The phrase "custom chip" has become ubiquitous in the technology industry, but its meaning has grown dangerously vague. In recent weeks, a wave of announcements from OpenAI, Anthropic, Meta, DeepSeek, and Zhipu has signaled that the race to build bespoke silicon is no longer confined to cloud hyperscalers. Model makers are now entering the design room, and the implications for Asia's semiconductor ecosystem are profound.
To understand why, one must first untangle the terminology. "Custom chip," "ASIC," "inference accelerator," and "custom silicon" are often used interchangeably, yet they describe at least five distinct categories: data-center AI accelerators, edge AI processors, networking chips, memory controllers, and connectivity parts for consumer devices. A deal for a Wi-Fi chip in a smartphone is not the same as one for a frontier inference accelerator, even if both are labeled "custom silicon." The confusion matters because the drivers behind each category differ radically.
Why Inference Is Reshaping Silicon Design
The core reason for this pivot lies in the changing economics of AI computing. For years, the bottleneck was training—the math-heavy process of building models. But as generative AI has moved into widespread use, inference—the act of running a model to answer a query—has become the dominant computational load. A single query requires pulling the entire model, often hundreds of gigabytes, across a narrow memory path, then performing only a small amount of arithmetic. The result is that the processing units sit idle while the memory interface is overwhelmed.
This structural shift means that faster processors no longer solve the problem. Instead, the solution lies in rearranging how data moves and is stored. That rearrangement is only possible if the chip designer knows the exact shape of the model it will run. Model makers, therefore, have a unique advantage: they can tailor silicon to their own architectures, reducing waste and improving efficiency.
Two areas are seeing the most innovation. First, logic chips are being redesigned to keep more data on-chip, reducing the need to fetch it from external memory. Second, memory itself is evolving to perform light computation, sending only what is needed and managing traffic with model-specific tools. These changes are not incremental; they represent a fundamental rethinking of the processor-memory relationship.
Asia's Role in the Custom Chip Wave
Asian companies are central to this story. DeepSeek and Zhipu, both Chinese AI labs, have announced custom chip efforts, joining a list that includes Baidu, Alibaba, and Huawei. In Japan, SoftBank's Arm architecture is increasingly used in custom designs, while South Korea's Samsung and SK Hynix are racing to supply the advanced memory that these chips require. India's semiconductor ambitions, though nascent, are also gaining traction as the government pushes for domestic fabrication.
The trend also has geopolitical dimensions. As the United States tightens export controls on advanced chips to China, Chinese firms are doubling down on domestic design capabilities. This is not just a matter of national pride; it is a strategic necessity. The widening AI access divide means that companies without their own silicon risk being locked out of the most efficient inference hardware.
Yet the barriers to entry remain high. Designing a competitive custom chip takes years and billions of dollars. Even for model makers, the path from announcement to production is fraught with technical and financial hurdles. The hyperscalers—Amazon, Google, Microsoft—have been at this for nearly a decade and are only now seeing results. For newer entrants, the timeline is even longer.
Still, the direction is clear. As models grow larger and inference demand explodes, the reward for custom silicon will only increase. The question is not whether more companies will design their own chips, but which ones will succeed—and what that means for the global semiconductor order.
