Introduction: Redefining Storage in the AI Era
As AI models scale to trillions of parameters and GPU/NPU computing power surges tenfold annually, storage technology has struggled to keep pace. This lag is becoming a critical bottleneck for AI’s rapid advancement. At the MemoryS 2025 Flash Memory Market Summit, Yeestor unveiled a revolutionary concept: the “AI Storage Chip.” This innovation reimagines storage not as a passive data repository but as an active, intelligent hub, setting a new standard for the chip industry.
In an exclusive interview with Jiwei.com, Yeestor’s Chairman, Wu Dawei, explained, “Our AI Storage Chip isn’t just an upgrade—it’s a complete rethinking of storage architecture. We’re transforming storage from a data container into a self-evolving intelligent system, where every bit of data fuels greater intelligence.”
Teaching Storage to Think #
Traditional storage chips handle basic read/write tasks, responding passively to instructions. Yeestor’s AI Storage Chip, however, integrates AI-driven storage agents at the chip level. These agents enable seamless collaboration between storage and computing units, making data processing smarter and more efficient. “We’re not just storing data for AI,” Wu Dawei noted. “We’re teaching storage to think.”
By embedding a Neural Network Processor (NPU) within the storage controller, Yeestor creates a collaborative storage-compute framework. This approach enhances the cognitive capabilities of storage media, offering a bold new path for next-generation storage architectures.
A New Blueprint for AI and Flash Memory #
Yeestor’s innovation draws inspiration from an Apple research paper, “LLM in a Flash: Efficient Large Language Model Inference with Limited Memory.” This work highlighted how storage can optimize Large Language Models (LLMs) in memory-constrained environments. Yeestor seized this idea, integrating flash memory with LLMs to unlock new possibilities for AI efficiency.
Meanwhile, SanDisk’s High Bandwidth Flash (HBF) technology has boosted data transfer speeds and storage efficiency, providing a robust foundation for large-scale AI models. This synergy of flash memory and AI aligns with Yeestor’s vision of sparse model architectures—moving from macro-level modularity to micro-level neuron sparsity. Such architectures leverage flash memory’s strengths, optimizing LLMs for efficiency and cost.
Yeestor is now developing AI storage chips tailored for neuron-sparse models, aiming to drive AI toward greater performance and scalability through innovative storage design.
The Technology Trifecta: Control, Connection, Integration #
Yeestor’s AI Storage Chip is built on three core pillars: advanced storage control, high-speed storage-compute interconnection, and storage-compute integration. Together, these technologies redefine how data flows and empower intelligent storage.
- Optimized Storage Control: Advanced scheduling maximizes bandwidth and minimizes latency, ensuring fast, smooth data access for AI applications.
- High-Speed Interconnection: Smarter, low-latency channels with Quality of Service (QoS) support enable rapid data transfer to computing units, boosting system responsiveness.
- Storage-Compute Integration: By performing computations within or near storage units, Yeestor eliminates unnecessary data movement, dramatically improving energy efficiency and reducing latency.
“Mass production is the real challenge,” Wu Dawei emphasized. “Our technical path transforms data flow, making storage chips not just faster but smarter.”
Powering the AI Ecosystem #
Yeestor’s AI storage solutions span smart devices, vehicles, and computing centers, creating a robust ecosystem across AI phones, PCs, cars, IoT, and servers.
- AI Phones: Yeestor’s eMMC and UFS 3.1 storage controllers, with read/write speeds exceeding 2GB/s, power smartphones for major manufacturers, delivering seamless AI experiences.
- AI Cars: Automotive-grade chips, including eMMC and upcoming UFS solutions, support brands like Dongfeng and Geely, meeting rigorous standards for smart, electrified vehicles.
- AI Servers: The AI-MemoryX solution expands memory capacity for large model training, with plans for CXL-based storage-compute chips to enhance data center efficiency.
- AI PCs: PCIe 5.0 controllers, supporting up to 14.5 GB/s, meet the high-performance demands of AI-driven edge computing.
- AIoT: Over 100 million eMMC, SD, and TF cards ship annually, enabling smart homes, security, and industrial IoT with reliable, low-power storage.
The Hetu Project: A Vision for Intelligent Storage #
In 2025, Yeestor launched the “Hetu Project,” a bold R&D initiative to advance storage-compute integration. This project aims to evolve storage from passive containers to intelligent carriers, capable of understanding, optimizing, and even mining data value autonomously. “Future SSDs will be thinking agents,” Wu Dawei said. “They’ll not only store data but also enhance its value, driving a true storage revolution.”
Leading the Future of Storage #
By giving storage chips an AI brain, Yeestor is redefining storage as a form of computing power. With its innovative technology and forward-thinking vision, Yeestor is poised to lead the storage industry into a smarter, more efficient future, delivering cutting-edge solutions for global AI advancement.