Asia-Pacific Investment Report 2024 — Positions in the Stack

The AI capital cycle described in the 2024 global report reached Asia-Pacific through a different channel than it reached the US, and the difference determined the region's exposure.

In the US, the cycle ran through the companies building models and applications and the hyperscalers funding infrastructure. In Asia-Pacific, it ran predominantly through the physical supply chain — semiconductor manufacturing, memory, advanced packaging, components, servers, cooling and networking equipment.

That is a fundamentally different position in the value chain, and it has a specific characteristic: the region owns a constraint rather than a thesis.

The demand for advanced computation must be met with physical hardware. That hardware requires manufacturing capability that exists in very few places — and advanced-node semiconductor fabrication is arguably the most concentrated critical capability in the global economy, with leading-edge production dependent on a small number of facilities and, further upstream, on lithography equipment produced by essentially one company.

Owning a constraint is a stronger position than owning a thesis. As the 2025 AI report describes, four incompatible theses about where AI value accrues were being funded simultaneously. The supply chain's revenue does not depend on which is right. It depends only on the aggregate volume of computation being built.

That is both the region's strength and the source of its specific risks:

• Demand is derived. If the parties funding the build-out reduce spending — which is the question the 2026 global report frames — orders fall regardless of how good the manufacturing is.
• Concentration attracts policy attention. A capability this concentrated is a strategic asset, which draws restriction, subsidy and relocation efforts from multiple governments.
• Capacity cycles are severe, per the 2017 report's analysis. Capacity added in response to shortage arrives in large increments and typically in excess.

Why owning the constraint is a strong position

The distinction between owning a constraint and owning a thesis is the report's central analytical point.

A thesis position depends on a specific outcome. An investment in a model developer depends on models remaining differentiated. An investment in an application depends on that application building defensibility. If the thesis is wrong, the investment fails regardless of how well it was executed.

A constraint position depends only on the activity happening at all. A manufacturer of a component required for any AI deployment earns revenue whichever layer eventually captures the value. The uncertainty about value accrual — which the 2025 AI report argues is the central open question — does not affect them.

The conditions under which a constraint position holds:

• The constraint must be genuine. There must be no adequate substitute at a reasonable cost.
• It must be difficult to replicate. Advanced semiconductor manufacturing requires capital measured in tens of billions, decades of process knowledge, and supplier ecosystems — meeting this test more comprehensively than almost any other industrial capability.
• It must not be circumvented by a change in approach. This is the real risk: an architectural change that reduces the compute requirement, or shifts it to a different type of hardware, moves the constraint elsewhere.

A constraint position is not permanent. It lasts as long as the constraint does, and constraints move when the technology changes or when enough capital is directed at removing them.

The current period has both forces operating. Enormous capital is being directed at building alternative capacity, with state support in multiple jurisdictions. And architectural change is continuous. Neither has displaced the constraint, and both are working on it.

Concentration as strength and risk

The concentration of advanced semiconductor manufacturing is worth stating plainly because it is unusual in the modern economy.

Most critical inputs have multiple sources. Advanced-node logic fabrication does not. Leading-edge production is concentrated in a small number of facilities in a small number of locations. The lithography equipment required is produced by essentially one company. Several critical materials have similarly concentrated production.

The commercial consequence is pricing power, and it is genuine — a supplier with no adequate substitute captures a large share of the value created downstream.

The strategic consequences are the ones that dominate policy:

• The concentration is geopolitically salient, since a capability essential to both economic and defence applications sits in a small number of locations.
• Governments have responded with subsidy programmes aimed at building domestic capability, in multiple jurisdictions and at very large scale.
• Restriction has been applied to limit certain parties' access to the technology.
• Companies have begun geographic diversification of production, at substantially higher cost.

The investment implications are genuinely two-sided:

• Existing capacity holders benefit from pricing power while the concentration holds.
• Subsidised new capacity may erode returns over time, since capacity built with state support does not need to earn a commercial return on the subsidised portion.
• Geographic diversification raises industry costs, since new locations lack the ecosystem the 2018 report describes.
• Political risk is elevated and difficult to model, as the concentration itself makes the assets strategically significant in ways that commercial analysis does not capture.

The general observation: a capability concentrated enough to confer pricing power is concentrated enough to attract policy intervention. The commercial advantage and the political risk have the same cause, which means they cannot be separated in an assessment.

Power and land as binding constraints

A second constraint emerged in this period that shaped where data centre capacity could be built, and it favours a different set of markets than the semiconductor constraint does.

Data centres require electricity at scale, continuously and reliably. A large facility's power requirement is comparable to a small city's, and it must be available at the site.

This makes power availability, not land or capital, the binding constraint in many markets. The consequences are direct:

• Markets with surplus generation capacity gained advantage. Several Asia-Pacific economies have generation capacity in excess of current demand, which became a competitive asset.
• Grid connection timelines became the critical path. In constrained markets, connecting a new large load can take years, which is longer than building the facility.
• Location decisions shifted toward power rather than proximity. For training workloads, which are less sensitive to latency than serving workloads, the facility can be located wherever power is available.
• Cooling requirements interact with climate, and much of the region is hot and humid, which raises cooling energy requirements and favours locations with cooler climates or access to water.

The investment consequences extend well beyond data centres:

• Power generation and transmission became investable in a way that is directly linked to technology demand — connecting the real assets thesis in the 2018 report to the AI cycle.
• Industrial land near grid capacity appreciated.
• Cooling and electrical equipment experienced demand growth from a new source.

The general point is that a constraint in one part of a system creates opportunity in adjacent parts. The AI build-out's power requirement is a constraint for data centre developers and an opportunity for generation, transmission and equipment — which is a more accessible way to gain exposure to the cycle than competing for the assets at the centre of it.

Restriction deepens and substitution accelerates

Technology restriction, which the 2018 report identifies as a distinct category, deepened through this period.

The measures extended to more advanced computing hardware, the equipment used to manufacture it, and in some cases the software used to design it.

The responses were predictable and have proved durable:

• Domestic substitution accelerated, with very substantial state support directed at building indigenous capability at every layer of the stack.
• Supply chains bifurcated further. Companies serving multiple markets built parallel arrangements to comply with different rules.
• Workarounds emerged and were closed, in a continuing cycle of measure and response.
• Investment in restricted markets acquired a different risk profile, since access to inputs depended on regulatory determinations rather than commercial terms.

The medium-term consequence is the one that matters for allocation. Substitution efforts, given enough capital and time, generally succeed to some degree — the question is how much capability gap remains and how long it persists.

A partially-substituted market has different economics for everyone. Incumbent suppliers lose a market. Domestic substitutes gain one, though often at lower margins and with state support that complicates return analysis. And the global market becomes two markets with different standards, different suppliers and different cost structures — which raises costs for everyone and reduces the scale economies that made the original concentration efficient.

For investors the practical requirement is the one the 2018 report identifies, now more demanding: assessing a technology company in the region requires assessing its exposure to restriction on both sides — which inputs it depends on, which markets it sells into, and whether either is subject to determinations that can change.

How constraints move

A constraint position is durable only while the constraint holds, and constraints move for identifiable reasons. Naming them is what turns "watch the constraint" into a monitoring practice.

Capital removes a constraint by building alternative capacity. This is the slowest route and the most visible. Advanced semiconductor fabrication requires capital measured in tens of billions and years of construction, plus process knowledge that does not transfer with equipment. The lag between the decision and the capacity is long enough to be tracked, and SEMI's free monthly equipment billings by region are the leading indicator — equipment ships before capacity produces.

Architecture removes a constraint by changing the requirement. The more dangerous route because it is faster and less visible. If a change in how models are built or served reduces the computation required, or shifts it to a different type of hardware, the constraint moves without any new capacity being built. This has happened repeatedly in computing and there is no reason to assume it will not again.

Substitution removes a constraint through determined effort. Where state support is directed at building indigenous capability at every layer of a stack, substitution generally succeeds to some degree. The questions are how much capability gap remains and how long it persists — and both are difficult to assess from outside, which means claims in either direction should be treated as uncertain.

Demand collapse removes a constraint by removing the demand. The least discussed and the most immediate. A constraint exists relative to demand; if the derived demand falls, the constraint disappears without anything changing on the supply side. This is the risk the 2026 reports frame, and its evidence appears in the buyers' filings rather than in the suppliers'.

The monitoring implication is that four different series matter: equipment billings for capacity, technical developments for architecture, disclosed programme progress for substitution, and buyers' capex guidance for demand. Only the last is a leading indicator on a short timescale, which is why the 2026 regional outlook points there.

A constraint position is a strong position and a dated one. The useful question is not whether it will end but which of the four routes will end it, because each has a different warning time.

What an allocator could act on

Distinguish constraint exposure from thesis exposure. A supplier whose revenue depends on the activity happening is differently exposed from a company whose value depends on a particular layer capturing the economics. Both are AI exposure and they resolve differently — which means a portfolio holding both is holding two things rather than a concentrated bet.

Accept that concentration's commercial advantage and political risk have the same cause. A capability concentrated enough to confer pricing power is concentrated enough to attract subsidy, restriction and relocation efforts. They cannot be separated in an assessment, and pricing the first without the second is incomplete.

Follow the binding constraint into adjacent industries. The AI build-out's power requirement is a constraint for data centre developers and an opportunity for generation, transmission and electrical equipment. That is frequently a more accessible route to exposure than competing for the assets at the centre, and it connects the real assets analysis in the 2018 report to the AI cycle.

Watch power and grid data, not just chip data. IEA electricity reports and national grid operators publish demand, reserve margins and connection queues free. In many markets the binding constraint on new capacity is a grid connection timeline longer than the construction schedule.

Treat restriction exposure as a standard diligence item. Which inputs a company depends on, whether they have controlled origin, which markets it sells into, and whether substitution is feasible. This is a regulatory and geopolitical assessment rather than a commercial one, and it has become unavoidable for technology investment in the region.

Expect bifurcation to raise costs for everyone. Two markets with different standards, different suppliers and different cost structures reduce the scale economies that made the original concentration efficient. That is a permanent cost increase borne across the system, not a transfer from one party to another.

What 2024 established for Asia-Pacific

• The region's AI exposure ran through the constraint, not the thesis, which is a stronger position for as long as the constraint holds.
• Concentration confers pricing power and attracts policy intervention — the same cause producing both, so they cannot be assessed separately.
• Power became the binding constraint on data centre development, creating opportunity in generation, transmission and equipment.
• Restriction deepened and substitution accelerated, moving toward two markets with different standards and higher costs for both.
• Constraint positions are durable but not permanent, and both capital and architecture are working on this one.