AI and Chips: Specialization, Integrated Storage and Computing, and Edge Multimodality

Multiple discussions summarize the future scenarios and technological trends of the integration of AI and semiconductors. - Scenarios: Embodied intelligence (robots, drones, autonomous driving) requires low-power real-time visual and control chips; edge multimodality will run large models on smartphones, cars, and AR glasses; AI-native scientific computing will accelerate drug and material discovery using AI chips; edge intelligent networks support industrial IoT and autonomous smart factories. - Three major chip trends: differentiation from general-purpose GPUs to specialized chips; integrated storage and computing to break through the memory wall; collaborative design of chips and algorithms (hardware-software synergy). - Industry dynamics: procurement and negotiations by Nvidia/Google/Meta, differentiation and positioning of domestic manufacturers (e.g., Biren, Suirian, Huawei, Horizon). - Economics and timeline: If integrated storage and computing or embodied intelligence chips mature, a trillion-dollar market and a new generation of smart hardware will emerge within 3-5 years. - Behavioral observation: One of the obstacles to the promotion of autonomous driving is that driving itself is a behavior with positive feedback (driving is both a problem and a reward).
Key insights
Embodied Intelligence Will Drive Demand for Specialized Low-Power Real-Time Chips
Discussions point out that current AI primarily operates within screens, while embodied intelligence (robots, drones, autonomous driving) requires chips that can process vision, decision-making, and control in real-time with extremely low power consumption. Such scenarios have strict requirements for latency and energy efficiency, which general-purpose GPUs struggle to meet. If embodied intelligence chips mature in the next 3-5 years, they will give rise to trillion-dollar markets for AI caregivers, AI workers, and AI delivery personnel, and drive a shift towards specialized chip design.
Edge Multimodality Will Drive Chip Development for Smartphones, Automotive, and AR
In the future, large models will shift from the cloud to smartphones, cars, and AR glasses, requiring edge chips to handle text, images, voice, and video simultaneously at a few watts of power. Manufacturers like Qualcomm, MediaTek, and Huawei have already laid out plans for edge multimodal chips. If realized, this will lead to a new generation of smart hardware, similar to the impact of the iPhone on smartphones.
Integrated Storage and Computing Viewed as Key Technology to Solve Memory Wall and Energy Consumption Bottlenecks
Currently, the bottleneck for large models is data movement rather than pure computing power. Integrated storage and computing reduces data movement by closely coupling computation and storage. If technological breakthroughs occur and mass production is achieved, this will significantly lower inference costs and make many currently unfeasible applications possible. Research in this direction is being pursued by universities and startups both domestically and internationally, with the next 3-5 years expected to be critical.
Joint Optimization of Chips and Algorithms Becomes Core Competitive Advantage
Future chip design will involve collaborative optimization with algorithms—algorithms understanding chip characteristics and chips optimizing for algorithms. Examples include the integration of Apple's Neural Engine with iOS and Tesla's FSD chip tailored for autonomous driving algorithms. Domestic manufacturers like Huawei are attempting to integrate Ascend chips with large models and systems, potentially forming a core competitive advantage in hardware-software integration.
Industry Procurement and Capital Layout Indicate Short-Term Market Preference for Large Suppliers
At the market level, Anthropic has reached a large-scale TPU procurement agreement with Google, and Meta is negotiating a multi-billion dollar TPU procurement with Google. Short-term market sentiment shows a preference for large cloud and chip suppliers, which affects supply chain and R&D investment directions.