Shenzhen – October 18, 2026 -- More than half of cloud data center capital expenditure is expected to go to memory and storage, excluding HBM, by 2027, according to a projection by Morgan Stanley's Daniel Yen presented at the 5th GMIF Innovation Summit. The forum, themed "The Future Built on AI Memory and Storage," was co-hosted by the Shenzhen Memory Industry Association (SMIA) and Peking University's School of Integrated Circuits at the Renaissance Shenzhen Bay Hotel.
Yen said global data center capex will stay strong in the coming years. He named six paths to overcome the "memory wall": process migration, design changes, advanced packaging, memory peripherals, PIM/CIM and new materials.
Inference shifts the cost equation toward memory bandwidth, capacity and efficiency
As AI moves from model training to large-scale commercial inference, the token is becoming a new unit of value. Inference requires low latency, high concurrency and sustained throughput, which makes memory and storage bandwidth, capacity and energy efficiency key drivers of system cost.
Prof. Yimao Cai, Dean of Peking University's School of Integrated Circuits, said HBM remains central to high-end AI compute, but yield, cost and thermal limits constrain broader adoption. NAND-based high-bandwidth technologies such as HBF are a major research focus because of their cost advantage. Hybrid multi-media architectures could balance performance, capacity and cost for models with hundreds of billions of parameters, he said.
Samsung and Sandisk cite agent-driven KV cache demand and a move from TCO to TVO
Kevin Yoon, CVP and CTO of Samsung Memory China, said AI agents are generating massive KV cache demand and driving rapid iteration of tiered storage. He cited Samsung's work on Z-NAND, PCIe Gen6 SSDs and ultra-high-capacity data center SSDs.
Maya Zhang, Senior Director of Product Marketing at Sandisk, called this the first year of the "inference boom." She said AI data centers are shifting from cost centers to value and capacity centers, and that evaluation is moving from TCO to TVO. In her view, the SSD is becoming a "token battery" that improves economics through token reuse.
Solidigm reports faster QLC SSD adoption as Arm pushes system-level balance
Benny Ni, GAR Sales VP at Solidigm, said high-capacity QLC SSDs are being adopted faster. He added that memory offloading has become a core data center design paradigm and that liquid cooling is a prerequisite for unlocking high-end storage performance.
John Xavier Lionel, Head of Global Storage Business at Arm, said AI-era competition is moving from chasing higher TOPS to balancing compute, memory, storage and data movement. Arm's cloud-edge-device framework weighs power, capacity, bandwidth, latency, security and cost at each tier.
At the edge, Silicon Motion AVP Stanley Huang said its PerformaShape technology allocates resources flexibly across agents and improves read/write consistency. BIWIN Chairman Sam Sun said inference is extending memory demand to edge and endpoint devices, pushing storage toward scenario-specific design.
Lenovo says fewer than 40% of enterprise agent projects scale with positive ROI
Tao Zhou, General Manager of the Server Division at Lenovo ISG China, said 80% of enterprises have deployed at least one type of agent, yet fewer than 40% of projects can scale and deliver positive ROI. Lenovo has built a full-stack "token factory" paradigm covering data governance, compute optimization and security management.
NEXWISE presented its KuberCloud token factory solution with heterogeneous compute management and token-based billing. Numbers Law showed its PMD (Parallel Magnetic Drive) technology, aimed at raising HDD IOPS and throughput. Infplane, Maxio and Genstoraige used AI SSD tiered offloading and memory expansion to lower the deployment threshold for large on-device models.
Forty exhibitors show PCIe Gen6 controllers, automotive UFS and 2.5D/3D packaging tools
The concurrent exhibition drew 40 companies across the value chain. Sandisk featured its iNAND AT EU752 automotive UFS 4.1, SN861 enterprise NVMe SSD and BiCS8 QLC NAND. Silicon Motion introduced the SM8466 PCIe Gen6 16-channel controller and the SM2524XT controller for edge SSDs.
Chinese controller makers including Maxio and YEESTOR showed PCIe Gen4/Gen5 and UFS controllers with enhanced LDPC. Lenovo exhibited its WenTian series of AI training-and-inference servers, and BIWIN showed enterprise SSDs and CXL products.
Attach Point, MICROVIEW and Maxwell Technologies displayed die attach and bonding equipment for 2.5D/3D advanced packaging. OKN Technology, Acroview Technology, Hongjingwei and Tytantest introduced test and burn-in systems for SSD, DRAM and NAND that reproduce high-load AI I/O scenarios. Hemei Jingyi and Silergy brought storage IC package substrates and dedicated power management ICs.
SMIA President Rixin Sun said GMIF will pursue a more specialized format connecting memory companies with AI compute and end-application players worldwide.