On Monday, the market experienced a fluctuating recovery, with the Shenzhen Component Index rising over 1% and the ChiNext Index gaining more than 2%. The total trading volume for the Shanghai and Shenzhen markets was 2.77 trillion yuan. In terms of sectors, concepts such as CPO, optical fiber, PCB, MLCC, and robotics performed actively. On the downside, sports-related concepts saw collective adjustments. By the close, the Shanghai Composite Index was up 0.43%, the Shenzhen Component Index increased by 1.63%, and the ChiNext Index surged 2.66%.
In today's morning brokerage briefings, CICC suggested that Tesla could propel the smart driving industry chain into a phase of rapid volume expansion. China Securities highlighted robotics as a key direction for AI applications, advising a focus on high-quality segments amid market volatility. Zhongtai Securities posited that inference is the future core of AI computing power, forecasting a significant revaluation for CPUs.
Key Brokerage Perspectives
Tesla Motors (ASX: TSLA)
CICC's analysis indicates that Tesla is poised to drive the smart driving supply chain into a high-growth cycle. The firm notes that Tesla, in its Q1 2026 earnings guidance, targeted securing approval for its Full Self-Driving (FSD) system in China by Q3 2026. Following this, on May 21, Tesla officially announced the inclusion of China in its global FSD rollout plan. Concurrently, the mass production of the Tesla Cybercab commenced in April 2026.
The accelerated global expansion of Tesla's FSD system is expected to hasten the worldwide adoption rate of intelligent driving technology. Furthermore, Tesla's advancements in high-level autonomous driving and Robotaxi services are creating a catalyst effect, potentially increasing consumer acceptance of smart driving. This dynamic is anticipated to push the entire industry chain into a period of substantial volume growth.
Focus on Robotics as an AI Frontier
China Securities emphasized that physical AI represents the next wave of artificial intelligence, with robotics serving as one of its most effective physical embodiments. The future is expected to see billions of autonomous systems and robots operating in the physical world.
Although the robotics sector has recently experienced some adjustments, this is primarily attributed to shifts in investor sentiment. The underlying narrative of physical AI, however, is a tangible and advancing industrial trend that warrants significant attention. Additionally, the mass production of Tesla's Optimus robot is drawing closer, with clearer guidance emerging on supply chain volume and specifications, validating its production ramp-up timeline. Upcoming events like the V3 product release and further production milestones remain crucial to monitor.
The ongoing progress of IPOs for domestic robotics companies could also lead to a revaluation of these entities. With continuous catalysts for the sector, the focus should remain on high-quality segments within the value chain.
The Central Role of Inference in Future AI Compute
Zhongtai Securities argues that the structure of computing power is transitioning from a training-centric model to one where inference is paramount. Currently, over 70% of computing power is dedicated to centralized training, but in the future, more than 70% is projected to be used for distributed inference. The scale of inference demand could reach 5 to 10 times that of the training phase.
The fundamental differences between training and inference determine the distinct roles of CPUs. In training scenarios, the CPU acts primarily in a "supporting" role, whereas in inference, it can become the "main force." This is based on two key points: First, according to Little's Law (Throughput = Concurrency / Latency), CPUs and GPUs have chosen different optimization paths—CPUs focus on reducing latency, while GPUs aim to increase concurrency. Second, training predominantly involves large-scale dense matrix operations, where GPUs handle most of the computation, and CPUs are responsible for data transfer and cluster scheduling, accounting for only 10-30% of the time. In contrast, inference tasks are characterized by fragmentation, long-tail distribution, and latency sensitivity. CPUs demonstrate greater competitiveness in various tasks such as decoding, sparse computation, long-context management, and embedding, potentially handling over 70% of the runtime workload.
Comments