AI Economics · · 5 minute read
Asia makes the chips, pours the concrete and writes the code for the AI boom. The profits sit with the first two. Most of the region's businesses are still on the paying side.
When people talk about the business of AI in Asia, they usually mean the models: which lab is ahead, whose open weights are cheapest, whether the region can build its own. That is the loudest part of the business, and the part where the least money changes hands.
Follow the cash instead and the picture is simple. Asia earns most at the bottom of the stack, where the chips are fabricated and the data centres are built. It earns less, and increasingly under pressure, in the middle, where services firms sell implementation by the hour. And at the top, where ordinary companies actually use the tools, most of the region is still spending ahead of the returns.
None of these layers is a secret. Lined up side by side, they tell an operator where the margin sits and where it does not.
TSMC reported second-quarter revenue of $40.20 billion, up 36.0 per cent on a year earlier, with a gross margin of 67.7 per cent and a net margin of 55.6 per cent. It guided the third quarter to between $44.6 billion and $45.8 billion (TSMC, 16 July 2026). August revenue alone was NT$514.81 billion, up 53.3 per cent year on year, taking the first eight months to NT$3,386.87 billion, up 39.3 per cent (TSMC monthly revenue report, 10 September 2026).
A manufacturer keeping more than half its revenue as profit has pricing power. Korea's memory makers are in a similar position, which I covered last week. Whatever happens to any individual model company, every one of them pays this layer first.
So what: The most profitable AI business in Asia is not an AI company in the sense most boards mean. If your strategy deck treats "Asian AI" as the model labs, it is looking at the wrong end of the value chain.
The second layer is power, land and buildings. Knight Frank's Data Centre Atlas 2026 puts Johor's live IT capacity at 1,110 megawatts, just behind Singapore's 1,118 and Tokyo's 1,473, with 8,542 megawatts in the pipeline and a co-location vacancy rate of 0.7 per cent. Singapore's vacancy is 4.9 per cent; Bangkok's is 23.3 per cent and Jakarta's 20.5 per cent (The Edge Malaysia, on Knight Frank's report of 30 July 2026).
A vacancy rate under 1 per cent is a landlord's market. It also shows how uneven the build-out is. Johor is full because it sits next to Singapore's power and land constraints; other Southeast Asian capitals have built ahead of demand.
So what: Data centres are where Southeast Asia captures real AI value, but it is property and utility economics, not software economics. Governments and developers outside Johor should read those vacancy numbers before approving the next campus.
India's IT services firms sit between the infrastructure and the end user, and AI is squeezing them from both sides. In April, HCLTech's chief executive C. Vijayakumar said "AI deflation" would cut 3 to 5 per cent from future revenue in the coming year, and TCS's K. Krithivasan acknowledged the same effect (The Register, 28 April 2026). Infosys then reported first-quarter revenue of $5,082 million, up 2.4 per cent in constant currency, said AI revenues were 8.2 per cent of the total, and narrowed its full-year growth guidance to 1.5 to 3.0 per cent (Infosys, 23 July 2026).
The new AI revenue is real. It is just smaller than the old work it is replacing, because a task that used to be billed as ten people for a month can now be done by two and a set of agents.
So what: If you buy IT services in Asia, you have leverage you did not have two years ago. Ask suppliers to price by outcome, not headcount, and expect them to agree. If you sell services, the hourly model is the part of your business AI is eating first.
At the top of the stack sit ordinary companies, and they are spending hard. Digital Realty's 2026 survey of 707 IT decision-makers across seven Asia Pacific markets found 59 per cent expect to raise AI investment by more than 25 per cent this year. Only 16 per cent said they were realising a return on it. Singapore did better, with 39 per cent reporting measurable business outcomes, and just 36 per cent of regional respondents had found repeatable, high-impact use cases (Digital Realty, 22 September 2026).
This is the part of the business I see from the inside. The companies that get a return are rarely the ones with the biggest budgets. They are the ones where the people who know the work were taught to redesign it, rather than handed a licence and left alone.
So what: Spending is not adoption. If more than four in five of your peers have not seen a return yet, the gap is not the tools, which are cheap and getting cheaper to access. It is whether your domain experts can actually use them.
The business of AI in Asia today is mostly a business of selling to the AI boom, not of profiting from using it. The fabs and the landlords are paid up front. The services firms are watching their unit of work shrink. The buyers are paying all of them and waiting. That will not last forever, and the operators who come out ahead will be the ones who stop asking which model to buy and start asking which of their own processes, run by their own people, can turn this spend into margin.
Keep reading
AI Economics
For three years operators budgeted AI as something that only gets cheaper. The first week of October says otherwise, and this region is both selling the scarce inputs and paying the higher price.
Model Economics
Asian buyers talk about the non-American models as one cheap bloc. The money says otherwise. One sells tokens, one sells apps, one sells a position ahead of a listing.
Enterprise Adoption
Adoption is up almost everywhere in Asia. Budgets are up too. Look at what the money is attached to and a quieter story appears: most organisations are still buying tools, and very few are buying redesigned work.