01. DRAM Supply to Stay Tight Through 2027, NAND to Ease
On July 30, TrendForce projected that DRAM supply will remain tight and prices firm through 2027, while NAND Flash supply may loosen in the second half of the year.
On the DRAM front, new capacity won't meaningfully come online until 2028 due to lengthy fab construction timelines. Moreover, HBM wafer consumption far exceeds that of commodity DRAM, constraining overall supply growth. North American cloud service providers continue to ramp up capital spending, and the commercialization of agentic AI is driving demand, widening the supply-demand gap by an estimated 1 to 2 percentage points.
For NAND, capacity is gradually ramping up while consumer electronics demand remains soft, raising the risk of a structural supply-demand reversal. However, if the widespread adoption of agentic AI boosts SSD demand, the market could move closer to balance.
02. Surging Memory Costs Reshape Smartphone Economics; SoC Shipments Drop 15% in First Half
On July 31, Commercial Times reported that Counterpoint data showed global smartphone SoC shipments fell 15% year-over-year in the first half of 2026, weighed down by climbing memory prices and extended replacement cycles. Both MediaTek and Qualcomm saw declines exceeding 25%. In the second quarter, mobile memory prices surged more than 300% annually, pushing memory costs above SoC costs across the board. Full-year SoC shipments are forecast to drop 14% year-over-year, with entry-level segments plunging over 30%.
03. AMD Unveils Kria AI System Module to Challenge Nvidia's Jetson Thor
On July 27, Commercial Times reported that AMD introduced the Kria AI system module at its Advancing AI 2026 conference, equipping it with the Ryzen AI Embedded X100 processor to take on Nvidia's Jetson Thor. Industry analysts noted that the two chipmakers' race to power the "robot brain" will drive demand for TSMC's advanced process nodes, industrial computing, and machine vision.
The X100 integrates up to 16 Zen 5 CPU cores, 40 GPU compute units, and a 50-TOPS NPU, manufactured on TSMC's 3nm/4nm nodes. Compared with Nvidia's Jetson T5000, it delivers up to 3.4x higher reliability in real-time control. Industrial PC makers including Aaeon, IBASE, and IEI Integration have signed on as launch partners, with products expected to hit the market in the fourth quarter.
04. Samsung Electro-Mechanics Raises Shenzhen MLCC Prices by 30%
On July 29, SemiMedia reported that Samsung Electro-Mechanics notified sales partners of a 30% price increase on all MLCC products supplied by its Shenzhen factory, effective August 1. The company cited structural growth in global electronics demand that has pushed supply chain pressures beyond what it can absorb internally.
Demand from AI servers and data centers is the primary force reshaping the MLCC market. Rising power consumption in accelerators is driving up demand for high-reliability MLCCs, but manufacturing complexity and lengthy qualification cycles are limiting supply expansion. The company's MLCC book-to-bill ratio reached 1.31 in June, with inventory for some general-purpose products falling below 30 days.
05. First U.S.-Made Nvidia GB300 Chip Rolls Off Production Line
On July 27, MyDrivers reported that the first Nvidia GB300 AI chip produced on U.S. soil has completed 4nm fabrication at TSMC's Arizona Fab 21, marking a milestone in America's capacity to mass-produce advanced AI GPUs.
However, the chips still need to be shipped back to Taiwan for CoWoS packaging. TSMC is planning to invest up to $265 billion in the U.S. and has already broken ground on two advanced packaging facilities in Arizona. Foxconn and Wistron are also setting up server assembly lines in Texas. Once CoWoS production comes online, AI chips could be manufactured entirely in the U.S., potentially reshaping the global semiconductor supply chain.
06. Moonshot AI Seeks Nvidia Blackwell Chips to Train Next-Generation Model
On July 30, Economic Daily News reported that Chinese AI startup Moonshot AI is seeking to secure additional Nvidia Blackwell chips to train its next-generation model, Kimi K4.
Earlier reports indicated that Moonshot AI has been running workloads on servers located in Thailand equipped with Nvidia GB300 chips—the Blackwell generation—likely for training its latest large model, Kimi K3.