AI Demand Stretches the Chip Stack AI需求拉紧芯片全栈

HBM scarcity is lifting Chinese AI-chip prices by as much as 50%, while record TSMC sales, an accelerated 14A timetable and fresh CPO funding show demand pulling capital through the wider stack. The same expansion is also drawing closer antitrust scrutiny. HBM短缺令中国AI芯片价格最高上涨50%；与此同时，台积电创纪录营收、提前的14A时间表以及CPO新融资显示，AI需求正把资本牵引至更广泛的芯片全栈。伴随市场扩张，反垄断审查也在同步收紧。

AI & Accelerators AI与加速器

Broadcom Rides the Shift Toward Custom AI Accelerators Broadcom乘上定制AI加速器浪潮

The Next Platform · 15h ago The Next Platform · 15小时前

Broadcom is benefiting as hyperscalers expand their use of custom AI accelerators alongside merchant GPUs. The shift matters because more workload-specific silicon can redirect data-center spending toward ASIC design, networking and packaging, widening the AI hardware market beyond Nvidia's general-purpose accelerator franchise. 随着超大规模云厂商在通用GPU之外扩大定制AI加速器部署，Broadcom正从这一趋势中受益。更多面向特定负载的芯片会把数据中心支出引向ASIC设计、网络与封装，使AI硬件市场不再只围绕Nvidia的通用加速器体系展开。

DeepSeek V4.1 Flash Targets Lower HBM and SSD Footprints DeepSeek V4.1 Flash瞄准更低HBM与SSD占用

Wccftech · 18h ago Wccftech · 18小时前

DeepSeek's V4.1 Flash is reported to cut HBM requirements by 3.8 times and SSD requirements by eight times while posting competitive coding and cybersecurity results. If those comparisons hold across production workloads, the model could reduce the memory capacity and storage bandwidth needed per inference deployment; the performance and cost claims still need independent validation. 报道称，DeepSeek V4.1 Flash在编码与网络安全测试中保持竞争力，同时将HBM需求降低3.8倍、SSD需求降低8倍。若这些对比能在生产负载中成立，该模型可显著降低每套推理部署所需的内存容量与存储带宽；其性能和成本结论仍需独立验证。

Foundry & Manufacturing 晶圆代工与制造

TSMC Reportedly Pulls 14A Trial Production Into April 2027 报道称台积电将14A试产提前至2027年4月

Wccftech · 14h ago Wccftech · 14小时前

TSMC could begin trial production of its 1.4nm-class 14A process in Taichung in April 2027, with mass production following in the second half, according to a report cited by Wccftech. That would move the leading-edge timetable roughly a year forward and tighten the execution window after the 2nm ramp, though the schedule remains reported rather than company-confirmed. 据Wccftech引述的报道，台积电可能于2027年4月在台中启动1.4nm级14A制程试产，并在下半年进入量产。此举将最先进制程时间表提前约一年，也会压缩2nm爬坡后的执行窗口；不过该进度目前仍属媒体报道，尚未获公司确认。

Leaked Apple A20 Pro Benchmark Shows a 25% Gain 泄露测试显示Apple A20 Pro性能提升25%

Tom's Hardware · 19h ago Tom's Hardware · 19小时前

An early Geekbench 6 result reportedly puts Apple's 2nm A20 Pro about 25% ahead of its predecessor, with CPU clocks approaching 5 GHz. The result offers an initial signal for the performance available from the first wave of 2nm mobile silicon, but a leaked single benchmark does not establish sustained performance, power efficiency or shipping-bin behavior. 一项早期Geekbench 6结果显示，Apple采用2nm制程的A20 Pro较前代快约25%，CPU频率接近5 GHz。这为首批2nm移动芯片的性能提供了初步信号，但单项泄露测试仍无法说明持续性能、能效或量产分档表现。

Memory 存储

HBM Crunch Lifts Huawei and Cambricon AI-Chip Prices by Up to 50% HBM短缺令华为与寒武纪AI芯片价格最高上涨50%

BigGo News · 17h ago BigGo News · 17小时前

Huawei and Cambricon have reportedly raised AI-chip prices by as much as 50% as constrained HBM supply increases system costs. The increase shows that memory availability, not just accelerator die output, is setting the effective price and deployment pace of China's domestic AI infrastructure. 报道称，受HBM供应受限推高系统成本影响，华为与寒武纪将AI芯片价格最多上调50%。这表明决定中国国产AI基础设施实际价格与部署节奏的，不只是加速器裸片产量，内存可得性同样关键。

Equipment & Materials 设备与材料

ASML Backs 12-Inch Photomasks for High-NA EUV ASML支持High-NA EUV采用12英寸光掩模

Simply Wall St · 13h ago Simply Wall St · 13小时前

ASML is backing a move toward 12-inch photomasks for High-NA EUV lithography. A larger mask format could give advanced-node patterning more room for complex layouts and process control, but adoption would require coordinated changes across mask making, inspection, handling and fab infrastructure. ASML正支持High-NA EUV光刻转向12英寸光掩模。更大的掩模格式可为先进制程的复杂版图与工艺控制提供更多空间，但其落地需要光掩模制造、检测、搬运及晶圆厂基础设施协同升级。

Policy & Geopolitics 政策与地缘政治

Reported DOJ Review Puts Nvidia's $17 Billion Groq Deal in Focus 据报美国司法部审查聚焦Nvidia与Groq的170亿美元交易

TechStock² · 11h ago TechStock² · 11小时前

A reported US Department of Justice probe is putting Nvidia's $17 billion Groq licensing deal under antitrust scrutiny, while Nvidia shares fell 2.4%. The review matters because licensing and talent arrangements can transfer strategic accelerator technology without a conventional acquisition, testing how regulators treat consolidation in AI hardware. 据报道，美国司法部正对Nvidia价值170亿美元的Groq许可交易展开反垄断审查，Nvidia股价随之下跌2.4%。许可与人才安排可在不进行传统收购的情况下转移关键加速器技术，因此此次审查将检验监管机构如何界定AI硬件领域的集中度。

Earnings & Markets 财报与市场

Altera Prepares an IPO That Could Raise More Than $2 Billion Altera筹备IPO，募资规模或超20亿美元

Reuters · 13h ago Reuters · 13小时前

Silver Lake- and Intel-backed Altera is preparing an IPO that could raise more than $2 billion as early as 2026, Reuters reported. A listing would establish a public valuation for the standalone FPGA business and give investors a fresh read on demand for programmable logic outside the largest AI accelerators. 据Reuters报道，由Silver Lake与Intel支持的Altera正筹备IPO，最快可于2026年募资超过20亿美元。上市将为这项独立FPGA业务建立公开估值，也让投资者重新评估大型AI加速器之外的可编程逻辑需求。

TSMC Posts Record $16.26 Billion August Revenue 台积电8月营收创纪录，达162.6亿美元

TechPowerUp · 20h ago TechPowerUp · 20小时前

TSMC reported record August revenue of $16.26 billion, with sales up 53% in the cited report as AI demand filled advanced-node capacity. The result reinforces the foundry's role as the clearest upstream read-through for accelerator demand, although one month of revenue does not by itself resolve the pace of the broader semiconductor cycle. 台积电8月营收创下162.6亿美元纪录，报道所引销售额增长53%，AI需求持续占用先进制程产能。该数据再次表明台积电是观察加速器需求最清晰的上游窗口，但单月营收本身仍不足以判断整个半导体周期的速度。

Challengers & New Silicon 新势力与新硅

Ayar Labs Takes 2026 Funding to $650 Million for CPO Manufacturing Ayar Labs 2026年融资增至6.5亿美元，用于CPO制造

HPCwire · 17h ago HPCwire · 17小时前

Ayar Labs has brought its 2026 funding total to $650 million to scale manufacturing for co-packaged optics. The capital signals that optical I/O is moving from laboratory demonstrations toward the supply-chain work needed for volume systems, where packaging yield and manufacturing capacity will matter as much as link performance. Ayar Labs将2026年融资总额提高至6.5亿美元，用于扩大共封装光学（CPO）制造。资金流向表明光学I/O正从实验室演示迈向批量系统所需的供应链建设，届时封装良率与制造能力将和链路性能同样重要。

d-Matrix Adopts Nvidia NVLink Fusion for Inference Servers d-Matrix推理服务器采用Nvidia NVLink Fusion