At Computex 2026, Phison unveiled its next-generation PCIe 6.0 SSD controller solution, the "PS5303-X3", which supports capacity of up to 2PB. According to the demonstration, the PS5303-X3 delivers sequential read and write speeds of 28,000 MB/s and random read/write performance of 6.8 million IOPS, doubling the performance of PCIe 5.0 solutions. Phison's X3 SSD prototype designs include E3.S and E1.S interface form factors. The solution is reportedly nearly finalized, with customer samples expected to be shipped in December, followed by mass production targeted for mid-2027.

According to foreign media reports, Google has commissioned Marvell to design a custom networking chip for its next-generation TPUs. The chip is expected to be manufactured using Intel's 18A or 18AP advanced process nodes. Mass production is planned for the end of 2027, and the chip will be paired with MediaTek-designed "Humufish" for Google.

Phison announced its operating results for May 2026. Consolidated revenue reached NT$22.828 billion, representing 13% month-over-month growth and 300% year-over-year growth, setting a new all-time monthly revenue record. Cumulative revenue for the first five months of 2026 reached NT$84.002 billion, up 229% yoy, also marking a record high for the same period in the Company's history.  K.S. Pua, CEO of Phison Electronics, stated that as AI inference adoption accelerates, demand for NAND storage remains exceptionally strong, with no signs of slowing. Meanwhile, NAND supply remains relatively constrained as manufacturers continue to take a cautious approach toward capacity expansion, resulting in a persistently tight supply-demand environment.

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.

In a recent interview, Arm CEO Rene Haas stated that the U.S. attempt to ban the export of AI CPUs to China is practically unfeasible. He pointed out that unlike Nvidia GPUs, which can be easily restricted by setting performance thresholds and memory bandwidth limits, CPUs are general-purpose core components widely used across various sectors, including servers, PCs, and mobile devices. Because it is difficult to accurately determine whether a CPU is exclusively dedicated to AI applications, enforcing such a ban would require the U.S. to impose comprehensive restrictions on all CPU exports, which lacks operational viability.

Regarding the evolution of the artificial intelligence market, TSMC Chairman C.C. Wei stated that AI is currently transitioning from generative query models to agentic and action-oriented modes. This trend has significantly increased token consumption and computing power demands for large language models, providing strong support for the advanced semiconductor industry. He emphasized that both downstream customers and end markets remain optimistic about AI prospects. Based on its technological leadership and extensive customer base, TSMC remains highly confident in the overarching AI development trend over the next several years, believing that semiconductor demand possesses solid fundamental support. Measured in U.S. dollars, the company expects its full-year revenue growth this year to exceed 30%.

According to media reports, nine US trade associations representing the automotive, medical, and retail sectors recently sent a joint letter to Treasury Secretary Bessent and Commerce Secretary Lutnick, highlighting the memory chip shortage crisis triggered by the AI boom. The associations pointed out that the excessive consumption of production capacity by AI data centers has led to soaring chip prices, which now threatens traditional manufacturing supply chains and could potentially drive up inflation. They urged the Trump administration to intervene by coordinating with manufacturers to expand production in the United States and its allied nations, while utilizing mechanisms such as the CHIPS Act to ensure supply security across all industries.

Google has unveiled Gemma 4 12B, a new unified multimodal model. Its biggest innovation is the complete removal of the "encoder" component, which has been essential in traditional multimodal models, achieving a qualitative leap in local deployment and inference efficiency on consumer-grade hardware. Reportedly, Gemma 4 12B can process text, vision, and audio on laptops with 16GB of RAM, and also has certain coding capabilities. It is well suited for on-device data analysis and information extraction, bringing agentic and multimodal capabilities to laptops.

Transcend Information reported May 2026 revenue of NT$6.298 billion, with a year-on-year increase of 469.83% and a month-on-month decrease of 15.5%. The combined revenue for April and May reached NT$13.75 billion, surpassing the NT$13.63 billion recorded in the first quarter. Transcend Chairman Shu Chien-wan stated that the results were primarily driven by high market prices. Currently, the company has a stable long-term supply of DDR4 from Samsung Electronics, which will continue at least through the end of 2027, with a high likelihood of further extension beyond that.

According to Taiwanese media reports, Vicky Tseng, Chairperson of Goldkey Technology, stated that AI servers are driving a surge in memory demand, with each AI server requiring approximately eight times the memory capacity of a traditional server. Moreover, three major memory suppliers continue to prioritize production capacity for HBM, squeezing the supply of DDR4 and DDR5. The industry has entered a new supercycle, with shortages potentially lasting until 2028. Currently, market order visibility has extended to 2028 while some customers have even begun negotiating supply plans for 2029–2030 with the manufacturers. Zeng Wanquan, Vice General Manager, added that suppliers are still prioritizing CSPs and large data centers, leaving the enterprise market supply tight. Based on current market supply and demand conditions, DRAM contract prices are expected to rise another 30% in the third quarter, while NAND Flash prices could increase by as much as 60% to 70%.

According to Korean media reports, Hanmi Semiconductor unveiled its "TC Bonder 4" for HBM4 production and "Wide-Type TC Bonder" for next-generation HBM manufacturing at Computex 2026, which began mass production this year,. The wide-type TC bonder supports the production of HBM using larger DRAM dies. Hanmi Semiconductor stated that as HBM chip area increases, the number of TSVs (Through-Silicon Vias) and I/O interfaces can be stably increased, memory capacity and bandwidth can also be improved in comparison to previous technologies.

Microsoft has unveiled the Surface Laptop Ultra, powered by the NVIDIA RTX Spark SoC, which combines an ultra-efficient CPU with a powerful RTX GPU to deliver up to 1 petaflop of AI compute. The Surface Laptop Ultra supports for up to 128GB of LPDDR5X unified memory and features a 15-inch mini-LED PixelSense Ultra touchscreen with a peak HDR brightness of 2000 nits, a pixel density of 262 PPI, high-precision color accuracy, and all-day battery life. Additionally, the device comes equipped with a full set of convenient ports, including USB-C, USB-A, HDMI, a headphone jack, and an SD card reader, no dongles necessary.

According to foreign media reports, Qualcomm may break away from its traditional business model of reaping high profits from premium chipset pricing and instead offer significant discounts to Samsung. Earlier reports indicated that Samsung plans to equip 50% of its Galaxy S27 series smartphones with the Exynos 2700 chip, with a sharp increase from the 25% share of the Exynos 2600. As a result, the Exynos 2700 is expected to replace a large number of Snapdragon chips next year, eating into Qualcomm's market share. The Snapdragon 8 Elite Gen 6 Pro will be the first to adopt TSMC's 2nm process, with a per-chip cost potentially exceeding $300, forcing phone makers to compromise on other components to maintain profitability. As Qualcomm's largest chip customer, Samsung is expected to benefit from a discount agreement to lower procurement costs. It is reported that Qualcomm may offer Samsung a 16% discount compared to the standard Snapdragon processor, making it 12% cheaper than the Exynos 2700.

During the Goldman Sachs event held at COMPUTEX 2026, MediaTek announced that its next-generation chips will exclusively adopt Intel's EMIB-T advanced packaging technology, moving away from TSMC's CoWoS solution. MediaTek revealed that tape-out for the project is targeted for the fourth quarter of 2026, with mass production planned for the fourth quarter of 2027. Although it is currently unclear which specific chips are covered under this project, industry insiders widely believe that custom AI chips and CPUs will be involved.

Phison CEO Pua Khein-Seng stated that as AI data centers continue ramping up shipments and AI PCs gradually enter the market, the volume of generated data will increase significantly, and storage demand will only continue to expand. He noted that the biggest issue facing the memory industry today is no longer demand, but rather supply that is far from keeping up with the growth of tokens and data volumes. Although major manufacturers are increasing production capacity in hopes of delivering more output, the added capacity still falls far short of the growing demand for tokens. He expects next year's storage shortage to be even more severe than this year's