Asia’s AI infrastructure boom is fuelling semiconductor profits and investment, reshaping economies and creating new opportunities—and risks—across the APAC region

How is the AI infrastructure boom changing the competitive landscape for semiconductor leaders in Taiwan, South Korea, and Japan, and what strategies are companies adopting to sustain growth beyond the current surge in demand?

“Right now, the whole industry is building for peak training demand. That is the risky bet. The surge is training-led, but the floor underneath it is inference, and inference has a very different shape.

“It helps to think of the chip makers as running different races. Nvidia, AMD and Huawei are not playing the same game, and the buildout in Taiwan, Korea and Japan is not one market either. The most durable position may not be logic at all. It is memory. High-bandwidth memory is already the chokepoint, and as workloads shift from training a model once to running it a billion times, memory bandwidth and cost gate everything. That favours the Korean and Japanese supply chain more than the headlines suggest.

“The strategy that survives the surge is not defending one architecture. It is competing on cost per useful token, and making sure your capacity is not built for a workload mix that is already changing.”

What new opportunities are emerging for businesses outside the semiconductor sector as AI infrastructure spending accelerates?

“The opportunity is in everything that sits around the chip. AI infrastructure spending creates demand for data centres, power, cooling, networking, construction, systems integration and the software layer needed to run models in production.

“The next issue for many companies will be inference cost. Training large models gets most of the attention, but the cost of running models repeatedly across real workflows is where a lot of enterprise spending will sit. That is where smaller, more specialised models become important.

“There is also a commercial opportunity for businesses that can make AI less dependent on one hardware route. Companies don’t want every use case to rely on the largest model running on the most expensive GPU available. In industries like manufacturing, logistics, financial services and healthcare, there is demand for models that can be run efficiently, closer to the data and with more control over cost.”

How are governments across Asia balancing industrial policy, national security concerns, and international partnerships to strengthen their positions in the global semiconductor value chain?

“The lesson a lot of governments are learning is that chips are necessary but not sufficient. Owning fabs does not mean you own your AI. Real sovereignty needs the full stack: open models you can inspect, inference you control, and hardware you can actually source. A country can have world-class fabs and still be dependent on someone else’s model and someone else’s software.

“Export controls are the other force worth watching. Restrictions meant to concentrate power are quietly doing the opposite. They are accelerating domestic silicon and alternative vendors like Huawei. The unintended result is a more multipolar hardware map, which is good news for anyone who refused to bet their entire strategy on one vendor.”

Could concentration of AI chip production create new geopolitical or supply-chain vulnerabilities, and how should businesses prepare?

“The AI market is becoming dependent on a small number of chip suppliers, foundries and regions. That makes supply-chain risk part of AI strategy, not just a hardware procurement issue. There is also a memory issue. High-bandwidth memory is scarce, and as inference demand grows, companies will not be able to assume that more compute will always be available at the right price, which changes how they should think about model design and use.

“Businesses should avoid building AI plans around a single model provider, chip type or cloud route. They need more flexibility across models and hardware, and they should look seriously at smaller models that use less compute and memory for specific tasks. That will not remove geopolitical risk, but it does reduce the extent to which a company is exposed to one part of the supply chain.”

As investment pours into AI data centres and advanced fabrication facilities, what pressures are being placed on energy supplies, water resources, and infrastructure, and how are companies addressing sustainability concerns?

“Efficiency is the sustainability story, and most people are looking at it from the wrong end. Running a frontier model for every single request is like lighting an entire building with incandescent bulbs and then arguing about how many power plants you need. The fix is not more power. It is the LED. A smaller, specialised model running the same job can cut the energy per useful task by an order of magnitude.

“So the sustainability question is not only how green the data centre is. It is whether you are using the right size model for the work at all. The brute-force path, bigger model on more GPUs for everything, is the least sustainable one available, and it is also the most expensive. Efficiency and sustainability point the same direction here.”

Are we seeing a long-term structural shift, or is there a danger of overcapacity and another boom-and-bust cycle?

“The demand for AI infrastructure is long-term, but that does not mean every investment being made now will be useful. A lot of spending is still based on the idea that bigger models and more hardware solve most problems, but, in reality, enterprise use is more mixed. Companies need document processing, search, software support, customer operations and internal knowledge tools. Many of those workloads do not need the largest model available.

“That is where the market could split. There may be overbuild in parts of the infrastructure market, while companies still struggle to run AI cheaply and reliably in day-to-day operations. The next phase will be focused on whether AI can be run efficiently at scale.”

Which sectors and markets across APAC stand to benefit most over the next five years?

“The first beneficiaries will be the sectors closest to the buildout: data centres, energy, grid infrastructure, cooling, telecoms, construction and advanced manufacturing. Beyond that, the opportunity is in sectors with large volumes of operational data and documents. Financial services, insurance, healthcare, logistics and the public sector all have use cases where AI can be applied without waiting for another breakthrough in model training.

“For APAC, the stronger position will sit with markets that can connect hardware capacity with deployment. Chips are only one part of the picture. Memory, power, networking, software infrastructure and inference efficiency will all matter. For investors and businesses, the better question is not only which chipmakers benefit from AI demand. It is which companies help make AI cheaper and easier to run once that hardware is in place.”

Eugene Cheah, CEO and Co-Founder of featherless.ai.

Eugene Cheah is CEO and co-founder of Featherless.ai, a San Francisco-based AI infrastructure company providing serverless access to open-source AI models. He previously co-founded UIlicious, where he was CTO, and has worked across software testing, web application development and machine learning infrastructure. Eugene is also co-lead of RWKV, an open-source AI model project under the Linux Foundation, and is focused on making AI models more accessible, affordable and independent of closed platforms.