Custom chips are rapidly reshaping the AI industry landscape, with the competition for computing power behind large models like GPT-5 and Gemini quietly shifting.
As the global AI race reaches a fever pitch in 2025, a transformation in computing architecture is quietly taking place. While NVIDIA GPUs continue to dominate in training, ASIC (Application-Specific Integrated Circuit) chips are rapidly rising in the inference market, forming a new paradigm of “GPUs for training, ASICs for inference.”
Industry Status: A Game of Giants and Billion-Dollar Orders
The most notable recent development in the ASIC field is Broadcom securing a $10 billion ASIC order. Broadcom CEO Hock Tan revealed that this “unnamed new hyperscale customer” was later reported by the media to be OpenAI.
Broadcom’s Q3 fiscal results for 2025 show that its AI semiconductor revenue reached $5.2 billion, a 63% year-over-year increase. The company expects its AI chip business revenue to reach $6.2 billion in Q4, an increase of 66% year-over-year. Broadcom predicts that the global AI ASIC market will reach $60-90 billion by fiscal year 2027.
Not only Broadcom, but tech giants like Google, Amazon, and Meta are also heavily investing in self-developed ASIC chips. Google’s next-generation TPU Ironwood achieves an FP8 computing power of 4614T, with an HBM3e capacity of 192GB, and overall performance approaching that of NVIDIA’s B200.
Technical Advantages: Why ASICs Are the New Favorite?
ASICs have significant advantages over GPUs in inference scenarios:
1. Improved Energy Efficiency: ASICs are deeply optimized for Transformer models and fixed AI models, eliminating redundant general-purpose computing modules found in GPUs, resulting in higher energy efficiency.
2. Cost Advantages: Once AI algorithms converge and models stabilize, ASICs can be deployed at scale, effectively spreading R&D costs, with the total cost of ownership (TCO) per unit of computing power far superior to that of GPUs.
3. Specialized Performance: Designed specifically for certain algorithms, ASICs perform exceptionally well in inference tasks.
Market Outlook: Shipments Expected to Surpass GPUs by 2026
In terms of shipment volume, Google’s TPU shipments are expected to reach 1.5-2 million units in 2025, while Amazon’s AWS T2 is expected to reach 1.4-1.5 million units, and NVIDIA’s AIGPU supply will exceed 5-6 million units.
Currently, the total shipment volume of Google + AWS’s AI TPU/ASIC has reached 40-60% of NVIDIA’s AI GPU shipment volume. With Meta starting large-scale deployment of its self-developed ASIC solutions in 2026 and Microsoft beginning large-scale deployment in 2027, it is expected that the total shipment volume of ASICs will surpass that of NVIDIA GPUs at some point in 2026.
Core Concept Stocks in A-shares
ASIC Design/Services
Chipone Technology 688521 – One-stop chip customization business accounts for nearly 90%, with ASIC business revenue expected to reach approximately 725 million yuan in 2024, a year-over-year increase of 47.18%.
Cambricon 688256 – The Siyuan 590 is listed among the first tier of domestic AI chips.
Allwinner Technology 688220 – Sufficient orders in ASIC business, having undertaken multiple projects for leading clients.
CanSemi 688691 – Semiconductor IP licensing and chip customization services.
PCBs and Materials
Shenghong Technology 300476 – Core supplier of AI server PCBs, product upgrades enhance single-machine value.
Huadian Technology 002463 – High-end PCB supplier benefiting from server upgrade demand.
Shennan Circuit 002916 – Leading company in the PCB industry with strong technical capabilities.
Optical Modules
Inspur 300308 – Global leader in optical modules, with multidimensional deep cooperation with Broadcom.
Newray 300502 – High-speed optical module supplier with product compatibility with Broadcom.
Liquid Cooling Technology
Invec 002837 – Leader in liquid cooling technology, benefiting from the heat dissipation demand of AI servers.
GaoLan Technology 300499 – Supplier of liquid cooling solutions with a comprehensive technical layout.
Core Focus Points:
1. Technical Strength: Companies with core IP and design capabilities are more competitive.
2. Customer Relationships: Companies with close cooperation with leading internet companies have more secure orders.
3. Mass Production Capability: Companies capable of achieving large-scale production and yield control have a competitive advantage.
The rise of ASICs is not to replace GPUs but is an inevitable result of the maturation of the AI industry. As AI applications transition from “training” to “inference” and from “general-purpose” to “specialized,” custom chips are opening up a new frontier.
Disclaimer: This article only summarizes publicly available information, and mentioning specific stocks does not constitute a recommendation or investment advice. The market has risks, and investment should be approached with caution.