Temasek, Singapore’s S$389 billion state investor, is shifting its AI exposure toward hard assets—semiconductor plants, data centres, and cooling systems—to hedge against disruption, its chief investment officer has indicated. The move mirrors a broader pivot by Asian funds, including Goldman Sachs Asset Management, away from AI applications and toward the physical infrastructure that powers them.
The shift signals deep unease about the sustainability of AI stock valuations. It also underscores Asia’s near-monopoly on the advanced chips and manufacturing capacity that global AI spending depends on.
Asian institutional investors are not just betting on AI. They are betting against the AI stocks everyone else is buying.
The evidence is in where the money is going: not into the next generative AI app, but into the factories, memory chips, and cooling systems that make those apps possible. Temasek, Singapore’s S$389 billion state investor, has flagged AI as both an opportunity and a disruption risk, and is steering its portfolio toward hard assets less exposed to business model upheaval. Goldman Sachs Asset Management’s Asia-Pacific private equity head, Stephanie Hui, says her team is focusing on “simple stuff”—liquid cooling, data centres—rather than trying to pick winners among AI applications.
The shift is not just a defensive play. It is a bet that Asia’s dominance in semiconductor manufacturing—it produces 72% of the world’s chips—will capture the bulk of the value from the AI build-out, regardless of which chatbot or search engine wins.
The hardware hedge
South Korea put a number on the race in April 2024: a KRW 26 trillion (USD 19 billion) support package for its semiconductor industry, with tax incentives and infrastructure backing for advanced chip plants. The government has since classified chips as a national strategic technology, unlocking credits of up to 40% on R&D expenses and 15% on facility investments.
Japan followed with its own arsenal. A revised tax system for fiscal 2024 allows companies to credit up to 30% of qualifying R&D expenses against corporate tax liabilities. A separate law, the Act on Promotion of Investment in Advanced Semiconductors, underpins subsidies for domestic chip plants, including TSMC’s Kumamoto facility.
India is offering the longest runway: multi-year tax holidays of up to 21 years for data centres, stamp duty exemptions, and power tariff concessions. Its production-linked incentive schemes for IT hardware and semiconductors run to 2027.
| Country | Current rule | New rule | Effective date |
|---|---|---|---|
| South Korea | Standard R&D tax credits for semiconductors | Up to 40% R&D expense credit and 15% facility investment credit for national strategic technologies | April 2024 |
| Japan | Existing R&D tax credits | Enhanced R&D credits up to 30% of qualifying expenses; subsidies for advanced fabs under new Act | April 2024 |
| India | Limited incentives | Multi-year tax holidays (up to 21 years for data centres), stamp duty exemptions, power tariff concessions; PLI schemes for IT hardware and semiconductors | 2022–2027 |
The incentives are pulling in private capital. A consortium led by KKR agreed earlier this year to acquire Singapore’s ST Telemedia Global Data Centres for $10.9 billion. Samsung Electronics and SK Hynix have outlined plans for more than $500 billion in new fabrication plants, part of a South Korean push that could mobilise $880 billion in private capital.
Industry association SEMI confirms that Asia, including Japan, accounts for the majority of global semiconductor fabrication capacity, with Taiwan and South Korea dominating advanced logic and memory used in AI accelerators.
The hardware that makes AI possible is not just chips. High-bandwidth memory (HBM) and liquid cooling are the unseen plumbing. HBM stacks DRAM vertically so data moves fast between memory and AI accelerators. Liquid cooling circulates coolant around hot servers, allowing denser racks and higher-power processors. Together, they let large AI models train and run at scale.
The race for hardware dominance is not just about technology. It is about control. South Korea leads in HBM and advanced DRAM through Samsung and SK Hynix. Taiwan controls advanced logic and packaging via TSMC and server ODMs like Hon Hai and Quanta. Japan supplies substrates and components through firms such as Ibiden. In data centres, Singapore and Japan compete to host regional hubs with aggressive incentive regimes. Winning this race means capturing long-term rent from global AI compute demand and strategic leverage over downstream software ecosystems.
The risk is that the entire structure rests on a fragile assumption: that Western hyperscalers will keep spending at current rates. Market data shows that a single export-control announcement can wipe out months of gains in Asian chip stocks. The June selloff that wiped 10% off the KOSPI in five days was a reminder that the hardware trade is not a one-way bet.
Fred Hu, chairman of Primavera Capital Group, has warned that soaring AI valuations and capital inflows raise sustainability questions, comparing parts of the boom to previous technology bubbles. Satoshi Ueyama of Bain Capital Japan has cautioned that some market segments are over-excited and not all investments will succeed at this stage.
The structural advantage
The shift toward hardware is not just a hedge. It is a recognition that Asia’s regulatory environment is tilted toward infrastructure build-out in ways that the West’s is not. South Korea classifies semiconductors as national strategic technologies with enhanced tax credits and streamlined permitting. Japan uses METI-administered subsidy schemes and tax reforms to support advanced fabs and data centres. Taiwan relies on export-control compliance and investment-promotion rules tailored to TSMC and key suppliers. In contrast, the US leans on the CHIPS Act and data-centre zoning rules, while the EU emphasises digital and AI acts targeting usage rather than hardware buildout.
Recent market data show strong flows into Asian semiconductor and data-centre equities, with regional chip indices outperforming broader benchmarks year-to-date. But volatility has spiked around policy headlines. An investor in a Nasdaq-100 ETF may not realise that a fifth of the index’s AI exposure is effectively a bet on Korean memory chips and Taiwanese servers—a link that turns a domestic portfolio into a play on Asian industrial policy.
The pivot to hard assets is, at its core, a bet that the physical bottlenecks of AI—energy, compute, connectivity—will be more durable than the software that runs on them. For Western investors who have ridden the AI stock rally, the message from Asia’s smartest money is clear: the picks and shovels may be the safer bet.
Beyond the headline
The Money Trail
Western cloud giants plan to spend $697 billion on AI infrastructure this year and $873 billion next. Much of that flows to Asian chipmakers and server builders, meaning the real cash-flow beneficiaries of the AI boom may be the hardware suppliers, not the app developers.
The Bigger Picture
The move into data centres and fabs is not just about returns. It is about owning the physical chokepoints—compute, energy, connectivity—that AI cannot function without. In an era of proliferating models, the owners of that infrastructure gain leverage over both economic productivity and data governance, much as oil fields did in an earlier age.
The Reach
A Western pension fund holding a broad AI index is, in effect, exposed to the reliability of South Korea’s tax credits and Taiwan’s export controls. A change in either can ripple through supposedly domestic portfolios via the hardware supply chain, turning a regional policy shift into a global repricing event.
The hardware bet and your portfolio
With Asian funds repositioning toward infrastructure, Western investors face a choice: follow the hardware trail or ride the application wave.
- Western institutional investor
Review the policy incentives driving Asian fab and data-centre projects. Japan’s METI semiconductor portal details subsidy frameworks that could boost returns from TSMC-linked suppliers and packaging firms. India’s production-linked incentive schemes for IT hardware, available on the MeitY programme page, offer long-term tailwinds for data-centre investments tied to localisation trends. Revisit your exposure to Korean memory makers and Taiwanese server ODMs; their earnings are increasingly tied to hyperscaler capex guidance.
- Retail investor in AI-themed funds
Check your fund’s holdings for Asian hardware names. Many AI ETFs are heavy on US software but light on the Asian suppliers that capture the infrastructure spend. A shift in export controls or a cut in hyperscaler budgets could hit these stocks before the applications they support. Consider whether your AI bet is really a bet on Korean fabs and Taiwanese packaging lines.
Explainer
- High-bandwidth memory (HBM)
- A type of DRAM memory stacked vertically and connected with tiny wires, allowing data to move extremely fast between memory and AI processors. It is essential for training large AI models because it prevents the processor from waiting on slower memory. South Korean firms Samsung and SK Hynix control over 90% of the global HBM market.
- Liquid cooling
- A method of removing heat from computer servers by circulating coolant around hot components, rather than using air. It allows data centres to pack more powerful processors into tighter spaces, which is critical for AI workloads. The technology is becoming a standard feature in new hyperscale data centres.
- Hyperscaler
- A large cloud-computing provider, such as Amazon Web Services, Microsoft Azure, or Google Cloud, that operates massive data centres. These companies are the biggest spenders on AI infrastructure, driving demand for chips, servers, and cooling systems. Their capital expenditure plans are closely watched by investors as a leading indicator for the hardware supply chain.
- Semiconductor fabrication plant (fab)
- A factory where silicon wafers are turned into integrated circuits. Building a leading-edge fab costs upwards of $20 billion and takes several years. Asia, particularly Taiwan and South Korea, hosts the majority of the world’s most advanced fabs, giving the region a near-monopoly on the chips used in AI accelerators.
RELATED STORIES
