As AI applications continue to mature, particularly with the rapid emergence of new scenarios like AI agents, multimodal interaction, long-context processing, and tool calling, inference chains are significantly extending. Model call frequency and token consumption are persistently rising, shifting industry focus from peak performance to per-token cost, system utilization, and total cost of ownership. The inference era is creating a clear window of opportunity for ASIC adoption. The main viewpoints of Northeast Securities are as follows.
The marginal returns from stacking GPUs are diminishing, while the value of specialized computing power is becoming increasingly prominent. Capital expenditures from leading cloud service providers remain high, but constraints related to power, cooling, physical space, and operational complexity for AI clusters are intensifying. Compared to GPUs, ASICs can be deeply customized for relatively predictable inference workloads, optimizing computing units, on-chip memory, data flow organization, and chip interconnects. This enables superior energy efficiency, lower latency, and reduced unit costs in target scenarios. Cloud-side deployment is expected to be the primary growth driver in the near to medium term, while the proliferation of on-device AI holds potential for further long-term expansion.
Platform capability and turnkey delivery are critical for ASIC success. AI ASIC projects are no longer just about developing a single chip but involve system-level engineering encompassing advanced process nodes, HBM, chiplet technology, advanced packaging, software stacks, and mass production. The truly scarce capabilities extend beyond front-end and back-end design to include first-pass success rate, mass production execution, key IP integration, and collaboration with foundry and packaging partners. From a global perspective, Broadcom exemplifies a platform-based approach combining "custom chips + high-performance interconnects," with value extending from the ASIC itself to AI networking and infrastructure platforms. Alchip represents a "light-IP, heavy-delivery" turnkey model, leveraging advanced nodes, advanced packaging, and production capabilities to engage in highly complex AI projects. These two models outline the primary growth paradigms for ASIC design companies.
Domestic manufacturers are facing a strategic opportunity, where their platform capabilities will determine their depth of engagement in AI ASICs. VeriSilicon Holdings Co., Ltd., with its proprietary IP platform and SiPaaS model, most closely resembles the platform-based AI ASIC service provider path. ASR Microelectronics Co., Ltd. is leveraging its experience in baseband SoCs and large-scale chip production to extend its product validation capabilities into ASICs. Brite Semiconductor follows a pure turnkey model, deeply integrating with the local manufacturing ecosystem. NationalChip focuses on the high-reliability, secure, and controllable sector, addressing specialized ASIC needs for major national requirements. Domestic ASIC design service providers are not simply interchangeable but correspond to different business models and capability boundaries, all poised to benefit from the accelerating trend of AI localization in China.
Risks include potential shortfalls in AI capital expenditures, slower-than-expected adoption of ASICs by cloud service providers, weaker-than-anticipated orders for domestic companies, and intensifying industry competition.
Comments