文章

AI Token支出暴跌 · PCB瓶颈致英伟达延期

#992 · 2026-07-06 · 21ZHAO Blog
AI Token支出暴跌PCB瓶颈致英伟达延期AILLM架构算力
Reading Path / ARTICLE 先抓主张，再转成行动 #992 · 21ZHAO Blog · 读完进入产品或下一篇

AI Token支出暴跌 · PCB瓶颈致英伟达延期

一、 权威必看

EN: The financial markets are currently witnessing a significant shift in the AI investment narrative, marked by a nearly 20% drop in the Silicon Data LLM Token Expenditure Index. This metric, which tracks the price paid per million tokens across the market, has fallen sharply from its May highs, suggesting that the era of unchecked capital expenditure and vendor pricing power may be coming to an end. Industry investors like Louis Navellier point out that as costs remain high, users are forced to curb their AI usage, indicating a potential disconnect between the massive capital inflows into the sector and the actual return on investment expected by clients.

中： 金融市场正在见证AI投资叙事的显著转变，硅基数据LLM词元支出指数近20%的跌幅是这一变化的核心标志。该指标追踪全市场每百万词元的支付价格，自5月高点以来大幅回落，表明无节制的资本支出和厂商定价权的时代可能即将终结。资深投资者Louis Navellier指出，由于成本依然高昂，用户被迫限制AI使用量，这揭示了巨额资金流入与客户端预期的实际投资回报之间存在潜在脱节。这一指数的疲软并不完全等同于服务变便宜，而是价格与使用量双重维度的综合反映，预示着AI行业正从“烧钱扩张”向“成本敏感”阶段过渡。

二、 深度与多元

EN: In a stark contrast to the high-tech financial turbulence, local governance in Qinzhou is adapting to extreme weather events with pragmatic flexibility. Affected by the circulation of Typhoon Maysak, which brought maximum rainfall of 604.7mm to some townships, the Qinzhou Flood Control and Drought Relief Command has issued a directive for flexible working arrangements. This policy exempts only essential roles such as flood control, emergency rescue, medical services, and utility maintenance, aiming to reduce employee commuting risks during heavy rains.

中： 与高科技金融市场的动荡形成鲜明对比，钦州地方政府正以务实的灵活性应对极端天气事件。受台风“美莎克”环流影响，部分乡镇最大降雨量达604.7毫米，钦州市防汛抗旱指挥部发布指令，要求采取弹性工作制。该政策仅豁免防汛值守、应急救援、医疗及水电保供等刚需岗位，旨在降低员工在强降雨期间的出行风险。这一举措不仅体现了基层行政对突发公共事件的快速响应能力，也反映了社会管理从刚性执行向人性化、精细化治理的深层逻辑转变，将公共安全置于经济活动之上。

三、 科技与财经

EN: The IPO market this week (July 6-10) is highlighting the resilience of specific biotech and manufacturing sectors. Two new stocks are scheduled for subscription: Longxin Intelligent Equipment (920117), a national specialized and sophisticated ’little giant’ enterprise in smart manufacturing, and Tino Medicines (688806), a leader in innovative blood product alternative therapies. Tino Medicines’ core product, Steptocidumab Injection, approved in China in February 2025, is the world’s first-in-class recombinant anti-tetanus toxin monoclonal antibody, marking a significant breakthrough in preventive medicine.

中： 本周（7月6日至10日）的新股申购市场凸显了特定生物技术和制造领域的韧性。两只新股安排申购：龙鑫智能（920117）是智能制造领域的国家级专精特新“小巨人”企业，提供智能化工厂整体解决方案；泰诺麦博（688806）则是创新血液制品替代疗法的龙头。其核心产品斯泰度塔单抗注射液于2025年2月在中国获批上市，为全球同类首创的重组抗破伤风毒素单克隆抗体药物，在预防医学领域具有革新意义。这表明资本市场在AI热潮之外，仍对具备硬核技术壁垒和临床价值的创新药及高端制造给予高度关注，细分领域的龙头效应正在IPO窗口期得到强化。

四、 国际视野

EN: A new open-source 3D engine named ThreeJSON has emerged on V2EX, leveraging the MIT license to offer a JSON-driven approach to 3D scene construction based on Three.js. Its philosophy is condensed into four principles: ‘Scene as Data,’ ‘Runtime as Engine,’ ‘Domain as Extension,’ and ‘Agent as Operator.’ This framework allows developers to describe the world via JSON text while AI agents handle creation and modification, with ThreeJSON handling rendering and execution. It represents a shift towards data-centric 3D development where AI acts as the primary operator.

中： V2EX社区涌现出一个名为ThreeJSON的开源3D引擎，基于MIT协议和Three.js构建，采用JSON驱动的方式描述3D场景。其核心理念浓缩为四句话：“场景即数据”、“运行时即引擎”、“业务域即扩展”、“Agent即操作者”。该框架允许开发者通过JSON文本描述世界，由AI代理负责创建和修改，而ThreeJSON负责渲染和运行。这代表了3D开发向数据-centric模式的转变，AI成为主要操作者。对于前端和图形学开发者而言，这种将场景定义与逻辑执行分离的架构，为AI生成3D内容提供了标准化的数据接口，降低了复杂场景构建的技术门槛。

五、 青年与生活

EN: The semiconductor supply chain faces a critical bottleneck as reported by SemiAnalysis, which suggests that Nvidia’s Kyber NVL144 rack architecture may be delayed by over 12 months, pushing its launch to 2028. The primary cause is identified as significant manufacturing challenges with the midplane PCB (Orthogonal Backplane), which requires a 78-layer ultra-high-multilayer design to replace copper cables and achieve higher compute density within a single rack. This delay highlights the physical limitations of hardware manufacturing in supporting rapid AI scaling.

中： 半导体供应链面临关键瓶颈，研究机构SemiAnalysis指出英伟达Kyber NVL144机架架构可能推迟超过12个月，发布时间延至2028年。主要原因被归结为PCB中板（正交背板）制造工艺的重大挑战。该部件需采用78层的超高多层设计，以实现计算托盘与交换托盘间的90度垂直互联，替代铜缆并提升单机柜算力集成度。这一延迟凸显了在AI算力快速扩张的背景下，底层硬件制造的物理极限和工艺复杂性已成为制约技术落地的关键因素，硬件迭代速度可能无法完全匹配软件算法的演进节奏。

【21ZHAO 综合判断】

EN: The convergence of these five dimensions reveals a macro-trend: the AI industry is hitting both economic and physical ceilings. The drop in token expenditure suggests that software monetization is struggling to keep pace with hardware costs, while the Nvidia PCB delay proves that hardware scaling is encountering severe manufacturing bottlenecks. For developers and investors, this signals a shift from ‘growth at all costs’ to ’efficiency and feasibility.’

  • For Developers: Focus on optimizing AI inference costs and exploring data-centric 3D frameworks like ThreeJSON to reduce development complexity in spatial computing.
  • For Investors: Diversify beyond pure AI hype into resilient sectors like innovative biotech (e.g., Tino Medicines) and monitor hardware supply chain constraints closely, as delays in key components like PCBs can significantly impact tech valuations.

中： 综合这五个维度，揭示了一个宏观趋势：AI行业正同时遭遇经济回报和物理极限的双重天花板。词元支出下降表明软件变现难以跟上硬件成本，而英伟达PCB延迟证明硬件扩展遇到了严重的制造瓶颈。对于开发者和投资者而言，这意味着从“不惜代价增长”向“效率与可行性”的转变。

  • 对开发者建议： 重点关注优化AI推理成本，并探索如ThreeJSON等数据-centric的3D框架，以降低空间计算领域的开发复杂度。
  • 对投资者启示： 在纯AI炒作之外，将资产配置多元化至具有韧性的领域（如创新药泰诺麦博），并密切监控硬件供应链约束，因为PCB等关键组件的延迟可能显著影响科技股估值。

参考来源