AI Hardware Is Struggling to Restart Asia's Manufacturing Escalator

Penang, 1972: half a century of assembly

In 1972 Intel opened its first assembly plant outside the United States in Penang, Malaysia. The mechanism was: labour, volumes, licence. Anyone reading AI growth forecasts for East Asia should start there.

In 2026 the same mechanism is at work, with servers in place of the early chips. The context is different; the structure is identical.

The outcome of that precedent is well known: Malaysia became a global hub for assembly and testing, while chip design stayed in other hands. Decades of volume delivered jobs and factory-floor expertise. The intellectual property stayed where it was born.

$1.4 trillion in exports, and what is inside them

The World Bank titled its October update on East Asia and the Pacific "Riding the AI Wave". According to Mia Mikic's analysis for the Lowy Institute, published on 9 October 2026^[1], China, Indonesia, Malaysia, the Philippines, Thailand and Vietnam exported $1.4 trillion of AI-related goods in the twelve months to April.

The same source reports that the upgrade to the region's 2026 growth estimate rests largely on chips, servers and routers.

The composition matters more than the total.

• Thailand: assembled computers, servers and routers are 69% of AI-related exports.
• Vietnam: the same category accounts for 60%.
• China: $785 billion of AI exports, largely intermediate goods shipped to ASEAN countries for assembly.
• Vietnam and Malaysia: AI exports at least double the 2020-22 average.

Design, software, intellectual property and data, the things that give these goods their value, originate and are captured largely elsewhere, Mikic writes. It is the old lesson of value chains: which task a country performs matters more than which product it ships.

One buyer and one investment cycle

The demand has a specific address. In early 2026, US imports of AI-enabling goods rose by almost 60%, while other imports fell by almost a quarter, the Lowy Institute reports.

I read these two numbers together. The region is growing because one customer is buying one category of goods, while the rest of its shopping basket shrinks. The engine is US AI capex: an investment cycle lasts as long as the credit that funds it.

The cyclical precedent is 2001: when US technology investment stalled, electronics exports from Singapore, Taiwan and Malaysia fell in the same year. Demand is cyclical, running three to seven years. Position in the chain is structural and lasts decades.

Manufacturing was the escalator that carried workers towards more productive jobs. The new cycle absorbs capital and creates few positions: in Malaysia a data centre requires 4.6 million ringgit of investment for every job created, according to the same analysis.

The bill goes beyond employment. The Bank for International Settlements devoted BIS Paper 174^[2] to the balance between AI innovation and climate sustainability. In my judgement that cost falls on the country hosting the facility, rarely on whoever collects its margins.

A Chinese data point shows where the rent goes. Analysts cited by the South China Morning Post^[3] expect a surge of more than 50% in Alibaba's AI cloud revenue, the payoff from its investment blitz. The margin rises to those who sell compute and software; those who build the racks are paid the price of a service.

The rules are written far from the factories. At the UN General Assembly in September, Brazil and France voiced concern about power concentrated in a few companies and a few states. Developing countries asked for compute, skills, infrastructure and a voice in standards, while the United States treated AI primacy as a strategic imperative.

Asking for a voice in standards is a trade policy demand, Mikic observes.

Three consequences for capital allocators

1. 12-18 month horizon. Growth in Vietnam, Malaysia and Thailand should be treated as exposure to US AI capex. Anyone already holding the big server buyers is doubling up on the same risk, and risk models deserve a dedicated scenario.
2. 36 month horizon. Value accrues where design, software and data reside. I choose countries and companies for their position in the chain before the volume they ship.
3. 5-10 year horizon. Standards, access to compute and data rules will decide who moves up the ranks. A country outside that table remains a capacity supplier, with margins set by the customer.

My thesis, and the data point that would overturn it

I argue that the market is mistaking dependence on a single investment cycle for structural convergence. The upgraded estimates measure volumes; the escalator measured wages and tasks.

I would change my mind in the face of one specific fact: the assembly share falling while total exports hold up. That would mean design, advanced components or software are entering the export basket of Vietnam and Thailand.

So far the available data is a snapshot: 60% and 69%.

What I expect by October 2027

By 31 October 2027 the share of assembled computers, servers and routers in Vietnam's AI exports will remain at or above 60%.

Confidence: Medium (65 out of 100). Horizon: 31 October 2027. Verification: the World Bank's October 2027 update on East Asia and the Pacific, or Vietnamese customs data.

The signal that proves me wrong: a published share below 60%.

The indicators to watch

• US imports of AI-enabling goods: a slowdown from the pace of almost 60% will show up quickly in Asian ports.
• The assembly share in the AI exports of Thailand (69%) and Vietnam (60%).
• The outcome of UN negotiations on standards and access to compute.

Volumes make growth. Rules make wealth.