Computex Opens; Nvidia Lands on the PC; US Closes the Loophole Computex开幕：英伟达登陆PC，美国关闭芯片豁免漏洞

Nvidia's RTX Spark puts Grace Blackwell silicon into Windows PCs, Supermicro debuts a Vera Rubin NVL72 rack with a new coolant, Intel ships Xeon 6+ Clearwater Forest — all on the same Computex Day 1 the US closes the loophole that let Chinese-owned foreign subsidiaries buy regulated AI chips. Computex第一天：英伟达RTX Spark把Grace Blackwell带进Windows PC，美超微发布配新型冷却液的Vera Rubin NVL72机架，英特尔出货Xeon 6+「Clearwater Forest」——同日美国关闭了中资境外子公司绕道购买受控AI芯片的漏洞。

AI & Accelerators AI与加速器

Nvidia launches RTX Spark — Grace Blackwell Arm SoC for Windows PCs 英伟达发布RTX Spark：面向Windows PC的Grace Blackwell Arm SoC

ServeTheHome · 2h ago ServeTheHome · 2小时前

At Computex 2026 Day 0, Nvidia launched RTX Spark — a Grace Blackwell-derived Arm SoC for Windows PCs, with first notebook partners disclosed and developer-workload benchmarks claiming a 54% lead over Apple's M5 on selected workloads. The strategic move matters more than the silicon: this is Nvidia's first deliberate push into the general-purpose Windows PC processor market, exactly the segment Qualcomm has been trying to establish with Snapdragon X / C platforms. RTX Spark forces AMD and Intel to defend their CPU sockets against a vendor that already owns the discrete-GPU pipeline. Intel's response (covered separately below): 'a healthy dose of paranoia.' Computex 2026 Day 0英伟达发布RTX Spark——基于Grace Blackwell的Windows PC用Arm SoC，首批笔记本合作伙伴同时披露，开发者负载基准测试声称在部分场景下较苹果M5领先54%。意义在战略而非硅本身：这是英伟达首次有意识地切入通用Windows PC处理器市场，正是高通过去几年用Snapdragon X／C平台试图建立的地盘。RTX Spark迫使AMD和英特尔在CPU socket上对抗一家已经掌握独立GPU管线的厂商。英特尔的回应（下文单列）：「健康剂量的偏执」。

Supermicro shows Vera Rubin NVL72 rack with new coolant — 1,000× the thermal headroom claim 美超微展示配新型冷却液的Vera Rubin NVL72机架——号称散热裕度提升1,000倍

Tom's Hardware · 2h ago Tom's Hardware · 2小时前

Supermicro debuted a Vera Rubin NVL72 reference rack at Computex, with a new immersion-style coolant the company claims delivers roughly 1,000× the thermal headroom of conventional single-phase liquids — important because the Vera Rubin parts' TDPs were already pushing established water-cooled solutions to their limit. The launch is also notable as Supermicro's first major product event after the co-founder's arrest last weekend in the $2.5B Nvidia GPU smuggling case: the company is signaling continuity, and Nvidia is signaling that the OEM relationship continues despite the headline risk. 美超微在Computex上展示Vera Rubin NVL72参考机架，采用一种新型浸没式冷却液，公司宣称散热裕度较传统单相液冷高出约1,000倍——重要之处在于Vera Rubin系列TDP已经把现有水冷方案推到极限。这场发布也是上周末该公司联合创始人因25亿美元英伟达GPU走私案被捕之后的首次重大产品活动：公司在传递「业务连续」的信号，英伟达则在传递「尽管头条风险存在，OEM合作仍在继续」的信号。

Foundry & Manufacturing 晶圆代工与制造

Intel ships Xeon 6+ Clearwater Forest — Intel 18A's first major commercial volume 英特尔出货Xeon 6+「Clearwater Forest」——Intel 18A的首次重大商用量产

ServeTheHome · 1h ago ServeTheHome · 1小时前

Intel formally launched Xeon 6+ Clearwater Forest, its all-E-core server CPU and the first high-volume part on Intel 18A with backside power delivery (PowerVia). ServeTheHome's launch coverage notes the part lands with usable hyperscaler customer references rather than a vendor-only narrative, which is the meaningful difference from past Intel server-CPU launches. The combination of an actual 18A ramp and Clearwater Forest going GA puts Intel Foundry's 2H26 narrative on its first real footing in years — and gives the Apple-Intel rapprochement story (covered yesterday) a credible silicon backdrop. 英特尔正式发布Xeon 6+「Clearwater Forest」——其全E-core服务器CPU，也是Intel 18A工艺（带PowerVia背面供电）的首款高量级产品。ServeTheHome的发布报道指出，该产品上市时已附带可用的超大规模客户参考案例，而非以往「只剩供应商一家叙事」的局面——这是与英特尔此前服务器CPU发布相比最实质的区别。真实的18A爬坡叠加Clearwater Forest正式上市，让英特尔代工2026下半年的叙事多年来首次站在真实基础上；昨日报道的苹果-英特尔回暖故事，也由此获得可信的硅片背景。

Nvidia and TSMC deploy AI in chip fabs to improve yields 英伟达与台积电将AI部署到晶圆厂，用以提升良率

qz.com (via Google News) · 58m ago qz.com (via Google News) · 58分钟前

Quartz reports a joint Nvidia-TSMC initiative deploying Nvidia accelerator-trained models directly into TSMC's fab process-control loop, aimed at yield improvement on advanced nodes (N2 and below). The framing matters: it's not 'AI assists humans planning experiments'; it's AI inference embedded in the production-control feedback loop. For the broader semi industry this is the more interesting AI-in-fab data point because the customer is also the silicon supplier — the alignment of incentives is tight, and the yield-curve evidence (if it materializes) will be hard to dismiss as marketing. Quartz报道英伟达与台积电的一项联合计划：将英伟达加速器训练的模型直接部署到台积电晶圆厂的工艺控制回路，目标提升先进节点（N2及以下）的良率。叙事关键在于：这不是「AI辅助人类设计实验」，而是把AI推理嵌入产线控制反馈回路。对更广义半导体行业而言，这是关于「AI进入晶圆厂」更有意思的数据点——因为客户同时是硅片的供应商，激励高度对齐；若良率曲线证据真的出现，将很难被当作营销叙事被打发。

Memory 存储

Toxic gas fire at SK hynix Cheongju memory plant — 7 hospitalized SK海力士清州存储厂发生有毒气体起火事故——7人送医

Tom's Hardware · 46m ago Tom's Hardware · 46分钟前

An incident at SK hynix's Cheongju 4th-campus advanced-memory plant sent seven workers to the hospital after a toxic-gas fire today. The early reporting frames it as a safety-systems failure rather than a production-disrupting event, but the plant houses HBM-relevant capacity at a moment when supply discipline is already tight. The market read will hinge on whether plant output is meaningfully interrupted — Cheongju's 4th campus is one of the production lines feeding the HBM3E and HBM4 ramps, so even a several-day production hit is materially visible at the contract-pricing level. SK海力士清州第四园区先进存储工厂今日发生有毒气体起火事故，7名工人送医。早期报道将其定性为安全系统失效而非产线中断事件，但该厂区在供给纪律已经紧张的当下承载着HBM相关产能。市场解读将取决于产线是否实质中断——清州第四园区是支撑HBM3E与HBM4爬坡的产线之一，即便几天的产量损失，也会在合约定价层面被实质看到。

DRAM revenue soars 81% to $97B as AI demand and supply constraints push prices higher DRAM收入年增81%至970亿美元——AI需求与供应紧张共同推升价格

TelecomLead (via Google News) · 4h ago TelecomLead (via Google News) · 4小时前

Industry data shows DRAM revenue jumped roughly 81% year-over-year to about $97B, with the bulk of the gain coming from contract-price reset combined with HBM mix shift rather than unit growth. Three-vendor concentration (Samsung, SK hynix, Micron) plus the P4-line HBM conversion (covered last week) plus today's Cheongju incident together translate to a 2H26 contract negotiation cycle where memory makers hold visible leverage. Buyers are starting to sign multi-year LTAs with price caps that historically would have been impossible to negotiate from the supplier side. 行业数据显示DRAM年度收入同比上涨约81%至约970亿美元，增量主要来自合约价重置叠加HBM组合切换，而非出货量增长。三家厂商集中（三星、SK海力士、美光）+ P4厂HBM产能转换（上周报道）+ 今日清州事故，叠加成2026下半年合约谈判周期——存储厂商手中议价权可见。买方开始签下带价格上限的多年LTA，而这类条款历史上对供应方而言根本谈不下来。

Policy & Geopolitics 政策与地缘

US closes loophole that let Chinese-owned foreign subsidiaries buy regulated AI chips 美国关闭中资境外子公司绕道购买受控AI芯片的漏洞

Tom's Hardware · 3h ago Tom's Hardware · 3小时前

The US Commerce Department finalized a rule closing the long-standing loophole that allowed Chinese-owned subsidiaries located outside China — e.g., Hong Kong, Singapore, or Cayman-domiciled entities controlled from the mainland — to buy regulated US-origin AI chips. The change brings the export-control perimeter in line with end-use-control intent rather than nameplate-shipment destination. Practical effect: every reseller, OEM, and channel partner in Southeast Asia must now reassess know-your-customer obligations on Chinese-linked buyers, and the volume of grey-market diversion routed through Asian intermediaries should compress meaningfully. 美国商务部最终敲定一项规则，关闭长期存在的漏洞：此前中资境外子公司——如设在香港、新加坡或开曼但由大陆控制的实体——可以购买受控的美国原产AI芯片。新规则把出口管制的边界与「最终用途控制」意图对齐，而非仅以「出货目的地名义」为标准。实际效果：东南亚每一家分销商、OEM与渠道伙伴，都必须重新评估针对中资关联买家的KYC义务；通过亚洲中间方的灰色市场转运总量将出现明显压缩。

Intel + 3DGS announce $3.3B India semiconductor facility in Odisha 英特尔与3DGS宣布在印度奥迪沙建设33亿美元半导体设施

Yahoo Finance (via Google News) · 3h ago Yahoo Finance (via Google News) · 3小时前

Intel and Indian partner 3DGS announced a $3.3B semiconductor facility in Odisha — likely an OSAT/packaging line rather than wafer fab, but materially relevant given India's parallel Tata-ASML / Tata-Gujarat fab build. The deal supports Intel Foundry's broader pitch that 'foundry + packaging at non-TSMC geography' is a real offering, not a press release. Combined with yesterday's Apple-Intel volume signal, Intel's 2H26 capacity narrative has more concrete data points behind it than at any time since the foundry strategy was announced. 英特尔与印度本土合作伙伴3DGS宣布在奥迪沙建设33亿美元半导体设施——更可能是OSAT/封装产线而非晶圆厂，但在印度并行推进的Tata-ASML/Tata古吉拉特晶圆厂建设的背景下意义重大。该交易支撑英特尔代工的更大叙事：「非台积电地理位置的代工+封装」是真实供货能力、而非新闻稿。叠加昨日苹果-英特尔的出货信号，英特尔2026下半年产能叙事拥有的具体数据点，比代工战略发布以来任何时候都多。

Equipment & Materials 设备与材料

Tessalia plans €250M advanced packaging site in France Tessalia计划在法国投资2.5亿欧元建设先进封装基地

gasworld (via Google News) · 1h ago gasworld (via Google News) · 1小时前

French startup Tessalia disclosed plans for a €250M advanced semiconductor packaging facility in France — a small line by Asian-OSAT standards but materially relevant for the European AI Act / Chips Act regulatory environment that increasingly conditions sovereign-relevant compute on EU-located packaging steps. Tessalia's pitch is not to compete with ASE/Amkor on cost but to position as a regulatory-compliant supplier for European AI accelerator startups. Watch for whether any of the Mistral / Cerebras-Europe / VSORA peer set commits to a Tessalia tape-out — that's the validation step that turns the announcement into supply. 法国创业公司Tessalia披露在法国投资2.5亿欧元建设先进半导体封装基地——按亚洲OSAT标准属于小型产线，但在欧洲AI法案／Chips Act监管环境下意义重大——该监管环境正越来越多地要求与主权相关的算力在欧盟境内完成封装。Tessalia的卖点不是在成本上挑战ASE/Amkor，而是定位为欧洲AI加速器初创的合规供应商。后续关注：Mistral／Cerebras-Europe／VSORA同类公司中，是否有人承诺在Tessalia流片——那将是把这条新闻从「公告」变为「真实供应」的验证步骤。

Invisix raises €20M for soft X-ray metrology — AI-era chip-manufacturing inspection Invisix完成2,000万欧元融资——面向AI时代芯片制造的软X射线计量

Pulse 2.0 (via Google News) · 3h ago Pulse 2.0 (via Google News) · 3小时前

European metrology startup Invisix raised €20M in seed funding to advance soft X-ray-based inspection technology for AI-era chip manufacturing — explicitly positioned for the High-NA EUV node window where conventional CD-SEM and OCD optical metrology are running into resolution and throughput limits. The same arXiv physics.app-ph community publishing the Wideband Balanced Photodetector work earlier this month produces the underlying device-physics chain that makes soft X-ray metrology newly viable. Watch for whether ASML or Lam Research takes a strategic stake — that's the route from seed funding to actual tool integration. 欧洲计量初创公司Invisix完成2,000万欧元种子轮融资，推进面向AI时代芯片制造的软X射线检测技术——明确瞄准High-NA EUV节点窗口，该窗口下传统CD-SEM与OCD光学计量已遭遇分辨率与吞吐量瓶颈。本月早些时候发布Wideband Balanced Photodetector工作的arXiv physics.app-ph社区，正是支撑软X射线计量从可能走向可行的底层器件物理链。后续关注ASML或Lam Research是否会战略入股——那才是从种子轮到真正工具集成的路径。

Roadmaps & Markets 路线图与市场

Intel admits 'a healthy dose of paranoia' over Nvidia entering PCs — RTX Spark is 'great' 英特尔承认对英伟达进军PC「有一份健康的偏执」——称RTX Spark「不错」

Tom's Hardware · 4h ago Tom's Hardware · 4小时前

Asked about RTX Spark, Intel's executive team publicly admitted 'a healthy dose of paranoia' and called the Nvidia product 'great' — a notable departure from past Intel responses to Arm-on-Windows competitors. The framing is strategically useful for Intel: acknowledging Nvidia as a genuine threat domestically helps justify the 18A / Xeon 6+ ramp investments and the Apple foundry win as part of a coherent counter-narrative. It also subtly contrasts with Qualcomm's current Snapdragon C selloff (covered yesterday): Intel is treating the PC-CPU competitive dynamic seriously while Qualcomm is taking a positioning hit. 在被问及RTX Spark时，英特尔高管层公开承认「有一份健康的偏执」，并称这款英伟达产品「不错」——这与英特尔此前对Arm-on-Windows竞争者的回应风格明显不同。这种表态对英特尔战略上有用：把英伟达定性为真实威胁，有助于把18A／Xeon 6+爬坡投入与苹果代工赢单纳入一致的反击叙事。同时也微妙对照高通当下的Snapdragon C抛售（昨日报道）：英特尔正认真对待PC CPU竞争动态，而高通在仓位上受挫。

Computex 2026 opens — every Day 1 update from Taipei Computex 2026开幕——台北Day 1全更新

Tom's Hardware · 4h ago Tom's Hardware · 4小时前

Computex 2026 opened today in Taipei with Tom's Hardware running rolling coverage. Beyond Nvidia RTX Spark and the Supermicro Vera Rubin rack, day 1 highlights include Intel's Crescent Island AI GPU launch plans, AMD's EXPO Ultra Low Latency DDR5 (13% uplift), 256GB dual-channel CUDIMM (Origin Code) hitting mass-market, and Asus rolling out ROG Edition 20 SKUs. The aggregate signal: Computex 2026 is shaping up as the year the PC platform finally absorbs the AI-accelerator economics that have until now lived on rack-scale infrastructure. Computex 2026今日在台北开幕，Tom's Hardware开启滚动报道。除了英伟达RTX Spark与美超微Vera Rubin机架之外，Day 1亮点还包括英特尔Crescent Island AI GPU发布计划、AMD EXPO Ultra Low Latency DDR5（13%提升）、256GB双通道CUDIMM（Origin Code）走向大众市场，以及华硕ROG Edition 20系列SKU上市。综合信号：Computex 2026正在成为「PC平台终于吸纳了此前只存在于机架级基础设施上的AI加速器经济学」的一年。