T.Media logo

Custom AI chips bring Taiwan more routes to growth

Reporter Richard Brown
Release time:2026/11/17 16:18
Last update time:2026/11/17 16:18
  • S

  • M

  • L

MediaTek is applying its chip design and high-speed interconnect expertise to data center ASICs. (Image: Richard Brown)
MediaTek is applying its chip design and high-speed interconnect expertise to data center ASICs. (Image: Richard Brown)

Taiwan’s AI infrastructure business is growing as the world’s largest cloud companies design more of their own processors. Google’s tensor processing units and Amazon’s Trainium chips are creating orders for custom chip design, advanced manufacturing, servers, and cooling, adding to the demand already generated by Nvidia systems. For Taiwan, the shift broadens the customers and technologies supporting its growth.

The expansion is already visible in suppliers’ forecasts. MediaTek expects its AI accelerator ASIC business to generate US$2.5 billion in fourth-quarter revenue, following initial shipments in September. Cooling supplier Asia Vital Components, or AVC, expects revenue associated with ASIC systems to exceed its GPU-related business in 2027 as cloud operators deploy more internally designed processors.

ASIC stands for application-specific integrated circuit, a type of chip that has long performed specialized functions in electronics. In today’s AI market, the term generally describes custom accelerators designed around one company’s computing requirements, including the calculations, memory access, and connections between processors needed to run its models.

A GPU, or graphics processing unit, developed from hardware for rendering images into a programmable processor capable of performing many calculations simultaneously. That makes it useful across AI, scientific computing, and other demanding applications, and Nvidia’s CUDA software platform gives developers an established environment for programming it.

Both kinds of chip are programmable. Modern AI ASICs can support different models, and Google’s TPUs handle training, fine-tuning, and inference, the process of using a trained model to produce an answer. What distinguishes them is range, since their hardware is optimized for a narrower range of computational requirements than a broadly deployed GPU platform.

For cloud operators, the attraction is the cost of delivering AI at scale. They can design processors around the workloads they expect to run most often, allocating computing resources, memory, and data connections accordingly. Improvements in energy efficiency or hardware utilization can produce substantial savings across thousands of servers.

Growing inference demand strengthens that incentive. Every interaction with an AI service consumes computing resources, so the cost of repeatedly generating answers becomes a major operating expense as usage increases. Amazon says Trainium3 offers 30–40% better price-performance than Trainium2 and reported earlier this year that its capacity was nearly fully subscribed. The comparison is within Amazon’s own product family, but it illustrates the commercial goal of each custom-chip generation.

Owning the design also gives cloud companies greater influence over product schedules and system architecture. Developing a competitive accelerator nevertheless requires substantial engineering investment, software support, and sufficient deployment volume to recover those costs. GPUs retain advantages where developers need broad software compatibility or flexibility as models change. Custom processors therefore add another major source of AI computing capacity rather than making GPUs unnecessary.

Taiwan’s chip design companies help turn those custom architectures into manufacturable products. Alchip and Global Unichip, or GUC, provide design implementation, reusable circuit technology, packaging expertise, and production support. Their work connects customers’ computing requirements with the physical constraints of advanced semiconductor manufacturing.

Alchip is developing 2nm ASICs and expanding its capabilities in chiplets and three-dimensional integration. GUC has introduced HBM4E memory interface technology for 2nm designs, helping customers connect processors to the high-bandwidth memory needed to keep them supplied with data. These capabilities become more valuable now that an accelerator consists of several interconnected pieces of silicon.

MediaTek adds a much larger competitor to the field. Best known for smartphone processors, it is applying its chip design and high-speed interconnect expertise to data center ASICs. Its first accelerator project is entering mass production, while a second is scheduled for high-volume production in early 2028. The company is also investing in advanced packaging and optical connections as its customers’ requirements extend beyond the processor itself.

TSMC supplies the manufacturing and advanced packaging on which many of these designs depend. Its CoWoS technology connects processing chips and high-bandwidth memory within a package, supporting both GPUs and custom accelerators. Taiwan’s packaging and testing companies, including ASE, assemble and validate these increasingly complex devices. A customer choosing its own processor still remains heavily dependent on Taiwan’s semiconductor industry.

The benefits continue into servers. Wiwynn, controlled by Wistron, has seen rising ASIC-server shipments alongside demand for general-purpose systems. Its third-quarter revenue reached NT$434.62 billion, approximately US$13.25 billion, up 62.9% from a year earlier. Quanta is also developing its ASIC-server business, while Foxconn says its work with major US cloud customers extends across components, modules, servers, and complete racks.

Each architecture requires boards, power supplies, cooling, and connections adapted to its design. AVC plans to expand liquid-cooling capacity by more than 50% in 2027. Lite-On identifies volume production of 8.5-kilowatt power supplies for ASIC applications as a fourth-quarter growth driver. Custom processors may improve computing efficiency, but large installations still require substantial electrical and cooling infrastructure.

For Taiwan, custom chips bring two benefits. Each new program adds work in chip design, manufacturing, packaging, and server assembly, and a supplier serving several programs is less exposed to a delay or change in any one of them. Two constraints remain. Growth still depends on cloud operators’ spending and custom chips draws on the same limited supply of advanced manufacturing and memory as GPUs do. Even so, whichever processors those operators choose, Taiwan firms are positioned to help, build, and assemble them.

US$ conversion uses an illustrative exchange rate of NT$32.8 per US$1.