综合整理自 TechCrunch、The Verge、IT之家、新浪科技、科创板日报、澎湃新闻等当日公开报道，聚焦 AI 资本走向、算力金融化、模型能力与安全边界、国产算力推理成本、具身智能度量体系与产业应用落地。📌 今日摘要
本期的关键词是**"算账"**——所有人开始算账。资本面上，OpenAI 最新一轮融资目标从 300 亿美元下调至 200 亿美元，缩水约 33%，《华尔街日报》披露目前仅募得 110 亿美元，部分大额投资者反应"冷淡"；美股 10 月最后一个交易日以惨烈收场，纳斯达克 100 指数单日下跌 3.5%，全月市值蒸发规模或达 2 万亿美元，苹果则凭借稳定的消费电子基本盘以 5.03 万亿美元市值反超英伟达，重夺全球市值第一。算力金融化是本期最值得玩味的现象：英伟达正洽谈为 Nscale 的 460 兆瓦数据中心提供超过 50 亿美元的"剩余价值担保"，芯片供应商就此升级为项目融资的兜底方。技术侧，两项研究把智能体的能力边界推到了新位置——Claude Sonnet 5 自主从 19 种候选基因治疗中选定方案并操控手术设备完成小鼠脑部注射，GLM 5.3 在 48 小时内自主构建出 6 个 Firefox 版本的完整漏洞利用链，单次攻击的 API 成本约 1000 美元。国产侧，天数智芯天垓 300 在特定长上下文与 MoE 负载下实测吞吐超越 H200，趋境科技联合智谱、摩尔线程发布国产异构推理"Token 工厂"方案。产业侧，智元推出 Worken（有效智能工时）度量衡，把具身智能的评判标准从"下线多少台"推向"干了多少活"。
The keyword of this issue iscounting the money— everyone has started counting. On capital, OpenAI has cut the target of its latest funding round from $30 billion to $20 billion, a roughly 33% trim, with the Wall Street Journal reporting only $11 billion raised so far and a "tepid" response from some large investors; US equities ended October's last trading session in brutal fashion, the Nasdaq 100 fell 3.5% in a single day, and as much as $2 trillion in market value may have evaporated over the month, while Apple reclaimed the top market-cap slot with a steadier consumer-electronics base at $5.03 trillion, overtaking Nvidia. The most interesting development of the issue is the financialization of compute: Nvidia is in talks to provide more than $5 billion in "residual value guarantees" for Nscale's 460-megawatt data center, turning the chip supplier into a backstop for project financing. On the technical side, two studies push the boundary of agent capability further — Claude Sonnet 5 autonomously selected from 19 candidate gene therapies and operated surgical equipment to complete a brain injection in a mouse, while GLM 5.3 autonomously built complete exploit chains for six Firefox versions within 48 hours, at roughly $1,000 in API cost per attack. On the domestic side, Iluvatar CoreX's Tiangai 300 posted measured throughput above H200 under specific long-context and MoE loads, and Qijing Technology released a domestic heterogeneous-inference "Token factory" solution with Zhipu and Moore Threads. On the industrial side, AgiBot introduced Worken (effective intelligent work hours) as a new metric, shifting embodied AI evaluation from "how many units shipped" to "how much work got done."
一、资本面换挡：OpenAI 融资目标缩水 100 亿美元，苹果市值反超英伟达事件概要
10 月 31 日，TechCrunch 报道，OpenAI 最新一轮融资并不如预期顺利。据《金融时报》当日披露，这家 ChatGPT 开发方已将融资目标从 300 亿美元下调约 100 亿美元至 200 亿美元，对应估值约 1.25 万亿美元——此前 9 月底一度传出"1.4 万亿美元估值募资 300 亿美元"，10 月 22 日已调整为 1.25 万亿美元估值，而本轮被定位为投资者在明年 IPO 前入股的"最后机会"。《华尔街日报》补充称首轮交割目前仅获得 110 亿美元，且部分大额投资者反应"冷淡"；首关由阿联酋国家支持科技基金 MGX 牵头的财团领投，红杉资本、Founders Fund、Coatue、a16z 与软银等亦计划参与，后续还可能面向富裕个人开放零售认购。同日 IT之家援引报道指出，本轮节奏还受竞争对手影响：Anthropic 预计下月上市，估值可能达到 1.5 万亿美元或更高，正持续吸引投资者资金。
市场同步剧烈反应。美股 10 月最后一个交易日以"惨烈"收场，纳斯达克 100 指数单日下跌 3.5%，英特尔与 Reddit 双双暴跌近三成；IT之家测算，仅 10 月一个月，纳斯达克可能蒸发约 2 万亿美元市值。10 月 31 日收盘后，苹果重新夺回全球市值最高公司头衔：苹果股价单日下跌 1.2%，总市值回落至 5.03 万亿美元，英伟达下跌 4.2%，收盘市值 4.93 万亿美元，两家均跌破 5 万亿美元关口，英伟达因跌幅更大而让出维持数月的第一位置。这是苹果自今年 7 月底短暂登顶以来首次再度反超。Counterpoint Research 同期数据显示，iPhone 18 Pro 与 Pro Max 发布后前三周全球销量同比增长 3.1%，其中中国市场同比增长 29%。
On October 31, TechCrunch reported that OpenAI's latest funding round is not going as planned. According to the Financial Times that same day, the maker of ChatGPT has cut its fundraising target by roughly $10 billion, from $30 billion to $20 billion, at a valuation of about $1.25 trillion — down from an early report on September 29 of a $30 billion raise at a $1.4 trillion valuation, itself revised on October 22, with the round positioned as a "last chance" for investors to get in ahead of next year's IPO. The Wall Street Journal added that the first close has so far brought in just $11 billion, with a "tepid" response from some large investors. The first close is expected to be led by a consortium backed by the UAE's MGX, with Sequoia, Founders Fund, Coatue, Andreessen Horowitz and SoftBank also planning to participate, and a subsequent retail offering for wealthy individuals could bring the total to the $20 billion target. IT之家 noted the same day that the pace of this round is also affected by a competitor: Anthropic is expected to list next month at a valuation of possibly $1.5 trillion or higher, continuing to pull investor money and attention.
The market reacted sharply in parallel. US equities ended October's final trading session in brutal fashion, with the Nasdaq 100 down 3.5% in a single day and both Intel and Reddit tumbling close to 30%; IT之家 estimated that as much as $2 trillion in Nasdaq market value may have evaporated over the month of October alone. After the October 31 close, Apple reclaimed the title of the world's most valuable company: its shares fell 1.2% in the session, bringing total market cap back to $5.03 trillion, while Nvidia fell 4.2% to close at $4.93 trillion. Both slipped below the $5 trillion mark, and Nvidia, hurt more by the larger decline, gave up the number-one position it had held for months. This is Apple's first retaking of the crown since it briefly did so in late July. Counterpoint Research data released the same day showed that global sales of the iPhone 18 Pro and Pro Max rose 3.1% year over year in the first three weeks after launch, with China up 29%.来源 / Sources
TechCrunch：OpenAI reportedly lowers funding target for new round
IT之家：AI 热潮降温？消息称 OpenAI 最新融资目标缩水 100 亿美元，降幅 33%
快科技 / 新浪科技：苹果市值再度反超英伟达！成全球市值最高公司
IT之家：英特尔、Reddit 双双暴跌近三成，美股 10 月最后一个交易日以"惨烈"告终
二、算力金融化：英伟达为 Nscale 460 兆瓦数据中心提供超 50 亿美元信用担保事件概要
10 月 31 日，多家媒体证实英伟达正洽谈为新兴 AI 云服务商 Nscale 提供数十亿美元的债务信用担保，用于建设一座 460 兆瓦的数据中心，该园区将专供 Anthropic 使用。据《The Information》报道与彭博社证实，协议尚未签署、条款仍可能变动；业内按"1 吉瓦供电规模数据中心土建约需 120 亿美元"推算，本笔信用支持规模大概率超过 50 亿美元。担保采取"剩余价值担保"形式——担保对象是数据中心资产本身，而非租约或芯片：若 Nscale 违约，英伟达将代为偿付建设款、临时接管该设施并协助寻找买家与新租户。这与英伟达以往主要覆盖租约与芯片采购的财务支持方式不同，意味着芯片订单、机房建设与债务结构被进一步捆绑。
交易背景同样值得关注。Nscale 招股文件披露，其已与 Anthropic 签订约 446 亿至 450 亿美元的算力租赁合同，覆盖西弗吉尼亚 Monarch 园区首座 460 兆瓦机房，内部将全部搭载英伟达下一代 Vera Rubin 芯片；双方还在洽谈第二栋规模相当的建筑。Nscale 正寻求以最高 350 亿美元估值上市，但其实际交付能力与合同规模存在明显缺口——截至 8 月实际交付容量仅 55 兆瓦，潜在开发规模达 10 吉瓦。英伟达迄今已通过投资与信用支持向 Nscale、CoreWeave 等新兴云厂商及大客户提供约 1700 亿美元资金，其中 9 月已向 Nscale 承诺 10 亿美元可转债。与此同时，中金公司在当日观点中提示 Neocloud 的脆弱性：高杠杆与现金流错配使其信用债收益率飙升至 15%，隐含违约率约 60%。
On October 31, multiple outlets confirmed that Nvidia is in talks to provide tens of billions of dollars in debt credit guarantees for the emerging AI cloud provider Nscale, funding the construction of a 460-megawatt data center that will be used exclusively by Anthropic. According to reporting by The Information, corroborated by Bloomberg, the deal has not been signed and terms may still change; industry estimates based on the roughly $12 billion needed to build the shell of a one-gigawatt data center put the likely size of this credit support above $5 billion. The guarantee takes the form of a residual value guarantee covering the data center asset itself rather than the lease or the chips: if Nscale defaults, Nvidia would step in, pay the construction obligations, temporarily take over the facility and help find a buyer and a new tenant. This differs from Nvidia's earlier financial support, which mostly covered leases and chip purchases, and it binds chip orders, facility construction and debt structure together far more tightly.
The backdrop deserves attention too. Nscale's IPO filing discloses a roughly $44.6 billion to $45 billion compute lease with Anthropic covering the first 460-megawatt building of its West Virginia Monarch campus, to be filled entirely with Nvidia's next-generation Vera Rubin chips, with talks underway for a second building of similar scale. Nscale is seeking to list at a valuation of up to $35 billion, yet the gap between contracted scale and actual delivery capacity is stark — as of August it had delivered just 55 megawatts against a potential 10 gigawatts of planned development. Nvidia has by now extended roughly $170 billion in investments and credit support to neoclouds such as Nscale and CoreWeave and to other large customers, including a $1 billion convertible commitment to Nscale in September. Against this, CICC noted in its same-day commentary the fragility of the neocloud model: heavy leverage and cash-flow mismatch have pushed credit-bond yields to 15%, implying a default probability of roughly 60%.来源 / Sources
TechCrunch：Nvidia weighs playing the cosigner for Nscale's AI data center in West Virginia
新浪财经 / 环球市场播报：英伟达拟为云服务商 450 亿美元 Anthropic 项目提供兜底支持
金十数据：从卖芯片到担保融资，英伟达与 Nscale 的合作正在跨过哪条边界？
三、模型能力与安全边界：Claude 自主完成小鼠脑部注射，GLM 5.3 攻破 6 个 Firefox 版本事件概要
10 月 31 日，多项研究披露了智能体能力边界的最新位置。其一，Claude Sonnet 5 自主从 19 种候选基因治疗方案中选定药物、脑区与注射坐标，并操控手术设备对癫痫小鼠实施注射，研究团队称这是首次由 AI agent 自主选择治疗方案并实施到活体哺乳动物；值得注意的是，术中检测到校准误差后该 agent 主动拒绝继续执行，即使重启校准也未改变判断，相关结果尚未经过同行评审。其二，Epoch AI 研究者发现，开放权重模型 GLM 5.3 能在安全补丁公开、用户尚未更新的"补丁空窗期"把公开修复转化为可用的 Firefox 漏洞利用链，48 小时内自主攻破 6 个 Firefox 版本，单次攻击的 API 成本约 1000 美元，测试期间没有触发任何被测提供商的护栏。
同期论文层面同样密集。arXiv 上提出的 ShallowDefend 框架指出大模型的安全对齐高度集中在最初几个 Token，主张通过强化学习让模型在生成回复前先输出可解析的安全决策 Token，在 14 个基准测试的多轮对话安全任务上取得受控对比中的最好表现，代价是损失少量帮助性。另一篇论文发布包含 8735 个临床假设的 STRATA 时序临床假设基准，并提出 ART 分阶段范式，使 31B 开源模型表现逼近前沿闭源水平，揭示模型的主要缺陷在筛选有效证据集而非逻辑推理本身。第三项研究提出在多智能体交互中对 LLM 输出生成前的隐层激活做线性探针检测，在模拟多轮施压测试中，对 Qwen3-32B 在 5% 误报率阈值下识别出 82% 即将泄露机密凭证的调用，但论文也指出同批数据上的词袋分类器已接近这一效果，且探针会误标合规披露。
行业规范层面同步推进。OpenAI 宣布将在 robots.txt 中遵守一项由多方支持的通用信号标准，使网站运营方能够跨多个参与机构统一声明生成式 AI 训练偏好，该策略令牌提案由 Anthropic 负责标准事务的技术负责人在 IETF 框架下牵头推进；OpenAI、谷歌、微软与 Anthropic 均为支持方。流量侧的反向压力也在显现：Reddit 首席执行官 Steve Huffman 表示，受谷歌 AI Overviews 与 AI Mode 直接给出答案的影响，Reddit 在 10 月每周损失超过 100 万次来自谷歌的搜索推荐流量。订阅政策方面，Anthropic 将 Claude Max 套餐重命名为 Max（5x）与 Ultra（20x），价格维持每月 100 美元与 200 美元不变；OpenAI 随后跟进，为每月 100、200 与 500 美元的高级订阅账户赠送等额 API 调用额度。
On October 31, a number of studies revealed the latest positions of the agent capability frontier. First, Claude Sonnet 5 autonomously selected the drug, brain region and injection coordinates from 19 candidate gene therapies, then operated surgical equipment to inject an epileptic mouse; the research team says this is the first time an AI agent has independently chosen a treatment and carried it out on a live mammal. Notably, after detecting a calibration error mid-procedure the agent refused to continue, and restarting calibration did not change its judgment; the results have not yet been peer reviewed. Second, researchers at Epoch AI found that the open-weight model GLM 5.3 can, during the window between a security patch being published and users actually updating, turn a public fix into a working Firefox exploit chain — autonomously compromising six Firefox versions within 48 hours, at roughly $1,000 in API cost per attack, with no guardrail from any tested provider triggered during testing.
The paper landscape was equally dense. An arXiv paper proposing the ShallowDefend framework observes that safety alignment in large models is highly concentrated in the first few tokens, and argues for training models via reinforcement learning to emit a parseable safety decision token before generating a reply; it achieves the best controlled-comparison performance on multi-turn conversational safety across 14 benchmarks, at the cost of a small amount of helpfulness. A second paper introduces STRATA, a temporal clinical hypothesis benchmark containing 8,735 clinical hypotheses, plus an ART staged paradigm that brings a 31B open model close to frontier closed models, revealing that the main deficiency lies in filtering the valid evidence set rather than in logical reasoning itself. A third study proposes running linear probes over the hidden-layer activations of an LLM before output generation in multi-agent interactions, identifying 82% of calls about to leak confidential credentials on Qwen3-32B at a 5% false-positive threshold — while noting that a bag-of-words classifier on the same data already approaches this result, and that the probe flags legitimately compliant disclosures.
Industry standards moved in parallel. OpenAI announced it will observe a jointly backed generic signal standard in robots.txt, allowing site operators to declare generative AI training preferences once across multiple participating bodies; the policy-token proposal is being led within the IETF framework by the Anthropic technical lead responsible for standards, with OpenAI, Google, Microsoft and Anthropic all backing it. Counter-pressure from traffic is also showing: Reddit CEO Steve Huffman said the impact of Google AI Overviews and AI Mode answering directly costs Reddit more than a million Google search referrals per week in October. On subscription policy, Anthropic renamed its Claude Max tiers to Max (5x) and Ultra (20x), keeping prices unchanged at $100 and $200 per month; OpenAI followed by granting matching API credit to premium subscription accounts at $100, $200 and $500 per month.来源 / Sources
SegmentFault 思否：AI 日报（2026年10月31日）
The Verge：Max, Ultra, Claude.
The Verge：The biggest names in AI are backing a new way to opt out of scraping
IT之家：史上最大 IPO 前夕，黑石突然改口称 Anthropic 或无法获得投资级信用评级
四、国产算力与推理成本：天垓 300 局部超越 H200，异构推理"Token 工厂"方案发布事件概要
10 月 31 日，国产 AI 芯片厂商天数智芯在媒体开放日公布第三方与实验室测试结果：在典型的 Prefill 与 Decode 分离部署下，其天垓 300 折算单卡端到端吞吐达到 H200 的 1.05 倍；官方数据显示，在 GLM-5.3 的 64K 长上下文场景中其 MoE 吞吐达到 H200 的 1.15 倍，在 8K 至 128K 上下文的 Agent 场景下首字延迟降低 10% 至 30%，并展示了进一步细化 Attention 与 FFN 运算的 AFD 分离推理方案。同日，2026 人工智能+大会开幕式暨主论坛在北京举行，聚焦国产智能计算生态建设与物理 AI 持续学习等方向，会上趋境科技牵头发布《国产异构推理高品质 AI Token 工厂解决方案》，联合智谱、摩尔线程等企业探索国产算力的协同应用——该方案通过计算层改造拆分不同计算任务，让不同类型的芯片承担各自更适合的工作，并结合推理优化与系统调度提升算力资源利用效率。
产业侧的效率焦虑同样被量化。在 2026 数据存储产业大会上发布的报告预测年底全球存储产业规模将突破 1.7 万亿美元，华为副总裁袁远指出，当前国内智算中心 GPU 平均利用率约 30%，企业可用数据资源开发率约 28.4%，存储需从单纯扩充容量转向承担数据治理与系统效率协同，重点支撑长上下文 KV Cache 加速与超大规模集群 Checkpoint 可靠性。存储供给端亦在加码：长鑫科技宣布基于 4F² 架构的第六代技术平台 DRAM 已完成工程样片验证，并已与多家头部客户协商长期供货协议；同期 IT之家报道称，OpenAI 奥尔特曼赴韩与约 50 名三星高管会面，谈论存储芯片短缺问题，并公开表示全球 DRAM 应优先供给数据中心，闲置在玩家 PC 里是浪费。更长周期看，IT之家统计称微软、谷歌、甲骨文等巨头已承诺投入超 3 万亿美元，AI 基建军备竞赛仍在持续升温。
On October 31, domestic AI chip maker Iluvatar CoreX published third-party and laboratory test results at a media open day: under a typical split Prefill/Decode deployment, its Tiangai 300 delivers per-card end-to-end throughput equal to 1.05 times that of H200; official figures show MoE throughput at 1.15 times H200 in a 64K long-context GLM-5.3 scenario, first-token latency reduced by 10 to 30 percent in Agent scenarios from 8K to 128K context, and it demonstrated an AFD split-inference scheme that further refines Attention and FFN computation. Also on October 31, the opening ceremony and main forum of the 2026 AI+ Conference took place in Beijing, focused on domestic intelligent-computing ecosystem development and continuous learning for physical AI. At the event Qijing Technology led the release of a high-quality AI Token factory solution for domestic heterogeneous inference, working with Zhipu, Moore Threads and others: the solution splits different computation tasks at the compute layer so that different chip types handle the work they suit best, and combines inference optimization with system scheduling to raise utilization of compute resources.
Efficiency anxiety on the industrial side was quantified as well. A report released at the 2026 Data Storage Industry Conference projects the global storage industry to exceed $1.7 trillion by year end, and Huawei VP Yuan Yuan noted that average GPU utilization at domestic intelligent-computing centers is only about 30 percent while enterprises' usable data resource development rate is about 28.4 percent — meaning storage must shift from simply expanding capacity toward jointly carrying data governance and system efficiency, with emphasis on long-context KV Cache acceleration and reliability of checkpoints in very large clusters. Supply is ramping: ChangXin Memory announced that DRAM on its sixth-generation 4F² architecture platform has completed engineering sample validation, and that it is already negotiating long-term supply agreements with several leading customers; IT之家 reported the same day that OpenAI's Sam Altman met with roughly 50 Samsung executives in Korea over the memory shortage, stating publicly that global DRAM should be prioritized for data centers and that leaving it idle in consumer PCs is a waste. Over the longer horizon, IT之家 tallied that Microsoft, Google, Oracle and other giants have committed more than $3 trillion, meaning the AI infrastructure arms race is still heating up.来源 / Sources
SegmentFault 思否：AI 日报（2026年10月31日）
新京报 / 新浪科技：从创新链到应用链，2026人工智能+大会共探AI产业化路径
IT之家：长鑫科技：基于 4F² 架构的第六代技术平台 DRAM 完成工程样片验证
IT之家：存储芯片短缺加剧之际，OpenAI 奥尔特曼赴韩与约 50 名三星高管会面
五、具身智能换度量衡：Worken 登场，"下线多少台"让位于"干了多少活"事件概要
10 月 31 日产业媒体集中复盘了具身智能行业最重要的度量体系转向。此前在 2026 智元创新日上，智元联合创始人、总裁兼 CTO 彭志辉发布首个定义物理 AI 价值的度量衡——Worken（有效智能工时），计算公式为"活跃机器人数量 × 自主工作时长 × 产出效能比"，用于衡量具身智能部署态的真实贡献。彭志辉的表述是："Token 衡量机器思考了多少，Worken 衡量机器人行动了多少。"他同时披露，智元今年上半年通用 AI 机器人累计下线 2 万台、营收近 20 亿元，占据全球人形机器人 43% 的市场份额，该营收已超过宇树科技今年上半年 11.52 亿元的水平。Worken 目前已落地成都宽窄巷子游客服务、龙旗科技产线上下料、广州地铁安检引导、珠海长隆文娱表演、爱仕达零售导购、香格里拉酒店服务等场景。
但数据也显示热度与落地的落差：上半年全球人形机器人六成以上卖给了舞台和实验室，真正进厂进仓的不到两成；10 月具身智能领域融资 33 起、101.05 亿元，较 9 月的 60 起、110.28 亿元笔数几乎腰斩。10 月 AI 投融资月报显示，当月我国人工智能领域共发生投融资事件 236 起，披露金额约 1391.55 亿元，位居全行业第一，同比约为去年同期的 10 倍；细分领域中具身智能与机器人以 50 起占据 21% 居首，AI 医疗 35 起居次，基础技术研发 29 起第三，10 亿元以上的大额事件共 7 起。大会层面，清雁科技联合北京航空航天大学等机构发布面向物理 AI 长程任务的机器人操作学习评测基准《ReAct-Manip》，含 16 类任务，重点考察智能体能否根据前序操作反馈调整策略；光轮智能则开源了全球首个十万小时全模态人类数据集。
单点融资与平台建设同样密集。机器最前线 10 月机器人行业月报显示，星尘智能完成 10 亿元股权投资、投控东海领投，估值约 150 亿元；启元第 5000 台个人机器人下线。智澄 AI 完成数亿元融资，由杭州资本、富浙科技等投资，公司由前 Meta 首席工程负责人胡鲁辉创办，聚焦世界模型与物理智能自进化，其 XR1 工业机器人已进入蔚来汽车工厂产线并面向大众汽车开启场景测试。宇树科技联合钱塘新区、索辰科技、中国计量大学、图岳控股五方签约，共建具身智能训练、验证与孵化平台。文远知行旗下专注物理 AI 基础设施的景烁科技完成数千万美元 Pre-A 轮融资。
On October 31, industrial media concentrated on reviewing the most important measurement shift in embodied AI. At the 2026 AgiBot Innovation Day, AgiBot co-founder, President and CTO Peng Zhihui released the first metric defining the value of physical AI — Worken (effective intelligent work hours), computed as active robot count × autonomous working duration × output efficiency ratio, used to measure the real contribution of an embodied-AI deployment. Peng put it as: "Token measures how much the machine thought; Worken measures how much the robot acted." He also disclosed that in the first half of the year AgiBot shipped a cumulative 20,000 general-purpose AI robots with revenue near 2 billion yuan, holding 43 percent of the global humanoid robot market share, a figure already above Unitree's 1.152 billion yuan in first-half revenue. Worken has already been applied in Chengdu's Kuanzhai Alley tourist service, Linekong's production-line loading and unloading, Guangzhou Metro security guidance, Chimelong's cultural shows in Zhuhai, ASD retail shopping assistants, and Shangri-La hotel service.
The data also shows the gap between hype and deployment: more than 60 percent of first-half global humanoid robot sales went to stages and laboratories, with fewer than 20 percent actually entering factories and warehouses; in October, embodied-intelligence funding came to 33 deals worth 10.105 billion yuan, against 60 deals worth 11.028 billion yuan in September — nearly halving the deal count. October's AI investment and financing report shows China's AI sector saw 236 financing events with roughly 139.155 billion yuan in disclosed amounts, ranking first across all industries and about ten times the same period a year earlier. By sub-sector, embodied intelligence and robotics led with 50 events (21 percent), AI healthcare was second with 35, and foundational AI technology research third with 29; seven events exceeded 1 billion yuan. At the conference level, Qingyan Technology together with Beihang University and other institutions released ReAct-Manip, a robotic manipulation learning benchmark for long-horizon physical-AI tasks, covering 16 task categories and focusing on whether an agent can adjust strategy based on prior action feedback; Guanglun Intelligence open-sourced the world's first 100,000-hour multimodal human dataset.
Single-deal financing and platform construction were equally dense. Robot industry monthly report from the Frontline of Robotics showed Stardust Intelligence completing a 1 billion yuan equity financing led by Tongkong Donghai at a valuation of about 15 billion yuan, and Qiyuan rolling off its 5,000th personal robot. Zhicheng AI completed a several-hundred-million-yuan round led by Hangzhou Capital and Fuzhe Technology; founded by former Meta chief engineering lead Hu Luhui, it focuses on world models and physical intelligence self-evolution, and its XR1 industrial robot has entered the NIO production line with scenario testing open for Volkswagen. Unitree signed a five-party agreement with Qiantang New District, Suochen Technology, China Jiliang University and Tuyue Holdings to jointly build embodied-AI training, validation and incubation platforms. Jingshuo Technology, a WeRide subsidiary focused on physical-AI infrastructure, completed a several-tens-of-millions-of-dollars Pre-A round.来源 / Sources
澎湃新闻：机器人越多≠生产力越多！智元彭志辉给"量产热"泼冷水
科创板日报：2万台、20亿元、43%全球份额！智元彭志辉：机器人越多生产力越多
极新 / 网易：DeepSeek 1055亿引爆，十月融资总额1391.55亿，同比约增10倍（2026年10月人工智能领域投融资分析·极新月报）
机器最前线 / 网易：10月机器人行业月报：星尘智能披露10亿元融资，启元第5000台个人机器人下线
商道创投网：景烁科技完成数千万美元 Pre-A 轮融资
六、产业落地：可灵增资 198 亿、百度搜索 API 打到 4.9 元/千次、混元开源世界模型事件概要
10 月 31 日，多项应用与商业化进展落地。快手旗下视频生成大模型可灵 AI 完成本轮增资交割，实际到账约 198.02 亿元，由金蝶、沐曦等接盘原投资方未按期缴款的份额，融资规模较上限略有缩减。百度 App 搜索与文心助手同步升级，文心助手从问答交互转向直接交付任务成果，覆盖 170 余种场景，同期升级的搜索 AI API 最低价打到 4.9 元/千次，价格低于 DeepSeek 搜索 API。腾讯混元则开放 Hy-Nomas 交互式世界模型的权重、推理与训练代码，并开源评测基准 Hy-Nomas-Bench，该模型以 8B 参数即可在单张 4090 上生成 60 秒可控世界。
开发工具与运维侧同步推进。字节跳动 TRAE 在整合 TraeCode 与 TraeWork 之后正式上线开发者个人助理，采用"全局助理 + 项目专属助理"双层架构，可跨项目记住历史上下文，并在本地仓库中自主拆解、并行执行并闭环验证多个子任务，推动 AI 编码工具从单指令生成转向项目级统筹交付。腾讯云升级智能体开发平台 ADP 并发布 AgentPortal 2.0，强化安全私网连接与租户权限隔离，支持将权限管理细化到员工 ID。华为在 HUAWEI CONNECT 2026 上海站行业运维专场推出面向企业生产系统的行业 O&M 解决方案，以 Agentic AI 重构运维范式。自动驾驶端，广汽领程与小马智行联合研发的 T9 ROBOTRUCK L4 级自动驾驶重卡首批量产车下线，自动驾驶重卡从测试验证进入规模化交付阶段；小鹏 MONA L05 上市，13.98 万元起售，搭载图灵第二代 VLA 模型，把高阶智能驾驶带入主流价位区间。
硬件与前沿科学端亦有进展。IT之家报道马斯克公布 Neuralink 脑机新产品进展，其首款产品为"心灵感应"，第二款产品"盲视"在技术上已准备就绪，只待监管批准；理想 AI 眼镜 Livis 完成 10 月 OTA 升级，个人 Agent 新增接入 Hermes；IT之家另报道中国爆款 AI 漫剧全球"刷屏"，海外用户已超 1.5 亿。海外侧，The Verge 报道谷歌开发者推出名为 Sprout 的机器人盆栽项目，这是用于展示 Google AI Studio 的"仅供娱乐"尝试：Sprout 会自主游走寻找最佳光照，并在植物需要浇水时发出提醒。The Verge 另一篇报道记录了一名开发者用 Waveshare ESP32 加 Claude vibe coding 抢先做出 AI 养成游戏 Muse Charm 的过程。
On October 31, a number of application and commercialization milestones landed. Kling AI, the video generation model under Kuaishou, completed the closing of its current capital increase with roughly 19.802 billion yuan actually received, with Kingdee and Moore Threads taking over the shares that original investors failed to pay for on schedule, leaving the round slightly below its cap. Baidu upgraded its App search and the Wenxin Assistant in step, moving the assistant from question-and-answer interaction to directly delivering task outcomes across more than 170 scenarios, while the upgraded search AI API floor price dropped to 4.9 yuan per thousand calls — below DeepSeek's search API. Tencent's Hunyuan opened the weights, inference and training code for the Hy-Nomas interactive world model and open-sourced the Hy-Nomas-Bench evaluation benchmark; with 8 billion parameters, the model can generate a 60-second controllable world on a single RTX 4090.
Developer tooling and operations advanced together. ByteDance's TRAE, having consolidated TraeCode and TraeWork, officially launched a personal developer assistant built on a dual "global assistant + project-specific assistant" architecture that remembers context across projects and autonomously decomposes, runs in parallel and closes the loop on multiple subtasks inside a local repository — pushing AI coding tools from single-instruction generation toward project-level orchestration and delivery. Tencent Cloud upgraded its agent development platform ADP and released AgentPortal 2.0, strengthening secure private-network connectivity and tenant permission isolation, with permission management refinable to the individual employee ID. At the HUAWEI CONNECT 2026 Shanghai session on industry operations, Huawei introduced an industry O&M solution for enterprise production systems, reframing operations around Agentic AI. On autonomous driving, the first production units of the T9 ROBOTRUCK L4 autonomous heavy truck, jointly developed by GAC Lingcheng and Pony.ai, rolled off the line in Conghua, marking a move from testing to scaled delivery; the Xiaopeng MONA L05 launched at 199,800 yuan with the second-generation Turing VLA model, bringing advanced assisted driving into the mainstream price band.
Hardware and frontier science also produced news. IT之家 reported that Musk detailed new Neuralink brain-computer products: the first is "Telepathy" and the second, "Blindsight," is technically ready and awaits regulatory approval. The Ideal AI glasses Livis completed an October OTA update, adding Hermes access for its personal agent. IT之家 also reported that Chinese hit AI short dramas are going viral globally with more than 150 million overseas viewers. On the international side, The Verge reported that Google developers have built a robot planter called Sprout, a "just for fun" project showcasing Google AI Studio: Sprout wanders autonomously to find the best sunlight and sends an alert when the plant needs watering. A separate Verge piece documented a developer's process of beating Meta to an AI virtual pet, building the Muse Charm gadget from a Waveshare ESP32 with Claude vibe coding.来源 / Sources
新浪科技：中国爆款AI漫剧全球"刷屏"：超1.5亿海外用户在看
IT之家：OpenCode 启动 Cheepseek 第四阶段：10 美元订阅享 100 美元 DeepSeek V4.1 Flash 额度
The Verge：Google made a robot planter.
IT之家：马斯克公布 Neuralink 脑机新产品进展：第二款产品"盲视"只待监管批准行业数据速览
指标
数据
备注
OpenAI 最新融资目标
200 亿美元（原 300 亿）
缩水约 100 亿美元，降幅 33%；估值约 1.25 万亿美元
OpenAI 首关已募资
110 亿美元
《华尔街日报》披露，部分大额投资者反应冷淡
纳斯达克 100 指数
周五单日 -3.5%
10 月单月市值或蒸发约 2 万亿美元
苹果总市值
5.03 万亿美元
10 月 31 日收盘跌 1.2%，重夺全球第一
英伟达总市值
4.93 万亿美元
跌 4.2%，让出维持数月的第一
iPhone 18 Pro 系列前三周销量
全球同比 +3.1%，中国 +29%
Counterpoint Research 口径
英伟达对新兴云厂商支持
约 1700 亿美元
投资 + 信用支持累计口径
Anthropic–Nscale 算力合同
约 450 亿美元
首栋 460MW，全数搭载 Vera Rubin
天垓 300 对比 H200
端到端吞吐 1.05 倍
GLM-5.3 64K MoE 场景达 1.15 倍
国内智算中心 GPU 平均利用率
约 30%
企业可用数据资源开发率约 28.4%
全球存储产业规模（年底预测）
突破 1.7 万亿美元
2026 数据存储产业大会报告
AI 基建巨头累计承诺投入
超 3 万亿美元
微软、谷歌、甲骨文等口径
智元上半年机器人下线
2 万台，营收近 20 亿元
全球人形机器人份额 43%
Worken 计算公式
活跃机器人数 × 自主工作时长 × 产出效能比
2026 智元创新日发布
10 月具身智能融资
33 起 / 101.05 亿元
9 月为 60 起 / 110.28 亿元，笔数近乎腰斩
10 月 AI 领域投融资
236 起 / 1391.55 亿元
同比约增 10 倍；具身智能占 21% 居首
人形机器人真实进场比例
不足两成
上半年六成以上售予舞台与实验室
GLM 5.3 漏洞利用链
48 小时攻破 6 个 Firefox 版本
单次攻击 API 成本约 1000 美元
百度搜索 AI API
最低 4.9 元/千次
文心助手覆盖 170 余种场景
快手可灵 AI 增资
实际到账约 198.02 亿元
交割完成，金蝶、沐曦等接盘
Hy-Nomas 世界模型
8B 参数，单张 4090 生成 60 秒可控世界
权重、推理与训练代码已开源
媒体人解读
这期日报最值得技术管理者反复咀嚼的，不是任何一个单点的模型发布，而是**"算账"这件事本身从边缘变成了主线**。
第一件事是资本结构变了。以前 AI 公司的融资叙事是"我们拿到了多少亿、估值多少万亿"，是能力证明；现在 OpenAI 把 300 亿的目标砍到 200 亿、只交割了 110 亿，这不再是能力问题，而是需求侧验证问题——当你能融到的钱少了，你的估值叙事就必须从"技术领先"转向"单位经济性成立"。对技术管理者的直接含义是：如果你的团队还在用"我们用了某某大模型"作为立项理由，接下来很难通过评审。换成"这个 Agent 每处理一单成本多少钱、替代了原来几个人的多少工时"，反而更立得住。
第二件事是算力的金融属性被摆到了台面上。英伟达为 Nscale 的机房做"剩余价值担保"，本质上是芯片供应商在用自己的信用替客户的资产负债表兜底。这个动作的信号意义大于金额：一方面说明 AI 基础设施的真实瓶颈已经从"有没有芯片"转移到"能不能借到钱建机房"；另一方面也说明这条链上的风险正在高度集中——如果担保集中违约，损失会同时传导到芯片厂商、芯片客户、云厂商和最终用算力的模型公司。做技术选型时值得多问一句：我的算力供应商，其资金来源是否也依赖类似的信用结构？
第三件事是度量的迁移。Worken 公式（活跃机器人数 × 自主工作时长 × 产出效能比）看上去朴素，但它是行业第一次给"机器人到底干了多少活"下了可核算的定义。这个思路可以立刻搬到软件工程侧——团队的产出也不该用"写了多少行代码""发布了多少版本"来衡量，而应该用"多少条工位被持续替代""多少工时被真正释放"。当一个行业的度量衡换了，淘汰赛的打法就会换，从抢参数变成拼运营。
第四件事是安全研究越过了"能不能做"的边界，开始戳"该不该做"。Claude 自主给癫痫小鼠注射并在校准误差出现时拒绝继续，GLM 5.3 在补丁空窗期把公开修复直接变成可用攻击链——这两件事的共同点是：模型的能力已经强到足以在缺少人工复核的情况下造成不可逆后果。对技术管理者最务实的动作有两个：一是把"补丁空窗期"当成一个新的资产类别去做补丁响应演练，二是把智能体的"拒绝执行"当成一等公民来设计流程，而不是把它当异常。
最后是国产侧的清醒与急迫。天垓 300 在特定长上下文 MoE 负载上跑赢 H200，这是真实的技术信号，但国内智算中心 GPU 平均利用率只有约 30%——芯片的账面性能和集群的实际产出之间，隔着一个巨大的工程鸿沟。同样，异构推理"Token 工厂"方案发布、DeepSeek V4.1 Flash 把 64GB 本地部署变成现实，说明推理成本正在快速下探。对应用层团队来说，11 月最值得做的一件事是重算自己的推理账单：今天的默认模型选择，很可能是按去年的价格假设做的。
The most instructive thread in this issue is not any single model launch but the fact thatcounting the money has moved from the margins to the center line.
First, the structure of capital has changed. The old AI narrative was "we raised X billion at a Y trillion valuation" — a proof of capability. Now OpenAI has cut its target from $30 billion to $20 billion and closed just $11 billion. That is no longer a capability problem but ademand-validation problem: when the money available to you shrinks, the valuation story has to shift from "we lead on technology" to "our unit economics work." For engineering leaders the implication is direct. If your next project proposal still leads with "we use model X," it is going to get harder to pass review. Leading with "this agent costs Y cents per transaction and replaces Z hours of human work" will hold up better.
Second,compute now has a balance sheet. Nvidia offering a residual value guarantee on Nscale's buildings is, in effect, the chip supplier underwriting its customer's credit. The signal matters more than the size: the bottleneck in AI infrastructure has moved from "do we have chips" to "can we borrow to build the room," and the risk along that chain is concentrating. If guarantees default, the losses propagate at once to the chip vendor, the chip customer, the cloud operator, and finally the model companies that consume the compute. When making technology choices, one extra question is worth asking: does my compute provider's own funding structure depend on similar credit arrangements?
Third,metrics have migrated. The Worken formula — active robot count × autonomous working duration × output efficiency ratio — looks plain, but it is the industry's first auditable definition of how much work a robot actually did. The same idea transfers immediately to software engineering. Team output should not be measured in lines written or versions shipped; it should be measured in how many workstation slots are durably replaced and how many hours are genuinely released. When an industry changes its yardstick, the elimination criteria change with it, and the fight moves from parameter leadership to operational discipline.
Fourth,safety research has crossed from "can it" toward "should it."Claude autonomously injecting an epileptic mouse, then refusing to continue when it detected a calibration error, and GLM 5.3 turning a public patch straight into a working exploit chain during the patch window, share one property: the models are now strong enough to cause irreversible consequences without human review. Two pragmatic moves follow. Treat the "patch window" as a new asset class and rehearse patch-response accordingly. And design agent workflows so that "refuses to execute" is a first-class, instrumented outcome rather than an anomaly to be debugged away.
Finally, the domestic picture shows both promise and urgency. Tiangai 300 outperforming H200 under specific long-context MoE loads is a genuine technical signal — but with average GPU utilization at domestic intelligent-computing centers around 30 percent, a wide engineering gap separates a chip's benchmark performance from a cluster's actual output. Likewise, the heterogeneous-inference Token factory solution and DeepSeek V4.1 Flash making 64GB local deployment practical show inference costs falling fast. For application teams the single most valuable thing to do this month is to recompute the inference bill: today's default model choice was very likely made on last year's price assumptions.参考来源英文来源
TechCrunch：OpenAI reportedly lowers funding target for new round
TechCrunch：Nvidia weighs playing the cosigner for Nscale's AI data center in West Virginia
TechCrunch：Anthropic expected to receive high 'junk' rating from credit agencies following IPO, FT report says
TechCrunch：Intel's stock crashes, knocking $12B from government's stake
The Verge：Google made a robot planter.
The Verge：Max, Ultra, Claude.
The Verge：The biggest names in AI are backing a new way to opt out of scraping
The Verge：I built an AI Tamagotchi before Meta could中文来源
IT之家：AI 热潮降温？消息称 OpenAI 最新融资目标缩水 100 亿美元，降幅 33%
IT之家：英特尔、Reddit 双双暴跌近三成，美股 10 月最后一个交易日以"惨烈"告终
IT之家：史上最大 IPO 前夕，黑石突然改口称 Anthropic 或无法获得投资级信用评级
IT之家：跌掉一个 SpaceX？AI 投资回报引投资者担忧，纳斯达克本月或蒸发 2 万亿美元市值
IT之家：AI 基建掀起军备竞赛：微软、谷歌、甲骨文等巨头承诺投入超 3 万亿美元
IT之家：长鑫科技：基于 4F² 架构的第六代技术平台 DRAM 完成工程样片验证
IT之家：存储芯片短缺加剧之际，OpenAI 奥尔特曼赴韩与约 50 名三星高管会面
IT之家：马斯克公布 Neuralink 脑机新产品进展：第二款产品"盲视"只待监管批准
IT之家：OpenCode 启动 Cheepseek 第四阶段：10 美元订阅享 100 美元 DeepSeek V4.1 Flash 额度
新浪科技：苹果市值再度反超英伟达！成全球市值最高公司
新浪科技：中国爆款AI漫剧全球"刷屏"：超1.5亿海外用户在看
新京报 / 新浪科技：从创新链到应用链，2026人工智能+大会共探AI产业化路径
新浪财经 / 环球市场播报：英伟达拟为云服务商 450 亿美元 Anthropic 项目提供兜底支持
金十数据：从卖芯片到担保融资，英伟达与 Nscale 的合作正在跨过哪条边界？
澎湃新闻：机器人越多≠生产力越多！智元彭志辉给"量产热"泼冷水
科创板日报：2万台、20亿元、43%全球份额！智元彭志辉：机器人越多生产力越多
极新 / 网易：2026年10月人工智能领域投融资分析·极新月报
机器最前线 / 网易：10月机器人行业月报
机器人前瞻 / 网易：前Meta首席工程负责人创业，杭州具身创企获数亿元融资
商道创投网：景烁科技完成数千万美元 Pre-A 轮融资
SegmentFault 思否：AI 日报（2026年10月31日）
本文内容综合整理自上述公开信息来源，媒体人解读仅代表个人观点，仅供参考。
OCR:IMG:CSDN@刘一说
OCR:IMG:LV CSDN@刘一说
OCR:IMG:国产芯片×超长上下文 异构工厂·协同计算 Token持续产出 混合专家负载·吞吐比肩并局部反超 计算层拆分·异构协同·Token工厂 国际旗舰芯片对比 上下文长带（超长序列） Token光带 稀缺激活节点 （持续产出） (MoE) 比肩·局部反超 性能对比 方形核 条形核 晶圆核 国产芯片 INTEL 市 CHINA 盟直型 调度轨道 任务交换 专属工位 与调度 上下文长度 推理吞吐量 延迟 GPU显存占用 吞吐量 延迟 Token/s 128Ktokens cokens 1.8xtokens/s 1.8x tokens/s 0.8×秒 1.2×倍 2.2x tokens/s 0.6×秒 3.1×倍 CSDN@刘一说
OCR:IMG:视频生成：动态帧流 搜索与助手：成果交付 视频生成模型，动态顿流渲染 从问答转向直接交付成果 助手 Q搜索 帮我查找最新的AI产业落地案例.. 专他再 最玻低 大额增资落地 资本加持，产业加 地市服珠 AI 山 产业落地 世界模型：60秒场景 中 项目级统筹：闭环验证 小参数量开源世界模型，六面立方展开为连续可操控的六十秒场景 从单指令生成转向项目级统筹交付 端到端调度 全局调 多任务并行 度中枢 闭环验证 世界 模型 EC B B % 子任务一 m 子任务二 3 求分析 方案设计 子任务三 数据处理 子任务四 子任务五 模型消练 评估验证 子任务 子任务七 部署上线 60秒 12秒 24秒 60秒连续可操控场景 CSDN@刘一说
OCR:IMG:下线多少台 有效工作时长 衡量价值 才是价值标尺 有效工作时长 有效工作时长 大硕烈航动 AIBot AlBot AlBot 号：2024001 号：22400 0 丽 8,888,88888 AiBot AB0E 88 电 HOTEL 活跃数量工时 新的度量方式 hhhhl 效率×时间=真正的价值 公众号·刘一份 CSDN@刘一说
OCR:IMG:W CSDN@刘一说
OCR:IMG:国产芯片×超长上下文 异构工厂·协同计算 Token持续产出 混合专家负载·吞吐比肩并局部反超 计算层拆分·异构协同·Token工厂 国际旗舰芯片对比 上下文长带（超长序列） Token光带 稀缺激活节点 （持续产出） (MoE) 比肩·局部反超 性能对比 方形核 条形核 晶圆核 国产芯片 INTEL 市 CHINA 盟直型 调度轨道 任务交换 专属工位 与调度 上下文长度 推理吞吐量 延迟 GPU显存占用 吞吐量 延迟 Token/s 128Ktokens 1.8xtokens/s 1.8x 0.8×秒 1.2×倍 2.2x 号3.1x剂一份 cokens tokens/s tokens/s CSDN@刘一说
OCR:IMG:CSDN@刘一说
OCR:IMG:公众号·刘份 CSDN@刘一说
OCR:IMG:下线多少台 有效工作时长 衡量价值 才是价值标尺 有效工作时长 有效工作时长 AIBot AlBot AlBot 大顺白联动 号：20401 号：2240001 ：01 8,888,8888 大国型区动 AIBot 火 88 AIBot 中 1号2 HOTEL 活跃数量×工时 新的度量方式 mhhhhhhhhiml 效率×时间=真正的价值 CSDN@刘一说
OCR:IMG:视频生成：动态帧流 搜索与助手：成果交付 视频生成模型，动态顿流渲染 从问答转向直接交付成果 助手 Q搜索 帮我查找最新的AI产业落地案例. 专他再 最玻低 大额增资落地 资本加持，产业加 地市服珠 AI 山 产业落地 世界模型：60秒场景 中 项目级统筹：闭环验证 小参数量开源世界模型，六面立方展开为连续可操控的六十秒场景 从单指令生成转向项目级统筹交付 端到端调度 全局调 多任务并行 度中枢 闭环验证 世界 模型 EC B B % 子任务一 子任务二 m 3 求分析 方案设计 子任务三 数据处理 子任务四 子任务五 模型消练 评估验证 子任务 子任务七 部暑上线 60秒 公众号 12秒 24秒 60秒连续可操控场景 CSDN@刘一说
OCR:IMG:LI CSDN@刘一说
OCR:IMG:LY 公众号刘一份 CSDN@刘一说
OCR:IMG:下线多少台 有效工作时长 衡量价值 才是价值标尺 有效工作时长 有效工作时长 大硕烈航动 AlBot AlBot AlBot 号：2024001 号：2400 0 品丽 8,888,88888 AiBot AB0E 88 电 HOTEL 活跃数量工时 新的度量方式 hhhhl 效率×时间=真正的价值 CSDN@刘一说
OCR:IMG:公众号·刘一份 CSDN@刘一说
OCR:IMG:CSDN@刘一说
OCR:IMG:535553535535 222222222222 公众号·刘一份 CSDN@刘一说
OCR:IMG:视频生成：动态帧流 搜索与助手：成果交付 视频生成模型，动态帧流渲染 从问答转向直接交付成果 助手 Q搜索 帮我查找最新的AI产业落地案例. 专他再 最玻低 大额增资落地 资本加持，产业加 前地品市加楼 AI 产业落地 世界模型：60秒场景 项目级统筹：闭环验证 小参数量开源世界模型，六面立方展开为连续可操控的六十秒场景 从单指令生成转向项目级统筹交付 端到增调度 全局调 多任务并行 度中枢 ①闭环验证 快 8 子任务一 子任务二 lm F 需求分析 方案设计 子任务三 环境搭建 数据处理 子任务四 子任务五 模型W练 评估验证 子任务六 子任务七 郝著上线 60秒 0秒 12秒 48秒 60秒 24秒 60秒连续可操控场景 CSDN@刘一说
OCR:IMG:535553559535 222222222222 CSDN@刘一说
OCR:IMG:国产芯片×超长上下文 异构工厂·协同计算 Token持续产出 混合专家负载·吞吐比肩并局部反超 计算层拆分·异构协同·Token工厂 国际旗舰芯片对比 上下文长带（超长序列） Token光带 稀缺激活节点 （持续产出） (MoE) 比肩·局部反超 性能对比 方形核 条形核 晶圆核 国产芯片 INTEL 十 CHINA nnna 盈直屋 4 etoi 调度轨道 任务交换 专属工位 与调度 上下文长度 推理吞吐量 延迟 GPU显存占用 吞吐量 延迟 Token/s 128K tokens 1.8×tokens/s 1.8× tokens/s 0.8×秒 1.2×倍 2.2xtokens/s 2.2x tokens/s 0.6×秒 3.1×倍 CSDN@刘一说
OCR:IMG:535553539539 22222222222 CSDN@刘一说