The AI build-out has become a macroeconomic event in Asia. South Korean exports reached $98.89 billion in July, up 62.8% on a year earlier, with semiconductor exports up 178.8% to $41.01 billion. Taiwan is on course for its first year of double-digit growth since 2010. Japan, Malaysia, Singapore and China all posted export growth above 20% in the same month.
Equity markets have followed. Korea's Kospi is up close to 60% this year; Japan's Nikkei 225 and Thailand's SET are each up roughly 25%.
Those are extraordinary numbers, and they are also narrow ones.
What the winners are selling
Korea's gain is memory. The high-bandwidth memory that sits beside an AI accelerator is a product only a handful of firms can make, and Samsung and SK Hynix make most of it. Boursel has reported this week that Samsung's quarterly operating profit reached a record and that both companies are returning tens of billions to shareholders.
Taiwan's gain is the accelerator itself, through TSMC, and the ecosystem of packaging and testing around it. Japan's is further upstream still, in the equipment, photoresists and photomasks without which nobody makes an advanced chip anywhere.
What these have in common is that they are difficult to replicate. Each rests on decades of accumulated process knowledge, and that is why the margin sits there.
What Southeast Asia is getting
Singapore, Malaysia, Thailand and Vietnam are also in the boom, but at a different point in the chain: data-center construction, power and land, and assembly and test rather than fabrication. Singapore has raised its growth forecast to between 4.5% and 5.5%.
Danny Quah of the Lee Kuan Yew School in Singapore put the problem precisely, describing the gain as a sugar rush that comes from supplying the supporting rather than the leading-edge parts, inputs he called commodifiable.
Commodifiable is the right word. Anything that can be sourced from several countries earns the price a competitive market allows, and the countries currently supplying it are competing with each other to do so.
Why a data center is a weak growth engine
The construction phase employs a lot of people for eighteen months to two years, and creates real demand for concrete, steel, electrical work and transformers.
The operating phase employs very few. A large facility runs with a modest technical staff, because the point of the design is that it needs almost no one. The tax base is real and the electricity bill is enormous, but the jobs are not there, and neither is the technology transfer: the valuable engineering happened somewhere else and arrives on a truck.
The constraints are also binding sooner than expected. Grid reliability and water availability are already limiting where these facilities can go, and in several of these countries the electricity being committed to data centers is electricity not available to domestic industry.
What would change it
Moving up the chain requires the thing the region has found hardest: engineers, and the reason for them to stay. Skilled workers leaving for Singapore and the West is a long-running pattern, and populations in several of these economies are ageing before the incomes have caught up.
The honest summary is that this boom is doing for Korea, Taiwan and Japan roughly what electronics did for them in earlier cycles, and doing for their southern neighbours what assembly work usually does: raising output while it lasts, without changing what the economy is able to make.
Whether that is a blip depends on decisions about education, power generation and industrial policy that are being taken now, while the construction money is still arriving.
