AI科技早报 · 2026年6月29日 周一AI Daily Bulletin · Monday, June 29, 2026


AI free token/AI免费积分

今日核心看点:谷歌对Meta实施Gemini算力限购,硅谷巨头上演硬核算力争夺战;首届全球太空算力大会今日在北京启幕;海外数据中心扩张遭遇民意阻力,算力能源模式迎来重构

一、头条重点:谷歌收紧Gemini算力额度,Meta研发节奏被迫放缓

据英国《金融时报》独家消息,谷歌自今年3月起正式向Meta下达算力上限通知,对Meta调用Gemini大模型的接口设置每周封顶额度,即便Meta愿意提高服务费,也无法追加算力资源,限流规则延续至今没有调整空间。

供需底层矛盾:谷歌云积压未交付订单规模庞大,一季度云业务营收突破200亿美元,但算力硬件、数据中心建设周期跟不上市场爆发式需求,企业内部优先保障Gemini自研业务算力供给,大客户配额被动压缩;

商业博弈层面:Meta一边采购谷歌外部算力支撑社交、广告AI优化业务,一边全力自研通用大模型,属于直接竞品。谷歌通过配额管控,既缓解自身算力挤兑压力,也间接放缓对手技术迭代速度;

Meta应对动作:内部下发算力节约通知,要求研发人员优化token使用效率,多款新一代AI产品上线计划被迫延后,整个硅谷科技圈都在感受到算力稀缺带来的项目推进压力 。

二、全球算力产业链动态

硬件新品节奏:英伟达新一代Vera Rubin芯片下半年正式出货,性能较上代提升3.3倍;微软Maia 200、平头哥真武810E等自研推理芯片陆续落地,2026年推理算力占比首次突破66%,成为算力消耗主力;

上游材料变革:玻璃基板开始大规模导入AI芯片封装,解决传统有机基板散热、布线短板,英特尔已实现量产,台积电、三星加速产线建设,成为算力硬件上游关键增量赛道;

电力配套新模式:雪佛龙和微软签订20年专属供电协议,打造离网式算力园区,绕过公共电网漫长的审批周期;美国超300个地区出台AI数据中心新建限制,高能耗争议倒逼厂商自建能源体系 。

三、国内AI产业要闻

重磅产业会议今日举办:首届全球太空算力大会在北京中关村开幕,现场将揭牌北京太空算力创新中心,配套三年50亿产业资金,发布天地一体化算力产业路线图,主攻星载芯片、在轨液冷等核心技术;

地方AI生态落地:成都公布第二批省级AI产业链主企业名单,本地链主企业数量占全省92%,同步放出79项产业应用场景清单,覆盖制造、文旅、交通等领域 ;

产业政策导向:工信部明确2028年城域1毫秒算力圈覆盖率目标不低于75%,重点攻坚CPO、高速光电芯片;科创板扩容第五套上市标准,全面接纳优质AI硬科技企业登陆资本市场;

光器件产能释放:头部光通信企业海外工厂有源产线完成客户审核,下半年进入放量周期,CPO配套产品稳定交付,行业物料紧缺情况有望逐步缓解。

四、大模型与海外科技快讯

  1. OpenAI GPT-5.6处于有限预览阶段,联合博通打造定制推理芯片,强化硬件自研能力,摆脱外部算力依赖;

  2. 国内头部AI企业DeepSea宣布全员扩编,全业务板块规模翻倍,押注通用人工智能落地,布局AGI全链条研发;

  3. 美光锁定千亿级长期芯片供货协议,长协模式改变存储行业周期属性,高端算力配套存储资源持续紧俏 。

Today's Key Highlights: Google imposes Gemini compute quotas on Meta, igniting a fierce rivalry for computing power among Silicon Valley giants; the first Global Space Computing Conference opens in Beijing today; overseas data center expansion faces public opposition, prompting a shift in computing energy models.

Headline: Google restricts Gemini usage quotas, forcing Meta to slow down its research pace.

According to an exclusive report by the Financial Times, Google officially notified Meta in March to set weekly upper limits on the utilization of the Gemini large model API. Even if Meta wishes to increase its service fees, it cannot obtain additional computing resources, and the throttle rules have remained unchanged to this day.

Underlying Supply-Demand Conflict: Google Cloud has a massive backlog of unprocessed orders, with cloud business revenue exceeding $20 billion in the first quarter. However, the hardware construction cycle of chips and data centers is struggling to keep up with the explosive market demand. Internally, Google prioritizes computing power for self-research projects while involuntarily compressing the quotas of large clients.

Business Rivalry: Meta procures external computing power from Google to support its social media and advertising optimization businesses, while simultaneously researching its own large models, making it a direct competitor. By controlling quotas, Google alleviates its own power shortage while indirectly slowing the opponent's technological iteration.

Meta's Response: Internally, it has issued notices to save computing resources, requiring developers to optimize token usage efficiency. The launch plans for several new-generation AI products have been postponed, reflecting the pressure that computing scarcity is placing on project timelines across the entire Silicon Valley tech industry.

Global Computing Chain Dynamics

Hardware Product Rhythm: NVIDIA's next-generation Vera Rubin chip is set to ship in the second half of the year, with performance improving by 3.3 times compared to the previous generation. Microsoft's Maia 200 and Alibaba's True Wu 810E self-developed inference chips are being deployed. By 2026, the proportion of computing power consumed by inference will exceed 66%, becoming the primary driver of consumption.

Upstream Material Evolution: Glass substrates are being used on a large scale for AI chip packaging, solving the heat dissipation and wiring issues of traditional organic substrates. Intel has already mass-produced it, and TSMC and Samsung are accelerating production line construction, making it a key upstream growth area.

Power Supply Innovation: Chevron and Microsoft have signed a 20-year exclusive power supply agreement to build an off-grid computing park, bypassing the lengthy approval process of the public power grid. More than 300 regions in the United States have imposed restrictions on the construction of new AI data centers due to high energy consumption, forcing companies to develop their own energy systems.

Domestic AI Industry News

Major Industry Conference: The first Global Space Computing Conference is being held in Beijing, where the Beijing Space Computing Innovation Center will be unveiled. It comes with a 3-year, 5 billion yuan industry investment plan and a roadmap for integrated space-ground computing, focusing on satellite-borne chips and in-orbit liquid cooling.

Local AI Ecosystem Landing: Chengdu announced the second batch of provincial-level AI chain leaders, which account for 92% of the provincial total. It also released a list of 79 industry application scenarios covering manufacturing, tourism, and transportation.

Industry Policy Direction: The Ministry of Industry and Information Technology aims for a coverage rate of 1 millisecond inter-city computing rings to be no less than 75% by 2028, with a focus on CPO and high-speed optical chips. The Shanghai Stock Exchange has expanded its fifth listing standard to accommodate high-quality AI tech companies.

Optical Component Capacity Release: Leading optical communication companies have completed customer audits for active production lines in overseas factories, entering a period of high shipment volume in the second half of the year. The supply of CPO-related components is stable, indicating that the industry bottleneck is gradually easing.

Large Models and Overseas Tech Updates

OpenAI's GPT-5.6 is in limited preview, and the company is collaborating with Broadcom to develop custom inference chips, strengthening its self-research capabilities and reducing dependence on external computing power.

Domestic AI leader DeepSea announced a full-scale expansion, doubling its team size across all business segments, betting on the full-chain research and development of general artificial intelligence.

Micron has secured long-term chip supply contracts worth hundreds of billions, changing the cyclical nature of the storage industry. High-end computing support storage resources remain scarce.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry