📰 DigiTimesDigiTimesDoubao · 实盘不做把握 7/10不做原因:无增量信息
北京时间·🕒 发布 10:48:50·⟳ 抓取 11:02:55 +14m05s·⚡ 决策 10-13 12:21:26 +1h18m·14.1 天前
緯創、緯穎於OCP 2026展示全液冷Rubin NVL72試產平台，屬產品展示無新增訂單
這是公司在OCP峰會的常規產品展示新聞稿，自稱Rubin NVL72首批試產夥伴僅證明供給/組裝能力，沒有新訂單、客戶、數量、份額或財務增量；且PR Newswire官方稿10-12已首發，DigiTimes為次日跟進，缺乏新預期差。A股/台股映射（液冷、CPO）亦無由此事件產生的新增訂單錨點。
📄 原文DigiTimes↗ 原文
Wistron, Wiwynn spotlight liquid-cooled AI compute, storage at OCP
Wistron and Wiwynn are showcasing next-generation AI compute and storage systems for AI factories at the 2026 Open Comp…
不做 緯穎 Wiwynn（6669.TW）
OCP產品展示屬能力/行銷事件，pilot build身份非新增訂單，官方稿前一日已首發，缺乏預期差與可量化營收增量；行情工具暫不可用亦無法核驗午盤開盤價，不構成入場edge
不做 緯創 Wistron（3231.TW）
同為展示性軟催化，大市值ODM邊際影響更弱，無訂單/份額增量，不做
展开详情(成交明细 / 研究档案 / 原文全文 / 决策记录链)
研究档案
结论摘要
緯創/緯穎於OCP 2026展示全液冷Rubin NVL72試產平台，屬能力展示與行銷事件，無新增訂單或財務信息，官方稿已於前一日首發，預期差不足，不具可交易edge，故回避。
消息传播链
- 緯創/緯穎OCP展示官方新聞稿首發，披露Rubin NVL72首批pilot build身份，無訂單金額
- 緯穎官網同日發布OCP展示頁面，印證內容
- DigiTimes次日跟進報導，無超越官方稿的新增事實
关键理由
- pilot build身份驗證兩家進入Rubin首批供應鏈，對2027 Rubin放量年的中長期份額是正面能力背書
- 但展示本身不構成bookings/backlog增量，且官方新聞稿已提前一天發布，短線缺乏重定价催化
主要风险
- 若會後有分析師給出上調份額或具體Rubin訂單量的跟進報告，可能補充增量信息，需重評
- Rubin出貨預測為模型推導、區間寬（2027約4-6萬櫃），兩家實際可獲份額不確定
反向证据
- 做多前查反向 · 来源 web · 影响 none
kb_search服務連接拒絕不可用；Web未見與本展示掛鉤的新增訂單或分析師上調，亦無明確反向利空，核心問題是無增量而非有利空
检索:Wiwynn Wistron Rubin 訂單 砍單 下修 2026
先例如何影响了这次决策
对本次决策的影响:先例與增量事實均不足，加劇了不採用小倉试探的傾向，最終AVOID
方向:未測量；ODM在GTC/OCP前後的產品展示通常為已知能力確認，缺乏訂單增量時短線超額收益不穩定
涨幅量级:無高相似、乾淨錨點先例可校準幅度
形态:無足夠樣本判斷見光死模式
工具执行结果 · 成功 5 · 降级 0 · 失败 3 · 取消 0
- search_processed_events 成功 stock_tools HTTP · 10 ms
{"result": "search_processed_events:agent query=`wistron wiwynn ocp 2026 vera rubin nvl72 pilot build partner | wiwynn wistron rubin rack cpo ethernet showcase | 緯創 緯穎 vera rubin nvl72 試產 液冷 ocp`；结构化索引覆盖完整 72 小时；旧记录兼容扫描近 0.0/72 小时、0 条本臂决策(完整)，返回 0 条候选。未命中只表示这些 agent query 表达式没有召回候选，不能据此证明消息首发或不是重复；按 - web_search 成功 doubao_search · 1,501 ms
豆包搜索结果（原始查询:Wistron Wiwynn OCP 2026 Vera Rubin NVL72 pilot build partner all-liquid-cooled；API SearchType:web） - kb_search 失败 stock_tools HTTP · 57 ms
kb_search 服务暂不可用；这不是一条知识库命中。 - query_modeling_kb 成功 stock_tools HTTP · 75 ms
{"result": "modeling_kb 命中 10 条 (库快照 2026-07-27 · AI 算力链 19 页 · 其中 ⟨预测⟩ 7 条) —— 这是**定期重生成的快照**,不是活数据;比快照更晚的事实以你手上的信号/kb_search 为准。\n\n• [B] Rubin GPU · 逻辑 wafer-bound R200 供给 · CY2026E = N3 有弹性(5%手机转产=+0.1M, 25%=+0.7M Rubin; 非绑定瓶颈) M 封装/年 ⟨预测⟩（模型推导/无直接引用）\n 依据: 将 2026 年智能机 N3 总投片的 5%（43.7 万片的 5%） - normalize_ticker 成功 stock_tools HTTP · 5 ms
未归一到唯一代码;请用 stock_quote 的 candidates 消歧,或直接给代码。 - normalize_ticker 成功 stock_tools HTTP · 4 ms
未归一到唯一代码;请用 stock_quote 的 candidates 消歧,或直接给代码。 - stock_quote 失败 stock_tools HTTP · 11 ms
行情服务暂不可用；本次无法核验该标的价格和可执行条件。请基于其余证据继续判断并明确标注行情降级；不得仅因工具故障自动给出 AVOID 或 ENTER，也不得编造价位。 - stock_quote 失败 stock_tools HTTP · 17 ms
行情服务暂不可用；本次无法核验该标的价格和可执行条件。请基于其余证据继续判断并明确标注行情降级；不得仅因工具故障自动给出 AVOID 或 ENTER，也不得编造价位。
原文全文
展开原文 · 2,335 字
Wistron and Wiwynn are showcasing next-generation AI compute and storage systems for AI factories at the 2026 Open Compute Project (OCP) Global Summit in San Jose, California, with all-liquid-cooled platforms and Nvidia's latest rack-scale systems taking center stage.
Under the theme "Create AI Beyond Limits," the companies are highlighting "Powering the AI Factory: Next-Generation AI Compute and Storage Systems," focusing on the compute and storage infrastructure required by AI factories.
One of the key highlights is the companies' readiness for Nvidia's next-generation rack-scale platforms. The companies said they are among the first pilot build partners for the Nvidia Vera Rubin NVL72 platform, which uses an all-liquid-cooled design optimized for AI training and inference applications.
Beyond GPU compute platforms, the exhibits include Vera CPU racks, contextual memory storage platforms, and Ethernet solutions integrating co-packaged optics (CPO). The companies are also presenting the HGX Rubin NVL8 platform for enterprise AI factories.
The showcased products also offer a glimpse into several emerging trends in AI infrastructure.
First, as agentic AI and reinforcement learning applications expand, CPU racks are being presented as a standalone category. Built on the Nvidia MGX architecture and integrated with Nvidia Spectrum-X Ethernet networking, a single rack can support more than 22,500 concurrent execution environments, completing workloads twice as fast as traditional CPU-based infrastructure, according to the companies.
Second, growing inference demand is driving changes in storage architecture. Nvidia's CMX platform is designed to expand the effective capacity of GPU key-value (KV) caches, enabling high-bandwidth sharing of contextual data within an AI pod while freeing up compute resources and reducing power consumption.
Third, the networking side is adopting 200G co-packaged optics to support 800GbE, combined with direct liquid cooling to address network congestion and improve data center efficiency.
Through the OCP platform, Wistron and Wiwynn aim to present an integrated AI factory system spanning compute, storage, and networking. Liquid-cooling designs run across multiple platforms, underscoring the rising demands for energy efficiency and scalability in next-generation AI racks.
Under the theme "Create AI Beyond Limits," the companies are highlighting "Powering the AI Factory: Next-Generation AI Compute and Storage Systems," focusing on the compute and storage infrastructure required by AI factories.
One of the key highlights is the companies' readiness for Nvidia's next-generation rack-scale platforms. The companies said they are among the first pilot build partners for the Nvidia Vera Rubin NVL72 platform, which uses an all-liquid-cooled design optimized for AI training and inference applications.
Beyond GPU compute platforms, the exhibits include Vera CPU racks, contextual memory storage platforms, and Ethernet solutions integrating co-packaged optics (CPO). The companies are also presenting the HGX Rubin NVL8 platform for enterprise AI factories.
The showcased products also offer a glimpse into several emerging trends in AI infrastructure.
First, as agentic AI and reinforcement learning applications expand, CPU racks are being presented as a standalone category. Built on the Nvidia MGX architecture and integrated with Nvidia Spectrum-X Ethernet networking, a single rack can support more than 22,500 concurrent execution environments, completing workloads twice as fast as traditional CPU-based infrastructure, according to the companies.
Second, growing inference demand is driving changes in storage architecture. Nvidia's CMX platform is designed to expand the effective capacity of GPU key-value (KV) caches, enabling high-bandwidth sharing of contextual data within an AI pod while freeing up compute resources and reducing power consumption.
Third, the networking side is adopting 200G co-packaged optics to support 800GbE, combined with direct liquid cooling to address network congestion and improve data center efficiency.
Through the OCP platform, Wistron and Wiwynn aim to present an integrated AI factory system spanning compute, storage, and networking. Liquid-cooling designs run across multiple platforms, underscoring the rising demands for energy efficiency and scalability in next-generation AI racks.
决策记录链 · 只追加
每次领取产生一条记录,重新决策追加一条并以 supersedes 指向被替代的记录(ADR 0014)。
065cb15c9f· 2026-10-13 12:21 · doubao / doubao-seed-evolving-latest-version · 策略包2226518bfdee当前