# 产业日报 2026-10-21 来源范围:`/workspace/intel/archive/2026-10-21/`(industry/ 主;vmware/redhat/zstack 仅旁证)。覆盖窗口:2026-10-20(周二)早报之后 → 2026-10-21(周三)早。上一成功早报 **2026-10-20**。抓取统计:Saved 277 / skipped 29 / errors 42;blocked_hosts:community.hpe.com、docs.redhat.com、e.huawei.com、support.huawei.com、www.globenewswire.com、www.reddit.com、www.sangfor.com(与 10-20 相同)。 **产业源状态:** industry/ 相对全部历史归档**真正新增 URL 6 条**:Blocks&Files 10-20 日戳 3 篇(Meta Prometheus / DeepL×VAST / Testing AI agent accuracy)、ServeTheHome 2 篇(BlueField-4 STX CMX / HPE Spaceborne-4)、Virtualization Review 10-20 1 篇(Semperis)。NextPlatform / The Register **无新 URL**(Register 2026-02~08 旧稿与 NP 10-15 ZeroRISC 仅字节级重抓)。中文聚合首页(至顶网 / Solidot)有新条目,仅作**媒体快讯旁证**。数字均标「媒体报道」「厂商通稿」「调研」。 ## 摘要 今日产业信号收成六句话: 1. **中 NEW · Meta Prometheus 训练集群存储从文件改为多层缓存对象(B&F,10-20):** 吉瓦级集群由数十万 GPU 向约 **100 万** GPU 扩;L1/L2 GPU 主机内存+本地闪存、L3 区域闪存 BLOB、L4 HDD Tectonic 为真源;ZippyDB 扁平元数据 O(1) 查找、「fat client」SDK 直读、复用 Owl 做主机空闲内存分布式缓存——**媒体报道(引 Meta 自述)**。对规划:AI 训练数据面「对象+分层缓存」继续压过 NFS 类文件接口。 2. **中 NEW · NVIDIA BlueField-4 STX Server for CMX(OCP 2026,STH):** 合作伙伴展示单 Vera 小板(SOCAMM 内存、M.2),相对此前双 Vera + 4×ConnectX-9 方案成本更低;STH 称照片中**未见** ConnectX-9;文中提及 NAND 价格过去一年明显上涨——**媒体现场报道**。对规划:上下文内存(KV)存储节点形态在 DPU 与服务器之间分化。 3. **中 NEW · Semperis 2026 勒索报告(VR,10-20):** Censuswide 调查 9 国 **1,200** 名 IT/安全人员,**仅限**遭受过造成业务中断勒索的组织;**84%** 遭遇以某种方式利用 AI 的攻击(AI 驱动 40% / 操纵受害方自有 AI agent 或系统 44%);身份设施在 **87%** 案例被攻击(AD 51% / Entra ID 48% / Okta 27%);**59%** 支付赎金——**厂商委托调研经媒体**,样本有偏。 4. **中 NEW · DeepL 选 VAST(B&F,10-20):** DeepL 自称按部署计划可从**现有硬件**获得约 **3 倍**存储容量;用 VAST SyncEngine 迁移数据、AI OS 同时承担文件与对象——**厂商通稿经媒体**;B&F 同日「Testing AI agent accuracy」引 DataHub 案例:Miro 900 题 text-to-SQL 基准从直连 Snowflake <40% 到加入上下文层后约 **90%**(客户方经厂商转述)。 5. **轻–中 NEW · HPE Spaceborne Computer-4 登月(STH):** 计划 2026 年底随 Griffin-1 着陆器、Astrolab FLIP 月球车运行 LLM RAG、视觉推理、LiDAR 处理;无风扇、低功耗 CPU+GPU,存储为 Kioxia BG 系 M.2 SSD;地面以 HPE Private Cloud + CloudOps 管理——**厂商计划经媒体**;边缘极端环境叙事。Solidot 快讯称字节/阿里/腾讯订购逾 **300 万**片、总值逾 **350 亿美元** RTX PRO 5500,NVIDIA 告知伙伴实际可得或仅 **5%–10%**——**单一媒体快讯,未核原始出处**,仅作供应紧张旁证。 6. **对规划一句:** 产业增量在 **AI 数据面(Meta 对象分层缓存 / VAST 复用旧盘 / BlueField-4 CMX)** 与 **勒索×身份×AI agent 攻击面(Semperis)**;软件面见友商(llm-d 1.0、VCF 9.1.1 容量预测与租户网络六连、ZStack Zentrix 故障转移复盘)。续档:内存/闪存超级周期、Grafana 自治信任鸿沟、VergeOS 26.2 等无推翻性新量化。 ## 原文摘录 ### 1. 中 NEW · Meta Prometheus(B&F,10-20) - 归档:`industry/meta-s-prometheus-using-object-storage-for-massive-ai-training-jobs-f54bec54.txt` - URL:https://www.blocksandfiles.com/ai-ml/2026/10/20/metas-prometheus-using-object-storage-for-massive-ai-training-jobs/5304117 - 口径:**媒体报道**(引 Meta 工程资料)。 > Meta has started talking about its Prometheus gigawatt-scaleAI training cluster and its storage system has migrated from file-based to multi-tiered cached object storage to reduce GPU idle time. > It is being scaled up from its current hundreds of thousands of GPUs to to around 1 million GPUs… > Meta says the core idea is to have a dataloading scheme leveraging the memory and flash on the GPU host as L1 and L2 caches, and regional BLOB-storage fabric backed by flash as the L3 cache. ### 2. 中 NEW · BlueField-4 STX CMX(STH) - 归档:`industry/nvidia-bluefield-4-stx-server-for-cmx-at-ocp-2026-c403dd20.txt` - URL:https://www.servethehome.com/nvidia-bluefield-4-stx-server-for-cmx-at-ocp-2026/ > For context memory platforms, a single Vera versus a dual Vera makes this a lower-cost platform, and that will likely help the STX/ CMX systems come to market as the price of NAND has been rising meaningfully over the past year. ### 3. 中 NEW · Semperis Ransomware Risk Report(VR,10-20) - 归档:`industry/semperis-ransomware-hits-ai-systems-and-employees-families-89404ad5.txt` - URL:https://virtualizationreview.com/articles/2026/10/20/semperis-ransomware-hits-ai-systems-and-employees-families.aspx - 口径:**厂商委托调研**(Semperis 售身份安全/恢复产品)。 > The report says 84% of respondents experienced attacks that exploited AI in some way. Its breakdown distinguishes AI-based attacks, reported by 40%, from attacks that manipulated victims' own AI agents or systems, reported by 44%. > Attackers targeted or exploited identity infrastructure in 87% of surveyed cases. > Overall, 59% paid a ransom. ### 4. 中 NEW · DeepL×VAST / Testing AI agent accuracy(B&F,10-20) | 条目 | 归档 | 口径 | |---|---|---| | German AI-powered language translation powerhouse goes VAST | `industry/german-ai-powered-language-translation-powerhouse-goes-vast-9ba0432a.txt` | 厂商通稿经媒体;3× 为客户预期 | | Testing AI agent accuracy | `industry/testing-ai-agent-accuracy-b82c505a.txt` | 媒体·DataHub 案例;数字为客户经厂商转述 | > Based on current deployment plans, DeepL expects to achieve roughly three times the storage capacity from its existing hardware… > DeepL will use VAST’s SyncEngine technology to migrate data from its existing environment onto the VAST platform… > After layering in DataHub Cloud as our context platform… we nearly doubled accuracy from close to 50 percent to around 90 percent.(Miro,经 DataHub/B&F) ### 5. 轻–中 NEW · HPE Spaceborne-4(STH)/ 中文快讯旁证 | 条目 | 归档 | 口径 | |---|---|---| | HPE Spaceborne Computer-4 … Astrolab FLIP | `industry/hpe-spaceborne-computer-4-will-bring-ai-and-hpc-to-the-moon-on-astrola-39c67512.txt` | 厂商计划经媒体 | | Solidot 首页(RTX PRO 5500 订单) | `industry/奇客资讯网-ed448ea7.txt` | 单一媒体快讯,**未核原始出处** | | 至顶网首页(CERT-In:AI 厂商智能体测试中越权访问数据) | `industry/至顶网-af3b259d.txt` | 媒体摘要 | > Once deployed, Spaceborne Computer-4 will run large language model retrieval-augmented generation, vision inference, LiDAR data processing for mapping and navigation, and an AI services container… > 字节跳动、阿里巴巴和腾讯已订购超过 300 万个、总值超过 350 亿美元的英伟达 RTX PRO 5500 工作站显卡。……他们实际能拿到的产品数量,可能仅为所提出数量的 5% 到 10%。(Solidot) ### 6. 续档(强信号) | 条目 | 状态 | |---|---| | Grafana Survey 85%/50%(VR 10-19) | **续档** | | VergeOS v26.2 / SK hynix H-TPU / IBM Ceph / Maia 300(B&F·NP 首获) | **续档** | | 内存/闪存/HDD 超级周期 | **续档**;BlueField-4 文旁证 NAND 涨价,无新独立量化 | | RTX Spark N1X / Helios×QFX5252 / Intel D270 / Marvell CXL / EPYC 8005 | **续档** | | NextPlatform / The Register | 本日**无新 URL** | | Reddit | host_blocked | ### 7. 对规划一句 AI 数据面向「对象+分层缓存+上下文内存节点」收敛(Meta / VAST / BlueField-4 CMX),HCI 需说清分布式块之外的对象/KV 承载边界;安全面勒索已瞄准身份设施与企业自有 AI agent(Semperis),与 VMware 安全月、Grafana 写权限信任鸿沟同向。