The Global Memory Innovation Forum 2026 (GMIF 2026) brought together 40 companies in Shenzhen to demonstrate how artificial intelligence is changing the commercial value of memory and storage.
Instead of concentrating only on drive capacity and headline transfer speeds, exhibitors presented technologies intended to improve AI inference, extend limited GPU memory, move data more efficiently and support increasingly demanding cloud, edge, automotive and on-device workloads.
The fifth GMIF took place in Shenzhen on September 22–23, 2026, under the theme “The Future Built on AI Memory and Storage.” It was co-hosted by the Shenzhen Memory Industry Association and Peking University’s School of Integrated Circuits, according to the official GMIF 2026 website.
The concurrent exhibition covered enterprise SSDs, embedded storage, memory controllers, AI servers, packaging equipment, testing systems, substrates, power-management chips and software-defined computing infrastructure.
AI inference is changing what storage must deliver
The central development visible at GMIF 2026 was the closer integration of storage with AI computing.
Large models rely on rapid access to model parameters, vector databases and previously processed information. During inference, systems also create a key-value cache, commonly known as a KV cache, to avoid repeatedly processing the same data.
When these caches exceed available GPU memory, they may be moved to system memory or fast storage. NVIDIA’s technical guidance identifies long-context applications, high-concurrency workloads and memory-constrained deployments as situations where KV-cache offloading can be useful.
Several GMIF exhibitors presented SSD-based architectures intended to support this process.
Infplane demonstrated an AI SSD architecture that uses storage for KV-cache persistence, model-data tiering and the preloading of frequently required mixture-of-experts parameters. Genstoraige presented an edge-inference system that coordinates HBM, DRAM and storage-class memory, while Quanxing Technology showed an SSD designed to operate both as conventional storage and as an external GPU memory pool.
The companies reported significant improvements in model capacity, token throughput and inference costs. However, these performance figures were provided by the exhibitors and should not be treated as independently verified comparisons.
Buyers assessing AI-storage systems should request benchmarks using their own models, context lengths, batch sizes and concurrency requirements. A solution that performs well under one inference workload may produce different results when the model, hardware platform or data-access pattern changes.
PCIe Gen6 controllers target the next generation of AI infrastructure
Enterprise storage suppliers are also preparing for substantially faster server interfaces.
Silicon Motion presented its SM8466 PCIe Gen6 enterprise SSD controller, which the company says can deliver sequential throughput of up to 28GB/s and seven million random-read input/output operations per second. The controller is designed to support SSD capacities of 512TB and above for data-intensive AI workloads.
These specifications are consistent with the company’s official SM8466 announcement, although actual drive performance will also depend on NAND configuration, firmware, thermal management and server architecture.
The underlying PCIe 6.0 standard supports a raw data rate of 64GT/s, twice that of PCIe 5.0. PCI-SIG identifies data centres, AI, machine learning, high-performance computing and automotive systems among its intended applications. The standard also introduces PAM4 signalling, forward error correction and a new flow-control-unit encoding system. PCI-SIG provides further technical details.
For infrastructure buyers, the move to PCIe Gen6 will involve more than replacing existing drives. Servers, backplanes, controllers, cooling systems and software must support the required interface and performance levels. Procurement teams should therefore verify system-level compatibility rather than comparing controller specifications in isolation.
Sandisk combines enterprise and automotive storage
Sandisk used GMIF 2026 to present products serving both enterprise AI infrastructure and intelligent vehicles.
Its SN861 NVMe SSD targets AI, machine-learning, high-performance-computing and mixed enterprise workloads. Sandisk’s published specifications list capacities of up to 15.36TB across U.2, E1.S and E3.S form factors.
For automotive applications, the company showed its iNAND AT EU752 UFS 4.1 embedded flash drive. Sandisk says the product uses BiCS8 NAND, supports capacities from 128GB to 1TB and is designed for advanced driver-assistance systems, vehicle computing and other data-intensive automotive functions. The company’s product announcement states that the device was shipping to selected customers and partners.
Automotive storage buyers should examine more than interface speed. Relevant considerations include operating-temperature range, endurance, qualification status, data retention, lifecycle availability and the supplier’s change-notification procedures.
Memory expansion connects processors, accelerators and storage
The exhibition also included technologies such as CXL, DDR5 server memory and high-capacity enterprise SSDs, reflecting the growing need to coordinate different tiers of computing memory.
The Compute Express Link Consortium describes CXL as a cache-coherent interconnect that enables processors, memory-expansion devices and accelerators to share resources while maintaining memory coherence.
For AI infrastructure, this can make it possible to use available memory more efficiently across servers and accelerators. However, deploying CXL-based systems requires compatible processors, devices, firmware, operating systems and workload-management software.
BIWIN presented storage and memory products spanning enterprise, edge, automotive and endpoint applications, including PCIe SSDs, CXL products, DDR5 registered memory and embedded-storage formats.
Lenovo exhibited WenTian servers for AI training and inference, connecting the storage and memory technologies displayed elsewhere at the event with complete computing platforms.
This combination made GMIF relevant not only to component buyers but also to data-centre operators, server manufacturers, AI-platform developers and companies planning integrated infrastructure projects.
Advanced packaging and testing become more important
AI memory development does not stop at the chip or SSD controller. Packaging, inspection and reliability testing formed another substantial part of the GMIF exhibition.
Participating companies demonstrated technologies including:
- Wafer-level thermocompression and hybrid bonding
- Equipment for HBM and ultra-thin die stacking
- Precision ball-placement systems
- Automated optical inspection
- Wafer grinding, polishing and laser dicing
- Memory burn-in and system-level testing
- Semiconductor defect inspection and metrology
- Reliability and environmental test chambers
- High-density IC package substrates
These areas are becoming increasingly important as AI hardware combines processors, memory and supporting components within more compact packages.
Equipment suppliers reported placement accuracies measured in micrometres, high levels of parallel device testing and support for advanced memory formats. Such figures should be verified against the required production process, substrate, package design and acceptance criteria.
Companies interested in the broader Chinese semiconductor manufacturing ecosystem can also read ChinaExpoHub’s coverage of ICES 2026, which examined integrated-circuit design, wafer fabrication, advanced packaging and testing equipment.
Edge AI and intelligent vehicles broaden the market
GMIF 2026 also illustrated how AI-storage demand is expanding beyond centralised data centres.
Exhibitors presented embedded memory, automotive-grade storage, industrial SSDs and compact products for:
- AI-enabled personal computers
- Smartphones and smart wearables
- AI and augmented-reality glasses
- Robotics and embodied intelligence
- Intelligent connected vehicles
- Industrial equipment
- Security and monitoring systems
These applications have different requirements from enterprise data centres. Power consumption, physical size, temperature tolerance, shock resistance and reliable operation with intermittent connectivity may matter more than maximum transfer speed.
International buyers should therefore define the intended operating environment before discussing products with suppliers. A high-performance consumer SSD is not automatically suitable for an industrial controller or vehicle platform.
What technology buyers should verify
The exhibition showed a wide range of technically ambitious products, but many reported performance figures originated from the companies presenting them.
Before selecting a memory, storage or infrastructure partner, buyers should request:
- Independent or reproducible benchmark results
- Workload and test-environment documentation
- Endurance and data-retention specifications
- Power-consumption and thermal-performance data
- Interface, firmware and operating-system compatibility
- Automotive or industrial qualification records where applicable
- Production capacity and expected lead times
- Product-change and end-of-life notification policies
- Warranty, firmware-update and technical-support arrangements
- Information about intellectual-property ownership and licensing
Buyers should also distinguish between a product demonstration and proven volume production. Prototype performance, commercial availability and consistent manufacturing yield are separate questions.
GMIF presents a specialised view of China’s AI-storage ecosystem
Compared with a general electronics exhibition, GMIF offers a more specialised view of the organisations developing memory controllers, storage products, AI servers, packaging equipment and testing systems.
The 2026 exhibition indicates that storage is increasingly being treated as an active part of AI-system design rather than a passive repository for data. Performance now depends on how effectively information moves between SSDs, system memory, accelerators and processors.
This shift creates opportunities for component manufacturers, infrastructure providers and international technology buyers. It also makes technical verification more important because the commercial value of a storage product depends on how it performs within a complete workload and system architecture.
ChinaExpoHub previously covered the forum’s speakers and programme in its GMIF 2026 preview. According to the organiser’s post-event report, the completed exhibition brought together 40 companies across the memory and storage value chain.
