AI科技热点日报 | 2026年7月3日

文章目录

[AI科技热点日报 | 2026年7月3日](#AI科技热点日报 | 2026年7月3日)
- [📌 今日摘要](#📌 今日摘要)
- [一、AI 芯片与算力竞赛：Anthropic 联手三星推进定制芯片](#一、AI 芯片与算力竞赛：Anthropic 联手三星推进定制芯片)
- - 事件概要
  - [来源 / Sources](#来源 / Sources)
- [二、AI 云服务市场洗牌：Meta 拟对外销售过剩 AI 算力](#二、AI 云服务市场洗牌：Meta 拟对外销售过剩 AI 算力)
- - 事件概要
  - [来源 / Sources](#来源 / Sources)
- [三、主权 AI 加速：韩国政府拟用 AI 税收红利投入本国算力](#三、主权 AI 加速：韩国政府拟用 AI 税收红利投入本国算力)
- - 事件概要
  - [来源 / Sources](#来源 / Sources)
- [四、监管与资本：OpenAI 提议向美国主权基金捐赠 5% 股权](#四、监管与资本：OpenAI 提议向美国主权基金捐赠 5% 股权)
- - 事件概要
  - [来源 / Sources](#来源 / Sources)
- [五、中国大模型动态：DeepSeek V4 上线在即，服务端出现波动](#五、中国大模型动态：DeepSeek V4 上线在即，服务端出现波动)
- - 事件概要
  - [来源 / Sources](#来源 / Sources)
- 行业数据速览
- 媒体人解读
- 参考来源
- - 英文来源
  - 中文来源
- [📚 相关阅读](#📚 相关阅读)

AI科技热点日报 | 2026年7月3日

综合整理自 The Verge、TechCrunch、Reuters Technology、Wired、MIT Technology Review、钛媒体、36氪、新华网科技、虎嗅网、雷锋网、IT之家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。

说明：因搜索引擎对若干原文页面返回 4xx/反爬，文末"参考来源"采用本日报检索期内可访问的镜像与转载来源；正文中英文叙述对应国际/国内原始报道。

📌 今日摘要

2026 年 7 月 3 日，全球 AI 行业呈现"算力自研 + 云服务重构 + 主权 AI 提速"三条主线并行格局。Anthropic 已被多家外媒证实与三星电子就定制 AI 芯片进入早期洽谈，并考虑采用三星 2 纳米制程，意在复制 OpenAI 自研芯片路径；Meta 正筹划对外销售 AI 算力的云基础设施业务，正式与 AWS、Azure、Google Cloud 同台对垒。政策与资本端，韩国政府准备将三星、SK 海力士带来的 AI 税收红利投入本国"主权 AI"建设，OpenAI 首席执行官 Sam Altman 被披露已向白宫提议向美国主权财富基金捐赠约 5% 股权，获白宫国家经济委员会主任公开支持。国内侧，DeepSeek 网页与 API 在 7 月 2 日下午出现明显性能下降，并波及"DeepSeek V4 正式版 7 月中旬上线、引入峰谷定价"的预期管理。

On July 3, 2026, the global AI industry moved along three parallel tracks: in-house silicon, the reshaping of cloud services, and the acceleration of sovereign-AI agendas. The Information, citing people familiar with the matter, reported that Anthropic has begun early-stage talks with Samsung Electronics on manufacturing a custom AI chip and is weighing Samsung's 2nm process. Meta is preparing a cloud-infrastructure business to sell AI compute to outside customers, putting it in direct competition with AWS, Azure and Google Cloud. On the policy and capital side, the South Korean government plans to channel the tax windfall from Samsung and SK Hynix into a homegrown sovereign-AI program, while OpenAI's Sam Altman has been disclosed as proposing to contribute roughly 5% of equity to a US sovereign wealth fund --- a plan publicly endorsed by NEC Chair Hassett. Domestically, DeepSeek's web and API services showed noticeable performance degradation on the afternoon of July 2, casting a shadow over its July-mid rollout of V4 with peak/valley pricing.

一、AI 芯片与算力竞赛：Anthropic 联手三星推进定制芯片

事件概要

7 月 3 日，《The Information》援引三位知情人士率先披露，Anthropic 已启动自研 AI 芯片的早期工作，并与三星电子就潜在制造合作展开洽谈，目的是效仿 OpenAI 的路线，把对昂贵算力供应链的控制权握在自己手里。报道称 Anthropic 正考虑采用三星 2 纳米制程及先进封装技术，目前项目仍处于规划阶段，尚未敲定芯片功能、算力规格与服务器部署方案，公司近期还招募了前 OpenAI 自研芯片项目核心工程师 Clive Chan，以加强内部设计能力。Anthropic 官方回应称，短期内仍将依托英伟达、谷歌、亚马逊的多元化算力体系支撑业务发展，并未否认与三星的合作传闻。

The Information, citing three people familiar with the matter, reported on July 3 that Anthropic has begun early work on its own AI chip and is in talks with Samsung Electronics over a potential manufacturing partnership --- a move that mirrors OpenAI's playbook of bringing its most expensive compute stack in-house. According to the report, Anthropic is evaluating Samsung's 2nm process and advanced packaging, and the project remains at the planning stage with no finalized decisions yet on functionality, performance targets, or server deployment. Anthropic has also hired Clive Chan, a former core engineer on OpenAI's in-house chip project, to strengthen its design team. In its official response, Anthropic said it will continue to rely on a diversified compute base spanning NVIDIA, Google and AWS in the short term, while declining to deny the Samsung discussions.

来源 / Sources

二、AI 云服务市场洗牌：Meta 拟对外销售过剩 AI 算力

事件概要

彭博社 7 月初报道，社交媒体巨头 Meta 正酝酿一项新的云基础设施业务，向外部客户开放其 AI 算力和模型访问权限，正式与亚马逊 AWS、微软 Azure、谷歌 Cloud 同台竞争。Meta 近年来持续加码自建数据中心，本轮举动意在把过剩算力转化为面向第三方的大模型推理与企业 AI 服务收入，对外口径接近"AI 时代的公用云"。此前，Meta 已经在 4 月与 CoreWeave 签下约 210 亿美元、截至 2032 年的 AI 云容量长单，进一步坐实其"既要自用、也要对外卖算力"的双轨战略。

Bloomberg reported earlier this month that Meta is preparing a new cloud-infrastructure business that would open access to its AI computing capacity and models to outside customers, taking direct aim at AWS, Microsoft Azure and Google Cloud. Meta has spent recent years aggressively building out its own data centers; the new push effectively turns surplus capacity into third-party inference and enterprise AI revenue, packaged as "public-cloud for the AI era." Meta had already anchored the strategy in April with an approximately $21 billion AI-cloud capacity agreement with CoreWeave running through 2032, locking in its dual-track stance of running AI internally while selling compute externally.

来源 / Sources

三、主权 AI 加速：韩国政府拟用 AI 税收红利投入本国算力

事件概要

7 月 3 日，《首尔新闻》《韩国经济日报》等韩媒综合报道，韩国政府计划利用 AI 浪潮带来的税收红利加速本国"主权 AI"建设。三星电子、SK 海力士及相关上下游厂商在 HBM、AI 内存与代工环节赚得盆满钵满，韩国政府由此获得一笔可观的额外税金收入，意图把这部分财政增量投向国家级 AI 算力底座与本土基础模型研发。背景是韩国近期面临 Anthropic、OpenAI、Google 等海外巨头加速本地化的压力，韩国政府希望在监管与算力两端同时下注，确保关键基础设施不全部依赖海外模型与海外芯片。

South Korean outlets Seoul News and Korea Economic Daily reported on July 3 that the South Korean government plans to use the tax windfall from the AI boom to accelerate the country's "sovereign AI" build-out. The surge in profits at Samsung Electronics, SK Hynix and their upstream/downstream partners --- driven largely by HBM, AI memory and foundry demand --- has generated a sizable extra-tax cushion, and Seoul now wants to steer that incremental fiscal capacity into national AI compute infrastructure and domestic foundation models. The move comes amid intensifying pressure from overseas players such as Anthropic, OpenAI and Google deepening their local presence, with Seoul aiming to hedge across both regulation and compute so that critical infrastructure does not depend entirely on foreign models and chips.

来源 / Sources

四、监管与资本：OpenAI 提议向美国主权基金捐赠 5% 股权

事件概要

7 月 2 日，《金融时报》援引两位知情人士披露，OpenAI 首席执行官 Sam Altman 已向美国政府提议，向一个拟议中的美国主权财富基金捐赠 OpenAI 约 5% 股权，并希望 Anthropic、Google、Meta 等其他主要 AI 公司同样出让约 5% 股权，注入类似阿拉斯加永久基金的载体，由政府持有并惠及公众。按 OpenAI 约 8520 亿美元估值，5% 股权价值约 426 亿美元，相关谈判仍处早期，正式落地可能需要国会立法授权。7 月 3 日早间，白宫国家经济委员会主任哈塞特（Hassett）公开确认美国政府正与大型 AI 公司讨论收益分享机制，OpenAI 的 IPO 时间表也可能因此推迟至明年。

The Financial Times, citing two people familiar with the matter, reported on July 2 that OpenAI CEO Sam Altman has proposed to the US government that OpenAI contribute roughly 5% of its equity to a proposed US sovereign wealth fund, and that Anthropic, Google and Meta be encouraged to contribute similar stakes into an Alaska Permanent Fund--style vehicle held by the government for public benefit. At OpenAI's approximately $852 billion valuation, the 5% stake is worth around$ 42.6 billion; the talks remain preliminary and formal implementation would likely require congressional authorization. On the morning of July 3, White House NEC Chair Kevin Hassett publicly confirmed that the US government is in discussions with large AI companies on benefit-sharing, and OpenAI's IPO timeline could be pushed to next year as a result.

来源 / Sources

五、中国大模型动态：DeepSeek V4 上线在即，服务端出现波动

事件概要

7 月 2 日下午至 7 月 3 日上午，陆续有用户在社交平台反馈 DeepSeek 网页与 API 服务出现性能下降，上传文档时频繁遇到"服务器繁忙"提示。DeepSeek 官方未披露根因，市场普遍将这次波动与"DeepSeek V4 正式版 7 月中旬上线、首推峰谷定价"的准备动作联系到一起------高峰时段 API 价格将为平时两倍，平时价位与之前 V4 API 持平。叠加 6 月 30 日已官宣的功能优化与定价调整，外界普遍预期 V4 正式版将在中旬前后平稳切换，但本次服务端波动让"模型迭代 + 资源重排"这一窗口期的稳定性受到市场额外关注。

From the afternoon of July 2 through the morning of July 3, users reported on social platforms that DeepSeek's web and API services showed noticeable performance degradation, with frequent "server busy" prompts when uploading documents. DeepSeek has not disclosed the root cause; the market largely links the wobble to the company's preparation for the official DeepSeek V4 launch in mid-July alongside its first peak/valley pricing scheme, under which peak-hour API pricing will be 2× the standard rate while off-peak pricing matches the current V4 API. Together with the feature and pricing adjustments already announced on June 30, V4 is still broadly expected to roll out smoothly mid-month, but the latest service blip has put the stability of this "model-iteration + resource-rebalancing" window under extra scrutiny.

来源 / Sources

行业数据速览

指标	数据	备注
Anthropic 估值	约 1.2 万亿美元	6 月以来市场口径，公司未官宣
OpenAI 估值	约 8520 亿美元	用于测算 5% 股权对应金额
OpenAI 拟捐赠 5% 股权价值	约 426 亿美元	8 520 亿 × 5% 估算
Meta--CoreWeave 长单金额	约 210 亿美元	截至 2032 年的 AI 云容量协议
三星 2 纳米制程	已在试产爬坡	Anthropic 在评估的合作工艺节点
DeepSeek V4 上线窗口	2026 年 7 月中旬	官方口径，正式版 + 峰谷定价
DeepSeek V4 峰值/平价倍率	2×	API 高峰时段价格翻倍
韩国 AI 内存相关税收增量	未披露具体金额	来自三星电子、SK 海力士

媒体人解读

今天的几条新闻放在一起看，本质上都在回答同一个问题------当 AI 变成像电力一样的"通用目的技术"之后，供给侧的话语权该往哪儿落？ Anthropic 谈三星，是要补上"算力自研"这一课；Meta 做对外 AI 云，是把"自建基础设施"做成新的营收；韩国政府用 AI 税做主权 AI，是把"算力税基"折算成国家战略资产；OpenAI 想捐 5% 股权，是在监管压力到来之前主动交出一部分控制权。这四条都不是孤立的"产品发布"，而是大模型公司在算力价格、监管预期、地缘政治三条线下的一次集中表态。

在产品层面，今天唯一让 C 端用户真正感知到的，是中国侧 DeepSeek 在版本切换窗口期出现的一次服务端波动。它提醒我们一件事：所谓"大模型已经稳了"，其实是特定算力、特定入口、特定 prompt 区间下的稳态。任何一次模型升级 + 计费重构 + 入口切换，都会重新把这层稳态撕开。对开发者而言，今天的实操含义只有两条------高峰期跑长 prompt 时段的脚本，要么切到 DeepSeek V3.2 兜底，要么把峰值任务排到谷价区间 ；对 OpenAI、Anthropic 这类海外 API 的依赖度，趁机再往下调一档，因为无论 5% 股权捐赠还是 Anthropic 自研芯片，短期内都会让他们的定价权、可用区域、合规口径出现新的不确定。

放到更长时间尺度看，今天全球 AI 行业真正的底层信号是："国家队 + 巨头 + 替代芯片" 三股力量同时进场。韩国想做主权 AI、中国做大模型 + AI 终端、Anthropic 和 Meta 各自往硬件侧走，每条路径的回报周期都在 3-5 年以上。这意味着 2026 年下半年不会是"哪家模型又刷榜"的叙事，而是"哪几家能把基础设施账算平"的叙事；C 端开发者短期看到的红利，仍来自 DeepSeek 峰谷定价、Claude Sonnet 5 这类更低门槛的 API，以及 Meta AI 云如果真的对第三方开放之后带来的推理价格二次下行------这些窗口期，才是我们今天值得抓住的东西。

Today's headlines, read together, are really one story: once AI becomes a general-purpose technology, where does supply-side leverage end up? Anthropic talks to Samsung to catch up on in-house silicon. Meta turns self-built infrastructure into an external cloud revenue line. Seoul uses an AI-generated tax surplus to bankroll a sovereign-AI program. OpenAI proposes to give up 5% equity before regulators force the issue. None of these are product launches; they're four simultaneous positioning moves by the model labs under pressure from compute economics, regulatory expectations, and geopolitics at the same time.

On the product side, the only thing C-end users actually felt today was the DeepSeek wobble during its version-switch window. It's a useful reminder that "LLMs have stabilized" is a statement about a specific compute stack, a specific entry point, and a specific prompt envelope --- and any major model upgrade, pricing refactor, or entry-point migration tears that equilibrium open again. The practical takeaways for developers are short: when running long-prompt jobs in peak hours on DeepSeek, either fall back to V3.2 or shift tasks into the valley-pricing window; and use this moment to dial API dependency on overseas providers like OpenAI and Anthropic down another notch, because the 5%-equity donation and the Anthropic-Samsung track alike introduce near-term uncertainty on pricing, regional availability and compliance posture.

Over a longer horizon, the deeper signal of today is that three forces --- nation-states, incumbents, and alternative silicon --- are all entering the field at once. Korea goes after sovereign AI, China doubles down on models and AI terminals, Anthropic and Meta each push further into the hardware side; each of these paths has a 3--5 year payback. That means the second half of 2026 is unlikely to be a "who topped the leaderboard" story --- it'll be a "who can make the infrastructure math work" story. For C-end developers, the near-term arbitrage is still real: DeepSeek's peak/valley pricing, lower-friction APIs like Claude Sonnet 5, and a potential second leg of inference-price compression if Meta's AI cloud opens up to third parties --- those are the windows worth leaning into today.

参考来源

受本日报检索期内可访问性限制，下列链接优先采用媒体官方域名/镜像以及新华网、IT之家、钛媒体、36 氪、虎嗅、雷锋网、网易科技等国内媒体页面；个别未被主要中文媒体完整转载的国际报道以原始媒体域名标注，请知悉。

英文来源

中文来源

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 7 月 3 日

AI科技热点日报 | 2026年7月3日

文章目录