Storage Becomes Core Infrastructure for AI

Open a server cabinet in a data center and you will often find a dozen or so solid-state drives (SSDs). These drives carry the burden of data storage, while the controller chip inside each SSD acts as its brain, efficiently and precisely scheduling data as it moves in and out of storage cells.

Storage is one of the key infrastructure layers of the large-model era. Wherever there is data, storage is needed. As data becomes a core resource in AI, storage technology determines how efficiently large models process data and affects both training and inference speed. As training datasets grow exponentially, balancing storage cost and performance becomes increasingly important.

“Storage is a definite large market,” said Zhong Xiaohui, CFO of storage-chip design company InnoGrit, in a recent interview. Computing power, storage power, and transmission capability often reinforce one another and develop together. The large-model boom is pushing storage technology forward, while also raising the bar for differentiated competition and iterative technology upgrades among storage-chip companies and SSD vendors.

Computing Power Matters, and So Does Storage Power

The main hardware components of an SSD include NAND flash, DRAM cache, and a controller chip. If the data to be stored is compared to cars, an SSD is a giant parking lot, the storage cells in flash chips are parking spaces, and the controller chip is the lot’s “manager,” directing every car to enter and leave its space accurately, quickly, and efficiently.

In other words, the controller chip is the brain of the SSD. Through firmware code, it performs complex operations such as reading, writing, and encrypting data across flash and other components. InnoGrit develops these core storage components; its main products include complete SSD solutions and the internal storage controller chips used in SSDs.

Globally, the enterprise SSD market has long been dominated by South Korea’s Samsung Electronics and SK hynix, whose combined market share exceeds 70%. China’s domestic enterprise SSD supply chain is still catching up rapidly. When InnoGrit was founded in 2017, mainstream storage technology was shifting from hard disk drives to SSDs, and data transmission interfaces were moving from SATA ports to the faster PCIe high-speed serial expansion-bus standard. This transition created opportunities for domestic startups.

Reference image of InnoGrit SSD controller chips and SSD modules used in PCs and servers.

“From 2017 to 2021, domestic manufacturers were solving the question of whether technology could be turned into products. From 2021 to 2024, after solving the product problem, everyone had to solve the customer problem,” Zhong said. Domestic substitution has objectively advanced China’s storage-chip industry, and the AI boom is now an even larger boost.

At the “Wukong Zhisu” 6876P computing center in Haizhou District, Lianyungang, Jiangsu, rows of server cabinets perform about 6.87 quintillion floating-point operations per second. After deep software-hardware co-optimization for the full-parameter version of DeepSeek, Wukong Zhisu achieved ultra-high throughput of more than 6,900 tokens per second and supports enterprises in quickly launching AI applications within three minutes.

Computing power is important, but storage power is also important. Under the current large-model boom, data volumes are growing rapidly. Cold data is becoming rarer, while more data is turning into warm data or even hot data. In the past, data in financial systems or traditional data centers might be archived and no longer used after five years. That has changed: once models are running, they need real-time data throughput, turning formerly cold and warm data into hot data.

Jiangsu Zhonghuan Yunkong IoT Technology is using Wukong Zhisu to develop a large model for sanitation services in Haizhou and explore agent-based applications. Sanitation workers wearing smart wristbands can transmit vital signs, location, and task progress in real time, allowing the system to automatically adjust work routes. Unmanned street sweepers and drones share road conditions and garbage distribution data in real time, refreshing operating strategies at second-level frequency. Through virtual-physical mapping, coordinated scheduling, and autonomous collaboration, the traditional sanitation operating model becomes a new intelligent-agent model.

“In the past, we called smart sanitation information-based management. Now we call it embodied agents. The difference is that the system is no longer just a brain that looks at data; it turns every device, every worker, and every operating chain into a digital life form that can think, converse, and evolve on its own,” said Xu Lei, executive director of Zhonghuan Yunkong.

At the same time, applications such as DeepSeek have opened the door to inference and edge computing. Lightweight model design, hardware adaptation optimization, and lower model deployment costs are shifting computing demand from training to inference: training tasks remain concentrated in the cloud, while inference tasks move down to edge devices. As massive data “heats up,” computing demand evolves, and users pursue better inference experiences with extremely low latency, storage power faces higher requirements as well.

Large Models Drive Storage Technology Upgrades

In surface-finishing industries, excellent process know-how is an important technical barrier. The value of AI lies in continuously accumulating process data, developing smarter robotic brains, and further optimizing the process.

Founded in 2018, Sophis Intelligent Technology (Shanghai) moved from robot agency services to self-developed products. Its industrial robots focus on manufacturing scenarios such as grinding, cutting, drilling, and deburring. Founder Du Ling said robots can become smarter only through AI. The team has developed an intelligent grinder that can display data on smart terminals such as phones and computers. Edge-side operation keeps workers away from harsh environments involving dust and noise, while also recording process data and key parameters left by experienced workers during grinding, including pressure, temperature, rotational speed, and material. In the future, the company plans to develop a grinding large model to handle different product-finishing needs and improve processes.

This also highlights the urgent need for both computing and storage power. According to InnoGrit, raw-data collection and inference logs generate enormous data volumes, requiring massive writes and high-speed reads. Data cleaning and model training require high-concurrency mixed reads and writes, with higher demands on random performance. Different data application scenarios have already begun to demand differentiated storage-chip capabilities.

Traditional data centers needed SSD capacities of 4 TB to 8 TB. After DeepSeek appeared, demand for flash capacity rose to 32 TB, 64 TB, and even 128 TB. The larger the flash capacity, the harder the development challenge. It is like constructing a taller building: the higher it rises, the stricter the structural requirements. InnoGrit’s products from the same generation have already branched into a variety of segmented applications, which in turn demands stronger differentiated competition and technology iteration from storage-chip companies and SSD vendors.

AI is in fact shaping the direction of storage technology. “In the past, many domestic data centers were still using hard disk drives. Over the past two years, speed requirements drove replacement with SSDs, from SATA to PCIe 4.0, and now we are entering the PCIe 5.0 era,” Zhong said. After ChatGPT arrived in 2022, application markets represented by AIGC began to demand higher storage performance and capacity. The emergence of DeepSeek has promoted practical large-model inference applications. New-generation PCIe 6.0 SSDs and CXL-based storage-class memory solutions are also beginning to attract attention. These technologies will support large-model data-center cloud services and locally deployed all-in-one machines in new ways, accelerating the adoption of open-source large models such as DeepSeek.

“The arrival of AI has accelerated SSD market adoption. It took us about a year to enter standard server manufacturers, and in the first half of 2024 shipments climbed more than tenfold,” Zhong said.

Computing power, storage power, and transmission capability often promote one another and grow together. Domestic AI-chip companies are using the more open RISC-V architecture and exploring deployments from edge servers to cloud servers. Meng Jianyi, CEO of chip-design company Zhihe Computing, said that for RISC-V to break through in high-performance computing, it must not only enter the high-performance arena at the general-computing level, but also integrate AI-enhanced computing at the architectural level to become AI-native.

“Storage has always followed computing power and transmission capability. If either end rises, you must keep up,” Zhong said. She believes differentiated storage solutions must be matched to different application scenarios to support computing demand. This year, her team’s focus is to develop storage controller chips and solutions that meet future AI needs. “There are still many storage-controller manufacturers worldwide. To open a gap in this market, gain a foothold, and stand firmly, we must achieve what others do not have during iterative upgrades.”

In the future, domestic manufacturers will need to focus not only on meeting domestic substitution needs and sustaining product iteration, but also on their ability to go overseas. For China’s domestic storage products, entering global markets will be an inevitable stage over the next three to five years, or even five to ten years.