中国光刻机进入「DeepSeek时代」?国产DUV量产消息引发全球芯片股重挫
Related
Institutions: Comprehensive upward revision of Q1 DRAM and NAND Flash price growth forecasts for all products.
According to BlockBeats' latest memory industry survey, released on July 6th, demand from AI and data centers will continue to exacerbate the global memory supply-demand imbalance in the first quarter of 2026, further increasing manufacturers' bargaining power. Based on this, TrendForce has comprehensively revised upwards its Q1 price growth forecasts for all DRAM and NAND Flash products. It predicts that the overall Conventional DRAM contract price will increase by 90-95% from the 55-60% increase announced in early January, while the NAND Flash contract price will be revised upwards from 33-38% to 55-60%, with further upward revisions not ruled out. Click the original link below to join the Beating · Lark AI news channel for 24/7 monitoring of global AI hot topics and news.
DRAM weakened, and the memory supply chain continued its correction, rebounding intraday before coming under pressure again.
According to BlockBeats, on July 6th, based on BIT (bit.com) market data, the DRAM chip memory ETF rose and then fell back today, weakening again. The current price is around 64.5, with a daily decline of 1.3%. Overall, it shows a high-level oscillating downward structure, with insufficient rebound momentum. The trading volume increased in stages during the decline, indicating that selling pressure is still being released.
Citrini Analyst: China's CXMT tests pilot production line for bonding DRAM; South Korean media claims its technology and development speed may be ahead of its South Korean rivals.
According to a Odaily by Citrini analyst Jukan on the X platform, South Korean media reports that China's CXMT is currently testing a bonded DRAM pilot production line in Hefei, aiming to achieve high-performance DRAM without using EUV lithography. Bonded DRAM is a technology that manufactures the memory cell array and peripheral circuitry on separate wafers and then bonds them together. This method can produce ultra-high-density DRAM using only multi-patterned deep ultraviolet (DUV) lithography, without requiring EUV equipment. Samsung Electronics is developing its own bonded DRAM under Project B1b, and SK hynix is also advancing similar technology. South Korean media warns that assessments suggest CXMT may currently be ahead of its South Korean competitors in terms of both the technology itself and the speed of its development.
Trendforce: Traditional DRAM prices are expected to rise by 13% to 18% in Q3.
According to BlockBeats, on July 4th, Trendforce predicts that contract prices for traditional DRAM products will increase by 13% to 18% quarter-over-quarter in the third quarter of 2026.
TrendForce: The DRAM market will continue to face severe supply shortages in Q3.
According to BlockBeats, on July 3rd, TrendForce, a high-tech industry research firm, released its latest memory price survey, predicting that the DRAM market will continue to face a severe and widespread supply shortage in the third quarter of 2026. Despite this, slowing consumer application demand and a high base effect will slightly moderate the rate of increase, with DRAM contract prices expected to rise by +13% to 18% quarter-over-quarter. For NAND Flash, TrendForce believes that AI inference demand and large-scale data center construction will continue to drive most of the demand. However, with contract prices already at historical highs and slowing consumer market demand, customers' ability to absorb further price increases has reached its limit. Therefore, TrendForce predicts that NAND Flash contract prices will rise by +10% to 15% quarter-over-quarter, a significantly smaller increase than in previous quarters.
DeepSeek's open-source inference acceleration framework DeepSpec, when deployed on DSpark, improves the speed of V4 models by up to 85%.
According to Beating's monitoring, DeepSeek, in collaboration with Peking University, released a technical report on DSpark, a framework for accelerating speculative sampling, and open-sourced its full-stack codebase, DeepSpec. DSpark is currently deployed in DeepSeek-V4's online operations. While ensuring lossless output, DSpark improves the single-user generation speed of the Flash version by 60% to 85%, and the Pro version by 57% to 78%. DSpark's performance surpasses the original single-token multi-branch prediction (MTP-1) baseline, significantly increasing the overall system throughput under strict latency constraints. Previously, multi-token speculative sampling was difficult to implement in online production environments. Autoregressive draft models were too slow to generate, while parallel draft models, due to independent predictions at each position, resulted in extremely low acceptance rates for the latter half of long sequences. Blindly validating multi-token drafts under high concurrency would waste significant computing power validating errors destined to be rejected, leading to a severe collapse in the overall system throughput. Therefore, the industry largely limited online implementations to single-token prediction (MTP-1). DSpark overcomes the throughput degradation bottleneck under high concurrency. DSpark first uses the DFlash parallel backbone to generate hidden states, then adds an extremely lightweight Markov head. The Markov head injects associations between adjacent words serially at extremely low cost through table lookups and a single matrix multiplication. Simultaneously, the system integrates a confidence prediction head and a posterior calibration algorithm. To perfectly support zero-overhead scheduling in production environments and prevent future information leakage, the scheduler employs an asynchronous mechanism, dynamically determining the candidate word pruning length using historical predictions from two steps ago, completely preventing large models from validating high-risk tail misspellings under heavy load. In addition to DSpark, DeepSeek's open-source DeepSpec codebase provides built-in support for large open-source models such as Qwen3 and Gemma. DeepSpec offers a complete Python toolchain from downloading prompt words, rebuilding large model caches, training draft models to benchmark evaluation. Developers can directly utilize open-source scripts to customize and deploy dedicated acceleration modules for different large open-source models locally.