AI科技热点日报 | 2026年06月28日

AI科技热点日报 | 2026年06月28日

综合整理自TechCrunch、The Verge、Reuters、第一财经、钛媒体、36氪、新华网科技、虎嗅、网易科技、雷锋网等国内外主流科技媒体当日报道，
聚焦人工智能领域核心动态，附资深媒体人综合解读。

📌 今日摘要

2026年6月28日，全球AI产业延续"芯片军备竞赛+大模型组织扩容"双主线：海外端，OpenAI与博通联合推出的自研推理芯片Jalapeño正式进入工程样品阶段，英伟达2026股东大会宣告"AI工厂时代"全面到来；国内端，DeepSeek宣布全部门规模至少翻倍、36个岗位同步开放，被业内视为这家以"技术优先"著称的初创公司正式迈入商业化深水区。同期，Anthropic最强网络安全模型Mythos 5获美国政府解禁，OpenAI GPT-5.6系列开启限量预览，华为联合中国电信在MWC上海发布星河AI网络Token化方案，AI产业正从"模型能力比拼"切换至"算力底座+应用场景"双轮驱动。

On June 28, 2026, the global AI industry continued its twin narratives of a chip arms race and large-model organizational scaling. On the international front, OpenAI and Broadcom’s jointly developed inference chip Jalapeño entered the engineering-sample stage, and NVIDIA’s 2026 Annual Shareholder Meeting officially declared the arrival of the “AI Factory Era.” On the domestic front, DeepSeek announced it would at least double the headcount of every department, opening 36 positions simultaneously — a move the industry views as the once research-first startup formally entering commercial waters. In parallel, Anthropic’s most powerful cybersecurity model Mythos 5 was unblocked by the U.S. government, OpenAI kicked off a limited preview of the GPT-5.6 family, and Huawei unveiled a Token-centric Xinghe AI Network solution together with China Telecom at MWC Shanghai. AI is shifting from “model capability contest” to a dual-driven “compute substrate + application scenarios” paradigm.

一、DeepSeek全部门规模至少翻倍：36个岗位开放，Harness团队公开招聘

事件概要

6月25日晚，深度求索（DeepSeek）在官方公众号及官网同步发布大规模招聘信息，明确将所有部门人员规模扩大至少一倍。本次招聘覆盖全栈开发、AI核心系统研发、运维、产品、模型数据策略、深度学习研究及综合职能等七大类、共计36个岗位，工作地点涵盖北京、杭州、乌兰察布，所有岗位均接受实习。值得关注的是，DeepSeek同步开始招募HR团队、行政团队以及AGI业务管培生等职能岗位，被业内解读为这家长期保持低调的公司开始搭建完整的商业化组织架构。

DeepSeek’s Harness团队负责人崔添翼在海外社交平台公开表示"部门仍然非常缺人，自己每天都在面试"，并强调没有"不招外国人"的规定，只需具备中文工作能力。BOSS直聘平台数据显示，DeepSeek目前在招职位已达121个，产品端除智能体应用外，还新增AI产品运营、情感智能数据产品经理、专业领域数据产品经理等岗位，被解读为DeepSeek正将竞争焦点从底层模型能力转向更细分、更专业的产品体验与行业应用。

On the evening of June 25, DeepSeek simultaneously published a large-scale recruitment notice on its official WeChat account and website, clearly stating it would at least double the headcount of every department. The recruitment covers seven major categories totaling 36 positions — full-stack development, AI core system R&D, operations, product, model data strategy, deep learning research, and general functions — with work locations in Beijing, Hangzhou and Ulamqab, and all positions open to interns. Notably, DeepSeek is also recruiting HR, administrative teams and AGI business trainees, which the industry interprets as the long-low-profile company beginning to build a complete commercial organizational structure.

Cui Tianyi, head of DeepSeek’s Harness team, publicly stated on overseas social media that “the department is still severely short-staffed, and I’m interviewing every day,” stressing there is no “no foreigners” rule and only Chinese working proficiency is required. Data from Boss Zhipin shows DeepSeek currently has 121 open positions; on the product side, beyond agent applications, new roles include AI product operations, emotional-intelligence data product managers, and domain-specific data product managers — read by the industry as DeepSeek shifting its competitive focus from underlying model capability to more segmented, professional product experience and vertical applications.

来源 / Sources

• AI周报|DeepSeek开启大规模招聘；OpenAI发布首款定制AI推理芯片（第一财经）
• DeepSeek开始造富（虎嗅网/腾讯新闻转载）
• DeepSeek大规模招聘 涵盖7个大类的33个岗位（界面新闻）
• DeepSeek Harness团队负责人崔添翼：目前人手还是很缺（腾讯新闻）

二、OpenAI × 博通联合发布首款自研AI推理芯片Jalapeño：推理成本降低50%

事件概要

当地时间2026年6月24日，OpenAI与博通联合发布首款定制AI推理芯片Jalapeño（中文意为"墨西哥辣椒"），这是一款专用集成电路（ASIC），专为大语言模型推理设计，从架构设计到流片仅用9个月时间，刷新高性能先进半导体领域的ASIC开发周期纪录。在合作分工上，OpenAI负责芯片架构设计与模型协同优化，博通负责流片与网络硬件，加拿大电子制造商Celestica提供板卡与机架集成配套。OpenAI硬件主管Richard Ho表示，该芯片专门针对大语言模型进行优化，可以更快、更高效地运行支撑各类AI应用的模型。

博通CEO陈福阳（Hock Tan）透露，Jalapeño性能与英伟达Blackwell芯片和谷歌TPU相当，相较传统AI GPU可节省约50%成本。工程样品已在GPT-5.3-Codex-Spark模型上跑通，最终版本芯片将于2026年底开始部署到微软及其他合作伙伴的大型数据中心。OpenAI与博通预计明年部署的芯片规模将超过此前预测的1.3吉瓦算力，长期目标剑指2029年实现10吉瓦算力规模。受此消息刺激，博通股价盘前一度拉升超过3%。OpenAI由此从"软件公司"正式迈入"模型+芯片全栈平台"，对英伟达中长期估值逻辑构成实质性挑战。

On June 24, 2026 local time, OpenAI and Broadcom jointly released their first custom AI inference chip Jalapeño, an ASIC purpose-built for large-language-model inference that went from architecture design to tape-out in just 9 months — a new record for ASIC development cycles in the high-performance advanced-semiconductor field. In the division of labor, OpenAI handled chip-architecture design and model co-optimization, Broadcom managed tape-out and networking hardware, while Canadian electronics manufacturer Celestica supplied board and rack integration. Richard Ho, OpenAI’s hardware chief, said the chip is optimized specifically for LLMs, enabling faster and more efficient operation of models underlying all kinds of AI applications.

Broadcom CEO Hock Tan revealed Jalapeño’s performance is on par with NVIDIA’s Blackwell chip and Google’s TPU, saving roughly 50% cost versus traditional AI GPUs. Engineering samples have already run on the GPT-5.3-Codex-Spark model, and the final version is slated to begin deployment in Microsoft and other partners’ large data centers by the end of 2026. OpenAI and Broadcom expect next year’s deployment scale to exceed the previously forecast 1.3 GW of compute, with a long-term target of 10 GW by 2029. The news pushed Broadcom’s stock up more than 3% in pre-market trading. With this move, OpenAI formally steps from a “software company” into a “model + chip full-stack platform,” posing a substantive challenge to NVIDIA’s medium- and long-term valuation logic.

来源 / Sources

• OpenAI和博通发布AI芯片，旨在更快、更经济地运行模型（钛媒体）
• OpenAI与博通联合发布首款自研AI推理芯片Jalapeño（财联社）
• OpenAI造芯片了！AI算力正在去"英伟达化"吗？（36氪/腾讯新闻转载）
• 对标英伟达和谷歌！博通携手OpenAI推出自研定制AI芯片（财闻海外资讯）

三、英伟达2026股东大会：黄仁勋宣告"AI工厂时代"全面到来

事件概要

北京时间6月25日凌晨，英伟达2026年度股东大会如期召开。CEO黄仁勋在演讲中正式抛出"AI工厂"（AI Factory）新概念，并明确将物理AI（Physical AI）定位为下一阶段企业乃至行业的核心增长方向。黄仁勋指出，AI数据中心的核心价值在于规模化生产词元（token），词元可转化为代码、解决方案、设计成果、智能行为及各类数字化服务，是具备商业价值的利润单元，“算力越多、token越多、收入也就越多”。

股东会披露的多组硬数据印证了黄仁勋的判断：英伟达国际市场收入超过300亿美元，同比增长约三倍；近40个国家在部署英伟达驱动的AI工厂；Blackwell平台token吞吐量比次优竞品高出30倍。同时，黄仁勋宣告下一代Vera Rubin平台即将推出，本轮AI基础设施建设周期将以"数十年"为单位来衡量，涉及电网、互联网等关键基础设施升级，“将成为人类历史上规模最大的基础设施建设”。这场股东大会虽未发布新品，但黄仁勋的表态直接影响了全球科技股情绪，A股硬科技链次日获得额外提振。

In the early hours of June 25 Beijing time, NVIDIA’s 2026 Annual Shareholder Meeting convened as scheduled. CEO Jensen Huang formally introduced the new “AI Factory” concept in his keynote and clearly positioned Physical AI as the next-stage core growth direction for both the company and the industry. Huang argued that the core value of an AI data center lies in the industrial-scale production of tokens, which can be converted into code, solutions, design outputs, intelligent behavior and various digital services — commercial profit units in their own right — adding that “the more compute, the more tokens, the more revenue.”

A series of hard figures disclosed at the meeting validated Huang’s thesis: NVIDIA’s international market revenue exceeded USD 30 billion, up roughly 3x year-on-year; nearly 40 countries are deploying NVIDIA-driven AI factories; and the Blackwell platform delivers 30x more token throughput than the next-best alternative. At the same time, Huang announced the upcoming next-generation Vera Rubin platform, stating this AI-infrastructure build-out cycle will be measured in “decades” and involve upgrades to power grids, the internet and other critical infrastructure — “the largest infrastructure build-out in human history.” Although no new products were unveiled, Huang’s remarks directly swayed global tech-stock sentiment, with A-share hard-tech chains receiving an extra lift the following day.

来源 / Sources

• 英伟达股东大会释放产业信号：物理AI将成为下一轮增长核心浪潮（腾讯新闻）
• 英伟达CEO黄仁勋股东大会放言：本轮AI基建周期长达数十年（华尔街见闻）
• AI行业晨报 2026-06-26：英伟达股东大会黄仁勋宣告"AI工厂时代"来了（搜狐）

四、OpenAI GPT-5.6系列限量预览：政府审批机制下"可信合作伙伴"优先

事件概要

OpenAI于6月26日正式开启GPT-5.6系列的限量预览。该系列包含三款模型：旗舰版Sol、面向日常工作的均衡版Terra，以及快速且经济实惠的轻量版Luna。不同于以往的公开发布，本次OpenAI并未立即全面开放，而是在美国政府要求下，仅向少量"可信合作伙伴"（trusted partners）提供有限预览权限。OpenAI在公告中披露，发布前已向美国政府展示了模型能力及发布计划，首批获得权限的企业约20家左右，部分客户的模型访问权限甚至需要美国政府逐一批准。

OpenAI同时强调，公司希望这种限制性发布不会成为长期默认做法，预计将在未来几周内全面开放GPT-5.6 Sol、Terra和Luna。这一表态被业界解读为OpenAI在政府监管压力下的一次折中妥协，反映出当前美国前沿AI模型在国家安全审查与商业化推广之间的张力。据TechCrunch报道，OpenAI此次限量预览是应美国政府审查要求所做出的安排，旨在确保模型能力不被"外国对手"获取。

On June 26, OpenAI formally began a limited preview of the GPT-5.6 family. The series includes three models: the flagship Sol, the everyday balanced Terra, and the fast, cost-effective lightweight Luna. Unlike previous public launches, OpenAI did not open the models up immediately but, at the U.S. government’s request, granted limited preview access only to a small number of “trusted partners.” In its announcement, OpenAI disclosed it had demonstrated the models’ capabilities and launch plan to the U.S. government prior to release, with the first batch of approved enterprises numbering around 20, and access for some customers requiring case-by-case U.S. government approval.

OpenAI simultaneously emphasized it hopes this restrictive release will not become a long-term default and plans to fully open GPT-5.6 Sol, Terra and Luna in the coming weeks. Industry observers read this as a compromise by OpenAI under government regulatory pressure, reflecting the current tension between national-security review and commercial promotion for cutting-edge U.S. AI models. According to TechCrunch, the limited preview is an arrangement made in response to U.S. government review requirements aimed at ensuring model capabilities are not accessed by “foreign adversaries.”

来源 / Sources

• OpenAI limits GPT-5.6 rollout after government request, says restrictions shouldn’t be the norm（TechCrunch）
• OpenAI发布GPT-5.6系列模型，应美政府要求限量开放（华尔街见闻）
• OpenAI：开启下一代模型GPT-5.6系列的限量预览（财联社）

五、Anthropic最强网络安全模型Mythos 5获美国政府解禁

事件概要

6月27日，Anthropic在社交平台X发文宣布，已收到美国政府通知，公司最强网络安全模型Mythos 5可重新部署至一批负责运营和维护关键基础设施的美国机构。这是自6月12日以来，Anthropic与美国政府持续紧密合作的成果，标志着此前因出口管制指令引发的纠纷取得阶段性突破。Anthropic同时表示，将继续与政府合作扩大Mythos 5的访问范围，并推动Fable 5模型重新面向大众开放。

回顾纠纷起源，6月13日左右，Anthropic收到一份出口管制指令，文件援引国家安全要求，勒令其全面禁止所有外籍人士使用最新两款AI模型Fable 5与Mythos 5，无论外籍人士身处美国境内还是海外。此后Anthropic多名高层赴华盛顿与特朗普政府官员会面，力求化解纠纷。此次解禁被市场视为Mythos项目重回正轨的关键节点，对Anthropic筹备中的IPO进程（约9000亿美元估值预期）具有积极意义。

On June 27, Anthropic announced on social platform X that it had received notice from the U.S. government that its most powerful cybersecurity model Mythos 5 may be redeployed to a batch of U.S. institutions responsible for operating and maintaining critical infrastructure. This marks a phased breakthrough in the dispute triggered by earlier export-control directives, the result of close coordination between Anthropic and the U.S. government since June 12. Anthropic also stated it will continue working with the government to expand Mythos 5 access and bring the Fable 5 model back to general public availability.

Recapping the dispute’s origin: around June 13, Anthropic received an export-control directive invoking national-security requirements to ban all foreign nationals from using its two newest models, Fable 5 and Mythos 5, regardless of whether those individuals were inside or outside the United States. Subsequently, multiple senior Anthropic executives traveled to Washington to meet Trump administration officials to seek resolution. This unblocking is viewed by the market as a critical node for Mythos returning to track, and carries positive significance for Anthropic’s planned IPO process (with a market expectation of roughly USD 900 billion valuation).

来源 / Sources

• Anthropic：Mythos 5获准向美国机构重新部署（财联社）
• Anthropic"闯关"成功：Mythos 5获美政府解禁，9000亿美元AI独角兽上市前迎转机（华尔街见闻）
• 财联社6月27日电：Anthropic表示，今日收到美国政府通知（腾讯新闻）

六、华为联合中国电信发布星河AI网络Token化方案，AI-Centric目标网架构亮相MWC上海

事件概要

6月24-25日，2026 MWC上海期间，华为以"迈向全面智能化"为主题，携"业网算协同创新，共赢Byte+Token双增长"主线亮相，与中国电信联合发布星河AI数据中心网络方案的升级版本，全面构建以Token为中心的安全智联底座。华为ICT Marketing与解决方案销售总裁刘康在"升级智能化网络，打造可保障连接，共筑Token优质体验"高峰论坛上表示，AI产业正式迈入规模化落地阶段，交互体验成为各类智能服务的核心竞争力，智能体、多模态应用爆发催生海量双向实时Token交互场景。

星河AI网络方案依托网算存协同能力，NPU与存储直通使传输带宽提升8倍，Token生产效率提升2-5倍。华为同步推出以Token为中心的系列解决方案及45款华为坤灵数通新品，覆盖商业与分销市场。同期，中国电信国际合作伙伴大会在上海召开，发布迭代升级的"OneGrowth 2026：共启AI Token时代新生态"战略，并揭牌组建AI Token全球服务生态联盟，聚合全球通信头部企业、优质生态伙伴及海内外重磅客户。IDC预测，Token消耗每年将增长超过34倍，数据中心正加速向"Token工厂"转型。

During MWC Shanghai 2026 (June 24–25), Huawei took the stage with the theme “Marching Toward Comprehensive Intelligence” and the main line “Business-Network-Compute Co-Innovation, Winning Byte + Token Dual Growth,” jointly releasing with China Telecom an upgraded version of the Xinghe AI Data Center Network solution to comprehensively build a Token-centric secure intelligent-connectivity substrate. Liu Kang, President of Huawei ICT Marketing & Solution Sales, said at the “Upgrade Intelligent Network, Build Guaranteed Connections, Co-Create a Quality Token Experience” summit that the AI industry has formally entered a stage of scaled implementation, with interactive experience becoming the core competitiveness of intelligent services, and that the explosion of agents and multimodal applications is creating massive bidirectional real-time Token interaction scenarios.

The Xinghe AI Network solution leverages network-compute-storage coordination, with NPU-to-storage direct connection lifting transmission bandwidth by 8x and Token production efficiency by 2–5x. Huawei simultaneously launched Token-centric series solutions and 45 Huawei Qianpeng>来源 / Sources

• 华为助力全球运营商以网为基，以智为翼，业网算协同创新，迈向Token经营新征程（华为官方）
• 共启AI Token时代新生态 中国电信国际合作伙伴大会举行（财联社）
• 华为，AI新动向：华为将支撑中国电信联合行业伙伴完善AI体验标准（腾讯新闻）

行业数据速览

媒体人解读

2026年6月最后一周，全球AI产业呈现"硬件竞速+组织扩张+合规博弈"三条主线交织的格局。从硬件层面看，OpenAI与博通联合发布Jalapeño、英伟达股东会宣告"AI工厂时代"，本质上是同一信号的两个投射：算力即权力，谁掌握专用推理芯片与超大规模Token工厂，谁就掌握下一代AI应用的成本结构与议价权。Jalapeño以50%成本优势对标Blackwell，意味着头部模型厂商自研芯片已从"防御性选项"升级为"战略性必需"，未来12-18个月我们或将看到更多厂商加入自研ASIC的浪潮。

从组织层面看，DeepSeek宣布全部门规模至少翻倍，是一个被市场低估的标志性事件。这家公司过去两年以"小而美"的150人团队、拒绝外部融资的姿态封神，如今开启大规模招聘、构建完整商业化组织、向Harness（智能体基础设施）等应用层延伸，标志着AGI竞赛进入"算力×组织×生态"综合实力比拼的下半场。崔添翼团队负责人公开"非常缺人"的喊话，更从侧面印证了智能体基础设施赛道的供需失衡。

从合规层面看，OpenAI GPT-5.6限量预览与Anthropic Mythos 5获解禁形成鲜明对照：前者是政府审查前置化趋势的体现，后者是政府与企业协商化博弈的产物。两种模式最终会走向何种均衡，将深刻影响美国AI产业的全球竞争力，值得长期跟踪。

The final week of June 2026 sees the global AI industry characterized by three intertwined main lines: a hardware race, organizational expansion, and regulatory wrestling. On the hardware front, OpenAI and Broadcom’s Jalapeño and NVIDIA’s “AI Factory Era” declaration are essentially two projections of the same signal — compute is power, and whoever masters dedicated inference chips and ultra-large-scale Token factories controls the cost structure and bargaining leverage of next-generation AI applications. Jalapeño’s 50% cost advantage over Blackwell means self-developed chips have moved from a “defensive option” to a “strategic necessity” for top model vendors; in the next 12–18 months we can expect more players to join the custom-ASIC wave.

On the organizational front, DeepSeek’s announcement that it will at least double every department is a symbolic event the market has undervalued. For two years the company has been mythologized as a “small and beautiful” 150-person team that refused external financing; now it is launching large-scale hiring, building a complete commercial organization, and extending toward applications like Harness (agent infrastructure), signaling that the AGI race has entered a second half of “compute × organization × ecosystem” comprehensive strength contests. Harness team lead Cui Tianyi’s public call that “we are severely short-staffed” further confirms the supply-demand imbalance in the agent-infrastructure track.

On the regulatory front, OpenAI’s limited preview of GPT-5.6 and Anthropic’s unblocking of Mythos 5 form a sharp contrast: the former embodies the trend of front-loaded government review, while the latter represents the product of negotiated bargaining between government and enterprise. What equilibrium these two models ultimately converge to will profoundly affect the global competitiveness of the U.S. AI industry and is worth tracking over the long term.

参考来源

英文来源

• TechCrunch - Startup and Technology News
• OpenAI limits GPT-5.6 rollout after government request（TechCrunch首页推荐）
• Google DeepMind Official Site

中文来源

• AI周报|DeepSeek开启大规模招聘；OpenAI发布首款定制AI推理芯片（第一财经）
• DeepSeek开始造富（虎嗅网）
• OpenAI和博通发布AI芯片，旨在更快、更经济地运行模型（钛媒体）
• OpenAI与博通联合发布首款自研AI推理芯片Jalapeño（财联社）
• 英伟达股东大会释放产业信号：物理AI将成为下一轮增长核心浪潮（华尔街见闻）
• 英伟达CEO黄仁勋股东大会放言：本轮AI基建周期长达数十年（华尔街见闻）
• OpenAI发布GPT-5.6系列模型，应美政府要求限量开放（华尔街见闻）
• OpenAI：开启下一代模型GPT-5.6系列的限量预览（财联社）
• Anthropic：Mythos 5获准向美国机构重新部署（财联社）
• Anthropic"闯关"成功：Mythos 5获美政府解禁（华尔街见闻）
• 华为助力全球运营商以网为基迈向Token经营新征程（华为官方）
• 共启AI Token时代新生态 中国电信国际合作伙伴大会举行（财联社）
• AI行业晨报 2026-06-26（搜狐）

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。

发布日期：2026年06月28日

📚 相关阅读

• D60: 给读者的信：你的下一步是什么？
• 开发中常用素材格式完整清单
• D47 《数据治理：从混乱到可控》：防护体系的「治理」篇
• 一人公司（OPC）典型案例与商业模式研究报告
• 主流 Claw 类产品技术与商业差异全景报告
• 常见误区：管理者最容易踩的 5 个坑
• 团队 AI 成熟度自评模型