The core objective of the plan is to increase monthly DRAM capacity from the current approximately 550,000 wafers (which includes around 200,000 wafers produced at the Wuxi plant in China) to roughly one million wafers by 2030. This expansion will be heavily concentrated in the Yongin Semiconductor Cluster.

According to foreign media reports, the latest analysis from SemiAnalysis suggests that NVIDIA may use a smaller 96GB SOCAMM for its Vera CPU, rather than the previously anticipated 192GB. As a result, the total SOCAMM DRAM capacity in the Rubin NVL72 is expected to drop from approximately 55TB per rack to around 28TB. The change lowers estimated rack cost from $7.6 million to $6.8 million and reduces total cost of ownership from $4.16/hour/GPU to $3.90/hr/GPU.

With its eighth-generation TPU, Google breaks from the tradition of single-chip iteration by splitting training and inference into two distinct chips: the TPU 8t designed for large-scale training, and the TPU 8i optimized for inference and AI agents. This enables precise matching of the different demands of AI scenarios.

Compared to the 400+ layer solutions pursued by some competitors, Kioxia's calculations show that its 332-layer design offers approximately 10% lower cost per GB, about 10% improved power efficiency, and roughly 35% higher cell reliability.

Chey reiterated that the Memory supply bottlenecks triggered by the widespread adoption of AI are expected to persist until 2030. He pointed out that not only is global capital pouring into AI data center construction, but Nvidia's upcoming AI PC architecture will also create rigid demand for high-capacity Memory, providing long-term growth momentum for the Memory market.

K.S. Pua also emphasized that AI development inevitably requires both DRAM and NAND Flash, and the larger the AI model, the higher the demand for Memory capacity. As enterprises invest heavily in purchasing GPUs, revenue generation leads to data creation, which in turn drives Memory demand. Although major manufacturers are continuously expanding production capacity, the newly added capacity still falls far short of the demand growth. He expects the tight supply-demand situation to remain unresolved in the short term, with the Memory shortage next year expected to be even more severe than this year.

At COMPUTEX 2026, Phison announced a collaboration with Intel to integrate its Pascari aiDAPTIV Memory expansion technology into Intel's AI PC platform. It overcomes the traditional limitation of DRAM capacity in PCs, enabling AI PCs to run larger Mixture-of-Experts (MoE) models and agentic AI applications locally. According to Phison, aiDAPTIV leverages high-performance NAND Flash working with DRAM to create a novel AI Memory architecture, reducing the system Memory capacity required for large models. Internal tests show that a system equipped with aiDAPTIV requires only 16GB of DRAM to run an AI model with 26 billion (26B) parameters. Without this technology, 32GB of DRAM would be needed to handle the same workload.

Consequently, Memory prices will retain some upward momentum; smartphone and PC Memory prices will follow suit passively, albeit with narrowing quarterly increases, presenting an overall trend of shrinking volume but rising prices.

NVIDIA CEO Jensen Huang introduced the RTX Spark at GTC Taipei on June 1. Co-developed by NVIDIA and MediaTek, the chip features a Blackwell RTX GPU with 6,144 CUDA cores and fifth-generation Tensor Cores supporting FP4 precision. It is connected via NVIDIA NVLink®-C2C chip-to-chip interconnect to a high-performance 20-core NVIDIA Grace™ CPU, and supports up to 128 GB of unified Memory. According to NVIDIA, RTX Spark-powered laptops and compact desktops will be launched in the fall of this year by OEMs including ASUS, Dell, HP, Lenovo, Microsoft Surface, and MSI, followed by Acer and GIGABYTE.

The company aims to kick off full-scale production in the second half of next year, positioning itself to aggressively capture the enterprise SSD (eSSD) market for AI data centers.

After roughly 18 months of development, Crescent Island is expected to begin customer sampling later this year. Intel is also exploring manufacturing options that could leverage its own fabs, potentially enhancing cost competitiveness through greater vertical integration.

Samsung Electronics announced that it has begun shipping the industry’s first 12-layer HBM4E samples to major global customers, further strengthening its leadership in the next-generation HBM market.

Alper Ilkbahar, CTO and EVP of Sandisk, said in a recent media interview that the global AI race is becoming increasingly Memory-centric rather than compute-centric, which could exacerbate an unprecedented tight supply of Memory chips. He also provided an update on HBF, revealing that the wafer is currently being designed, samples of the HBF chip will be available by the end of this year, and the complete product with a controller is expected to launch next year.

According to media reports from China's Taiwan region, Acer Chairman Jason Chen recently pointed out that the shortage of key components in the PC industry remains severe. Among them, CPUs have replaced Memory as the supply chain's most critical "bottleneck."

Driven by supply-demand mismatches, contract prices for DRAM and NAND rose significantly quarter-on-quarter, and the bargaining power of original manufacturers improved markedly. The tight-supply-driven price increase pattern remained highly sustainable. According to CFM analysis, the global DRAM/NAND Flash market size reached US$137.14 billion in the first quarter of 2026, an increase of 81.6% quarter-on-quarter and 245% year-on-year, setting a new all-time quarterly high.