Morgan Stanley's latest semiconductor industry chain survey report highlights that China's AI chip sector is undergoing significant structural adjustments, with inference computing demand gradually replacing pure computing power stacking as the core driver of market growth.
According to market intelligence, Morgan Stanley's on-the-ground research in Asian markets on the 8th revealed new divergences in China's AI accelerator market. Some domestic AI chip design companies are proactively adjusting their technical roadmaps, focusing on inference applications by developing mid-range specification products.
This shift has led to a more pragmatic approach in technical specifications. The report indicates that supply chain feedback shows local chip designers, including Tianshu Zhixin (燧原科技), are transitioning to TSMC's 6nm or 7nm processes for inference chip designs. Meanwhile, to accommodate the energy efficiency and cost-sensitive nature of inference tasks, these chip solutions are increasingly adopting LPDDR memory instead of supply-constrained and expensive HBM, carving out new growth opportunities in specific application scenarios.
On the foundry side, Morgan Stanley maintains an optimistic outlook for TSMC while favoring SMIC among domestic foundries. Analysts believe robust global AI demand will continue to support TSMC's compound annual growth rate exceeding 40% over the next five years.
**Rise of Inference Chips: Mid-Range Approach Becomes Mainstream** Morgan Stanley's supply chain survey reveals that Chinese AI chip companies are adopting more flexible strategies to adapt to the rapid expansion of inference scenarios.
The research notes that China's AI computing focus is accelerating its shift from training to inference. Currently, available computing power mainly comes from NVIDIA's 5090-level consumer GPUs, early Hopper architecture products, and some domestic AI accelerator chips. In this landscape, local players like Tianshu Zhixin are accelerating the iteration of inference-specific chips using more mature 6nm and 7nm nodes.
Memory supply chain data further confirms the rise of inference demand. Morgan Stanley points out that a new generation of domestic AI inference chips is increasingly adopting LPDDR as the primary memory solution. This approach offers significant advantages in cost, power consumption, and deployment flexibility, better aligning with the high concurrency, low latency, and low-power requirements of inference applications.
**Reshaping Market Supply-Demand and Localization Window** Regarding external supply, Morgan Stanley observes a tiered pattern in China's AI computing demand, characterized by "tight training capacity but expanding inference opportunities." The report suggests that certain new AI accelerator categories have failed to gain strong traction in China due to pricing, energy efficiency, or ecosystem compatibility issues, creating a window for domestic solutions.
The survey indicates that China's current AI inference computing power is primarily supported by consumer GPUs, early Hopper chips, and domestic alternatives. This dynamic provides local chip designers with greater iteration space, encouraging them to focus on breakthroughs in "inference performance" through differentiated approaches.
In the domestic foundry sector, Morgan Stanley updated its scenario analysis. The report suggests that if Chinese cloud service providers (CSPs) increase the adoption of domestic designs in specific scenarios or Taiwan-process-based domestic chips accelerate production, some existing demand could be redistributed in the short term.
Despite this potential diversion effect, Morgan Stanley maintains an "Overweight" rating on SMIC, viewing its prospects more favorably than Hua Hong Semiconductor. Analysts emphasize that SMIC's N+2 (7nm) node could become a critical technology point for domestic AI chip production in 2025, while the N+3 (5nm) node in 2026 will further enhance local supply capabilities.
**Market Outlook: Self-Sufficiency Rate and Revenue Forecast Revised Upward** Based on expectations of strong inference demand growth, Morgan Stanley has raised its revenue forecasts for China's AI GPU market. Under the baseline scenario, analysts have increased 2026 and 2027 revenue projections from RMB 94 billion and RMB 136 billion to RMB 113 billion and RMB 180 billion, respectively.
Data models suggest that as domestic manufacturers fill supply gaps, China's AI GPU self-sufficiency rate could reach 50% by 2027. Despite complex external challenges, the structural resilience of domestic AI computing demand remains intact, with inference applications poised to be the biggest winner supporting market growth in the coming years.
Comments