市场对「无上限投入」的耐心下降但未改变资本开闸：Microsoft 宣布 capex 后一周股价下跌 10%；Oracle 表示今年将借款最高 500 亿美元投入 AI 基建，且股价自去年 9 月高点回落 50%，反映债务与回收期疑虑。回报时间仍不确定，但晶片与算力供应链已先收割 2026 成长，且成长叙事从云端训练延伸到推论与终端：edge AI 被视为 2026 的拐点，推论下沉到手机与 PC 可降低云端成本，同时云端仍承担重负载；Qualcomm、Lenovo 等预期 2026 年 AI 手机与 AI PC 用例与出货动能更强，涵盖创作、效率、文件自动化、搜寻与影音编辑等场景。
「physical AI」把推论需求带入车、工厂与机器人，扩大晶片消耗面：Synopsys 把设计流程与实体世界模拟绑定，并以 San Francisco 的 Waymo 与 Las Vegas 的 Zoox robotaxi 作为商业化例证；Arm 亦把机器人、制造与汽车视为需求来源。中国则在 2026–2030 五年规划中把半导体自给置顶，带动国产 GPU 相关公司资本市场热度：Biren Technology、Moore Threads、Meta X 近期上市且首日股价大涨，Tencent 支持的 Suiyuan Technology（Enflame）准备上市。供给侧成为 2026 的硬限制：ASUS 的 Jackie Hsu 指出泡沫尚未到来，主因供需未平衡；Samsung 与 SK Hynix 认为记忆体供给吃紧将延续至 2027，Samsung 被迫选择性签约并优先供应伺服器 DRAM，SK Hynix 表示其 2026 年产能已售罄。
Big Tech shows no sign of slowing AI capex in 2026: Meta says it will nearly double capex this year to as much as $135 billion; Tesla plans to spend more than $20 billion, up from $8.5 billion in 2025; Microsoft posted $37.5 billion of quarterly capex in the three months ending December, up 66% year over year, with Amazon and Google also pressing ahead. China’s tech leaders are expanding in parallel: ByteDance reportedly has preliminary 2026 capex plans of 160 billion yuan (about $23 billion) for AI infrastructure; Alibaba is weighing lifting its three‑year AI infrastructure and cloud plan from 380 billion yuan to as much as 480 billion yuan; Tencent logged 59.7 billion yuan of capex in the first three quarters of 2025 and is accelerating product and talent moves, including the overseas chatbot Dola exceeding 10 million active users by end‑2025 and hiring former OpenAI core researcher Yao Shunyu.
Investor patience with open‑ended spending is thinning without shutting the capex spigot: Microsoft shares fell 10% in the week after its capex announcement; Oracle says it will borrow up to $50 billion this year for AI infrastructure and its shares are down 50% from a peak last September, highlighting debt and payback concerns. Returns remain uncertain, but chip and compute suppliers already capture 2026 growth, and the growth story is broadening from cloud training to inference and endpoints: edge AI is framed as a 2026 inflection point, with more inference running on phones and PCs to reduce cloud cost while the cloud still handles heavy workloads. Qualcomm and Lenovo also expect stronger 2026 momentum for AI smartphones and AI PCs, spanning creativity, productivity, document automation, search, and video editing use cases.
Physical AI pushes inference demand into cars, factories, and robots, widening the chip consumption base: Synopsys ties design flows to real‑world simulation and points to Waymo in San Francisco and Zoox in Las Vegas as commercialization evidence; Arm likewise targets robotics, manufacturing, and automotive demand. China’s 2026–2030 plan elevates semiconductor self‑reliance, boosting capital‑market interest in domestic GPU players: Biren Technology, Moore Threads, and Meta X have recently listed with first‑day share surges, and Tencent‑backed Suiyuan Technology (Enflame) is preparing to list. Supply becomes the binding constraint in 2026: ASUS’s Jackie Hsu says the bubble has not arrived because demand and supply are still imbalanced; Samsung and SK Hynix say the memory crunch will last until 2027, Samsung is selectively contracting and prioritizing server DRAM, and SK Hynix says its 2026 capacity is already sold out.