AI ASICs offer significant advantages in energy efficiency and cost for inference tasks, which is expected to accelerate their adoption among major internet companies. The report indicates that other leading domestic chip manufacturers will coexist with chips developed in-house by internet giants, benefiting top-tier chip firms with proprietary architectures and ecosystem synergy, as well as the upstream and downstream supply chain for internet AI ASICs. Based on this, a "recommended" rating is maintained for the computer industry.

AI ASIC Overview: Clear Advantages in Inference, Long-Term Coexistence with GPUs

AI ASICs are chips custom-designed for neural network training and inference. By removing general-purpose redundant modules and incorporating systolic arrays or dedicated tensor units, they achieve higher computational density, better energy efficiency, and lower cost. These chips come in architectures like TPU, NPU, and LPU. According to Semiconductor Industry Perspective, over 90% of AI frameworks natively support NVIDIA's CUDA for training, but inference scenarios have relatively fixed model algorithms, requiring less flexibility and prioritizing energy efficiency and cost, where AI ASICs excel. Global cloud giants and model developers are exploring AI ASIC solutions. DIGITIMES Research projects global high-end AI ASIC shipments to reach 7.23 million units in 2026, with a year-over-year growth rate exceeding 40%, and Counterpoint Research expects shipments to triple by 2027 compared to 2024. In the supply chain, AI ASIC cooperation often involves joint development, where major companies define chip specifications and architecture, while front-end physical design is completed with specialized ASIC design firms like Broadcom, Marvell, Alchip, and MediaTek. Broadcom and Marvell hold key positions due to their critical IP and system solutions. AI ASICs and GPUs are not substitutes but represent a trade-off between "efficiency" and "flexibility," with multi-chip strategies becoming standard. TrendForce estimates that AI ASIC server shipments will account for 27.8% of the total AI server market in 2026, rising to 39.5% by 2030.

Overseas AI ASICs: North American CSPs Iterate In-House Chips, Now in Mass Deployment

Google's TPU is currently on its seventh generation, Ironwood (TPU v7), which delivers 4.6 PFLOPS of FP8 dense compute and includes 192 GB of HBM3E memory. The eighth-generation TPU (split into 8t and 8i versions) was announced in April 2026, with mass production expected by the end of 2027. DIGITIMES Research forecasts Google TPU shipments of 3.326 million units in 2026, representing 46% of the entire AI ASIC market. According to Wall Street News, AWS's Trainium 3 has entered the ramp-up phase, offering 2.52 PFLOPS of FP8 compute per chip, with volume shipments expected to begin in Q3 2026, totaling around 1.47 million units for the year. The Trainium 4 is slated for a December 2026 launch, with single-chip FP8 dense compute boosted to 7.6 PFLOPS. Meta has unveiled a four-generation chip roadmap (MTIA 300, 400, 450, 500) covering two years, with FP8 compute ranging from 1.2 PFLOPS to 10 PFLOPS. The MTIA 450 and 500 are planned for large-scale deployment in 2027. Microsoft's Maia 200, released in February 2026, is optimized specifically for inference, delivering 10.15 PFLOPS of FP4 compute and 5.07 PFLOPS of FP8 compute. It is used in Azure data centers to support applications like OpenAI's GPT-5.2.

Domestic AI ASICs: Chinese CSPs Have All Entered the Field, Targeting Partial AI Cloud Computing Needs

ByteDance's rationale for developing its own chips is to reduce computing costs through customization. It has laid out four product lines: AI chips, CPUs, VPUs, and DPUs. Its AI inference chip, SeedChip, uses an NPU architecture and is expected to produce 100,000 units in 2026, with a long-term production capacity plan to ramp up to 350,000 units. This chip will primarily serve its own business scenarios like Douyin and Doubao. Simultaneously, it is developing a new generation of inference chips inspired by Groq's LPU, incorporating RRAM resistive storage technology to bypass HBM supply constraints. Alibaba's T-Head covers AI chips, server CPUs, and other product lines. The Zhenwu 810E (PPU) follows a GPGPU route, with comprehensive performance comparable to NVIDIA's H20, and has been deployed in multiple 10,000-card clusters within Alibaba Cloud. In May 2026, the new generation Zhenwu M890 training-inference integrated chip was released, offering approximately three times the performance improvement over the 810E. It also launched a 128-card Panjiu super-node solution. As of May 2026, cumulative AI chip shipments have reached 560,000 units, serving hundreds of customers across over 20 industries. Baidu's Kunlun Chip has completed three chip generations based on its proprietary XPU architecture. Its main product, the P800, uses a 7nm process and has been delivered to multiple 10,000-card clusters. The M100 was released in July 2026, targeting large-scale cloud inference and comparable to NVIDIA's H20, with an official market launch planned for 2026. The M300 is positioned as a dedicated chip for ultra-large multimodal model training and is expected to hit the market in early 2027. Kunlun Chip has initiated its listing guidance on the STAR Market and is simultaneously advancing an "A+H" dual listing plan. Tencent has laid out its AI ASIC strategy across three tracks: inference, video transcoding, and networking. Its Zixiao series of AI inference chips is co-developed with Enflame. The Zixiao 1.0 has been mass-produced and deployed in several key businesses on Tencent Cloud. In July 2026, the Zixiao 2.0 was released, targeting large model training and inference, with a 4x improvement in training performance and a 3x boost in inference throughput. Tencent plans to increase the proportion of self-developed Zixiao chips to 30% in 2026, expanding its deployment scale to 100,000 cards by year-end. Additionally, Tencent supplements its computing power by directly purchasing from Enflame, with procurement reaching 830 million RMB in 2025, accounting for over 80% of Enflame's revenue.

Risk Factors: Downstream demand recovery may fall short of expectations; AI large model development may underperform; risks of raw material price fluctuations; intensified market competition; exchange rate fluctuation risks; key company performance may not meet expectations; products are not fully comparable; and related data and benchmarks are for reference only.