Author: Sophia Original
On July 16, ChangXin Memory Technologies (CXMT), a leading memory chip manufacturer, officially launched its IPO subscription. Public data shows that the issue price was set at 8.66 RMB per share. With 6.688 billion shares issued, the total fundraising amount reached approximately 57.919 billion RMB, bringing the listing valuation to nearly 580 billion RMB.
If the issuance goes smoothly, CXMT will surpass SMIC (Semiconductor Manufacturing International Corporation) to become the largest IPO in the history of the STAR Market, and simultaneously rank as the fifth-largest IPO in the history of China's A-share market (with total fundraising second only to Agricultural Bank of China, PetroChina, China Shenhua, and China Construction Bank). In terms of shipments and sales, CXMT has already become the No. 1 domestic and No. 4 global DRAM (Dynamic Random Access Memory) manufacturer.
In many people's minds, DRAM is merely the memory stick in a computer or a part of a smartphone's running memory. Unlike the CPU, which handles computing, or the GPU, which drives AI (Artificial Intelligence) training, it has long played a low-profile 'behind-the-scenes supporting role.' However, given the accelerating integration of AI and IoT (Internet of Things), this perception is being redefined. As more devices gain sensing, computing, and autonomous decision-making capabilities, memory chips are evolving from basic components in consumer electronics into the data infrastructure of an intelligent world.
Smart vehicles continuously process massive amounts of sensing data from cameras, radars, and LiDARs; humanoid robots run Vision-Language-Action (VLA) large models in real time; the Industrial Internet continuously accumulates equipment operation data and digital twin models; and edge AI devices increasingly execute inference tasks locally to reduce latency and enhance reliability. Although these scenarios vary in form, they share a common characteristic: devices are continuously generating, retrieving, and processing data, and all these processes rely heavily on the support of high-speed memory systems.
Two Major Key Terminals for the Future Memory Market
In the next decade, what determines the scale of the memory market will no longer be just how many devices are connected to the network, but how many devices begin to possess intelligence. Smart vehicles and humanoid robots are precisely the two most representative terminals under this trend.
Smart Vehicles
As intelligent driving technology accelerates its evolution toward L2+ and higher stages, vehicles are transforming from traditional mechanical products into mobile intelligent terminals equipped with sensing, computing, and decision-making capabilities, a shift that is significantly boosting in-vehicle memory demand. In the past, automotive electronic systems mainly focused on powertrain control, safety assistance, and infotainment, with relatively limited memory requirements. However, in the era of intelligent driving, vehicles need to continuously collect environmental information through various sensors such as cameras, millimeter-wave radars, and LiDARs, and rely on in-vehicle computing platforms to perform real-time object recognition, path planning, and driving decisions. This has led to rapid growth in the data processing capabilities and memory capacity of vehicles.
Especially in L2+ intelligent driving models, vehicles have already acquired capabilities similar to 'edge AI nodes.' On the one hand, massive amounts of data generated by sensors need to be cached and processed in real time at the vehicle end to reduce data transmission latency and ensure driving safety. On the other hand, as intelligent driving algorithms are steadily advancing, in-vehicle systems also need to store increasingly complex AI models, software systems, and data continuously updated via OTA (Over-The-Air). Therefore, compared to traditional vehicles, L2+ and above models have a substantially increased demand for high-speed memory products such as DRAM and LPDDR.
Data provided by Micron shows that the average memory capacity of current L2+ models has already reached more than five times that of ordinary models. As the penetration rate of intelligent driving steadily improves, the proportion of L2+ models is expected to exceed 20% this year, and is projected to further increase to over 40% by 2030. It is foreseeable that the wave of vehicle intelligence will become a critical force driving the growth of memory demand.
More importantly, the intelligence trend in the automotive industry is changing the position of memory chips in the entire vehicle value chain. In the past, memory was more often viewed as an auxiliary component in automotive electronic systems. However, as intelligent driving becomes a core competency in the competition for NEVs (New Energy Vehicles), memory has gradually become critical infrastructure determining vehicle computing performance and intelligent experience. In the future, automotive competition will no longer be solely about battery capacity, powertrain performance, and mechanical design; it will also increasingly rely on intelligent foundational capabilities, including computing chips, memory chips, and data processing capabilities.
Humanoid Robots
If smart vehicles represent the development direction of mobile intelligent terminals, then humanoid robots may become a more prominent emerging field for future memory demand growth. Unlike vehicles, which primarily focus on environmental perception and path decision-making around driving scenarios, humanoid robots need to achieve a closed loop of perception, understanding, decision-making, and action in more complex and open environments, posing higher requirements for computing and data processing capabilities.
A humanoid robot needs to simultaneously process data streams generated by multiple sensors, including vision, touch, force, and posture, and rely on the Vision-Language-Action (VLA) large model to perform continuous real-time actions such as environmental understanding, fine hand manipulation, whole-body balance, and autonomous obstacle avoidance. These computing tasks cannot be entirely reliant on the cloud, as every movement of the robot requires millisecond-level response, and network latency could directly affect control precision or even lead to task failure. Therefore, a massive amount of multimodal sensing data, model parameters, and iterative data during motion must be read and written at high speeds locally, posing higher demands on the capacity, bandwidth, and latency capabilities of memory chips.
Meanwhile, the application environments of robots further amplify their memory requirements. L2+ smart vehicles mainly operate in relatively structured environments such as roads and traffic rules, whereas humanoid robots will need to enter unstructured scenarios like homes, factories, and warehouses in the future. They must not only perceive the surrounding environment in real time but also continuously learn and optimize their own behaviors. This means that the volume of data generated by robots will continuously accumulate as task complexity increases, creating long-term demands for local memory capacity and data processing capabilities.
In the future, robots will not only need to execute preset actions but also leverage large models to understand environments, learn tasks, and make autonomous decisions. This implies that more complex AI models need to run behind the scenes, including visual recognition models, natural language interaction models, and world models used for action planning and environmental understanding. All these models require high-speed, large-capacity memory as support.
According to Micron's forecast, the average memory capacity of humanoid robots could reach about 10 times that of L2+ smart vehicles. As the embodied intelligence industry gradually enters a stage of scaled development, memory chips are expected to become critical infrastructure supporting the enhancement of robots' intelligence capabilities. In terms of market size, a special report by Barclays estimates that the current global humanoid robot market is valued at approximately $2 billion to $3 billion, which could grow to $10 billion to $25 billion by 2030, and even reach $200 billion in an optimistic scenario by 2035. Along with the rapid development of the robotics industry, memory chips will usher in another major growth scenario following smart vehicles.
Strengthening the 'Foundation of China's Intelligent Manufacturing'
If we shift the perspective from a single memory chip company back to the entire IoT (Internet of Things) industry chain, the significance of CXMT's IPO becomes even more noteworthy. Over the past decade, China's IoT industry has experienced rapid development from 'device connectivity' to 'intelligent applications,' forming a relatively complete industrial system in segments such as sensors, communication networks, and cloud computing platforms. However, at the same time, the underlying memory segment that supports data flow and intelligent computing, especially critical components like high-performance DRAM, remains a significant bottleneck in the industry chain. The development of CXMT is precisely aimed at strengthening this critical link.
The core logic of the IoT industry is essentially to endow physical objects in the real world with sensing, connectivity, computing, and decision-making capabilities. A complete intelligence chain typically includes: sensors responsible for collecting environmental data, MCUs (Microcontroller Units) or processors handling basic control, communication modules managing data transmission, and edge computing nodes undertaking real-time analysis. Memory chips provide underlying support for data caching, model execution, and intelligent applications, ultimately driving AI models to achieve intelligent control over devices. In this process, any capability shortfall in a single segment could limit the overall intelligence level of the system.
From the perspective of the industrial landscape, China has already established strong competitiveness in the fields of IoT front-end devices and communication infrastructure. Whether in smart homes, industrial equipment, or smart city applications, a large number of domestic sensors, controllers, and terminal devices are being rapidly adopted. Meanwhile, communication technologies such as 5G, NB-IoT, and LoRa are steadily maturing, providing a network foundation for massive device connectivity. The development of cloud computing, big data platforms, and AI (Artificial Intelligence) algorithms has also further enhanced data processing and application capabilities. However, as IoT enters the AIoT stage, devices are no longer just simply collecting and uploading data; they need to perform real-time computing and intelligent decision-making at the edge or even the terminal side. At this point, the importance of underlying memory capabilities begins to stand out.
Especially in scenarios like the Industrial Internet and intelligent manufacturing, the scale of data generated by equipment is growing rapidly. A smart machine tool may continuously collect data on operating status, vibration, and temperature; an automated production line needs to process machine vision, quality inspection, and equipment predictive maintenance information in real time; and a digital twin system requires long-term accumulation of equipment operation data and model parameters. Not all this data can rely on cloud processing, which is constrained by network latency and involves industrial data security issues. Therefore, an increasing number of intelligent systems are developing toward edge computing and local intelligence, making high-speed, large-capacity, and highly reliable memory chips a critical foundation for this trend.
From this perspective, a truly independent and controllable IoT system requires not only independent sensors, independent communication networks, and independent computing platforms, but also a data storage foundation with independent assurance capabilities. The development of CXMT is precisely about filling this underlying capability gap. As the only domestic enterprise in mainland China to achieve large-scale mass production of DRAM, CXMT's listing marks a breakthrough from 0 to 1 in the most severely 'chokehold' segment of memory chips, completing the last core piece of the puzzle for independent and controllable semiconductors, following logic chips (SMIC) and packaging and testing (JCET).
The global DRAM market has long been monopolized by three giants (with a combined market share of over 90%). Entering the market with an approximate global market share of 7.67%, CXMT not only provides a domestic substitution option but also mitigates the risk of external supply disruptions at the supply chain security level, reshaping regional supply chain resilience. Currently, CXMT's products serve not only traditional markets such as PCs and servers but also play a vital role in emerging fields like smart vehicles, industrial control, edge computing, and AI terminals. Moving forward, its fundraising will focus on the R&D of HBM (High Bandwidth Memory) and DDR5/LPDDR5X, directly addressing the urgent demand for high-bandwidth, low-power memory from AI servers, humanoid robots, and advanced intelligent driving, thereby resolving the structural pain point of relying on imports for high-end memory power.
Conclusion
From connecting everything to bearing intelligence, the IoT industry is entering a new stage of development. Future competition will be not only about the number of devices but also about data processing capabilities and intelligence levels. For China, filling the gap in this critical underlying memory capability is not only related to the perfection of the semiconductor industry chain but will also provide a more solid technological foundation for the development of intelligent manufacturing, embodied intelligence, and AIoT. The growth of CXMT is precisely an important industrial footnote in China's stride toward the intelligent era.
References:
Micron: Humanoid Robot Memory Capacity Reaches 10 Times That of L2+ Vehicles, Potentially Sparking the Next Supercycle — Humanoid Lecture Hall
Memory Supercycle Erupts: How Standard and Niche DRAM Empower Through Differentiation and Benefit from Resonance? — IT Home
CXMT, the Largest IPO in STAR Market History, Opens for Subscription Today; Product Market Share Expected to Continuously Increase — China Blue News
Largest IPO in STAR Market History: CXMT Opens for New Share Subscription Today, Raising Approximately 57.9 Billion RMB, with Alibaba Holding About 5% — Modern Express
CXMT Listing: A Critical Leap for China's Memory Chips — Red Star Capital Bureau