in this chain—from Taiwan's leading-edge foundry to Korea's memory duopoly—while spurring China to build a parallel, self-sufficient ecosystem. The resulting supply tightness is generating durable pricing power that pushes against the industry's historically deflationary tendencies, echoing the higher-for-longer inflation dynamic we see across the broader economy. Investors would do well to own scarce assets that are the source of higher prices, rather than areas which may lose pricing and purchasing power.

Each market offers a distinct expression of the same thesis. Taiwan is the irreplaceable manufacturer of leading-edge compute. South Korea is the play on an unprecedented memory shortage, complemented by fragmentation-driven opportunities in industrials. China is the contrarian opportunity—a self-sufficiency build-out that the market has yet to fully price, with structural advantages in networking, power, and increasingly competitive AI models. Underpinning all three is an earnings picture that remains underappreciated (thanks to still reasonable valuations) relative to the scale and duration of demand.

The risks are real and warrant monitoring. A sharper-than-expected capacity ramp-up could soften the memory cycle, geopolitical escalation could disrupt supply chains, and any moderation in AI capital expenditure would call current lofty estimates into question. But on a twelve-month horizon, we believe the balance of risk and reward favors being invested into these markets.