社交媒体巨头Meta引发了广泛的国际讨论，全球市场热议人工智能算力是否步入过剩阶段。

Meta is reportedly formulating a cloud infrastructure business plan that would enable it to sell access to AI compute and models to external customers. The company not only intends to sell rights of access to various AI models hosted on its existing infrastructure but is also considering direct sales of the use of raw compute power.

消息一出，7月1日，Meta股价大涨8.8%。而美股半导体股票全线大跌，费城半导体指数暴跌超6%。存储芯片巨头美光、闪迪跌超10%。为Meta提供算力的新云厂商CoreWeave和Nebius分别暴跌14%和17%。紧张情绪迅速蔓延至全球，导致三星一日跌7%，SK海力士跌9%，韩国KOSPI指数一度触发交易暂停。

值得注意的是，就在Meta发出相关消息的之前，外媒已经报道了OpenAI通过软件优化将AI推理成本降低了一半。目前，推理占据了所有AI计算成本的三分之二，这意味着OpenAI将能够大幅减少为维持ChatGPT运行所需的GPU数量。

韩国存储巨头与政府共同宣布了总金额高达4800万亿韩元的中长期投资计划，其中约3700万亿韩元直接投入到了存储相关领域。部分解读认为，存储产能即将获得天量释放，导致供应严重过剩，从而终结了本轮存储涨价周期。

连跌两个交易日后，7月6日，美股半导体板块止跌回升，推动整体市场走强。截至收盘，标准普尔500指数涨0.72%，报7537.43点；纳斯达克综合指数涨1.12%，报26121.16点；道琼斯工业平均指数涨0.29%，报53055.91点。费城半导体指数收涨2.17%。

在此之前，AI芯片的繁荣发展一直基于同一个假设，即对算力的需求增速将保持在最高位。今年以来，微软、亚马逊、谷歌和Meta的总资本开支承诺已经飙升至7250亿美元。

One side, since the beginning of the year, the market's anxiety over AI concepts has kept expanding.

Meta opened its “magic box” at this moment, and the reality that was released was either the bursting of a bubble or a foretaste of the industry’s reshuffling?

硅谷科技巨头的“新旧世界”背后，算力过剩争论正通过现实情况揭晓其深层含义。

For Meta itself, with its long-term massive investments in AI, the decision to sell compute capacity may represent a long-anticipated necessity.

From October of last year, Meta has begun to preview its entry into the cloud computing industry. In the earnings call at that time, Meta's CEO Mark Zuckerberg (Mark Zuckerberg) mentioned that if Meta were to eventually face an excess of compute resources, the company would be able to make a portion of it available to third parties for use.

今年5月，在股东大会上，面对Meta是否有可能与进入云计算领域竞争的问题，扎克伯格回应称，这绝对在考虑范围内。

上海财经大学特聘教授胡延平对澎湃新闻记者表示，Meta转租算力，其背后原因和SpaceX转租Colossus算力给Anthropic和谷歌是相似的：两者都在AI前沿大模型竞赛中落后，在AI领域的科技、产业和市场地位都边缘化。

同时，流量与核心用户仍持续向GPT、Claude和Gemini集中，AI竞赛已导致科技产业以及市场出现明显分化。胡延平分析认为：“虽然都是科技巨头，但已经显著分化为新世界和旧世界。”

Meta is reportedly formulating a cloud infrastructure business plan that would enable it to sell access to AI compute and models to external customers. The company not only intends to sell rights of access to various AI models hosted on its existing infrastructure but is also considering direct sales of the use of raw compute power.

Since Meta's social platform, digital advertising business, and corresponding AI application ecosystem cannot absorb the large number of AI chips and computing clusters accumulated in advance through its substantial investments, the company must rent out old chips as soon as possible. Otherwise, the value of these substantial investments would be rapidly diminished.

SpaceX pioneered the practice of leasing compute clusters, which prompted Professor Tan Jianye of the School of Digital Media and Design Arts at Beijing University of Posts and Telecommunications to remark that Meta was merely “following suit.” This development marks a turning point in the industry: compute has formally transitioned from an accumulated internal asset into a rentable asset.

From a more positive perspective, if Meta can seize this opportunity, it will achieve a transformation of its business model.

一位深耕半导体领域的投资界人士对记者指出，这是Meta一次商业蜕变的叙事，将AI算力出租给第三方，本质上也是在为公司筛选未来潜在的合作乃至收购对象，同时给华尔街资本一个交代。

Jim Cramer, the well-known economic commentator from the United States, holds the same optimistic view toward this development: “At present, Meta is a company that relies excessively on advertising from consumers, whereas cloud infrastructure directed toward enterprises represents a superior business model that is not time-sensitive, which helps multiply Meta’s remarkably low price-to-earnings ratio.”

谭剑强调，“算力过剩”的说法完全把市场读反了。研究机构杰富瑞的调查数据显示，Meta内部算力利用率约65%，闲置的35%是因为大模型从训练转到推理时的物理规律所致，而非经营失败：“菜市场大白菜卖不动时，摊主敢提价吗？有人愿意溢价买，恰恰说明算力基建还是缺的。”

AI infrastructure has ushered in the next act, and whether it can transition from single points to multiple points for a breakthrough remains to be seen.

本次Meta所引发的市场震荡，更像是在长期压力下的一种过激反应。

国内行业视角来看，算力目前尚未达到供过于求的程度。一位国产服务器厂商的高管对澎湃新闻记者表示，目前，国内并没有出现算力过剩的情况，各家还在明确抢购GPU。

国联民生证券在近期研报中写道，Meta入云不是算力过剩的信号，而是算力从“资本开支承付”走向“商业价值变现”的转折点。算力租赁市场延续高景气，国内算力仍供不应求。

胡延平指出，全球半导体和存储芯片股价大跌的主要原因是前期炒作推高了股指本身存在回调动能，Meta的转租以及三星海力士的扩产只是加速了市场的反应。大模型训练与推理对算力存储网络模块的需求实际上还在大幅度增长算力总体并未过剩。

面对市场近日的表现，券商机构纷纷劝投资者保持冷静。美银最新预计显示，到2027年，全球云计算与AI设施资本支出将达到1.5万亿美元。

野村证券分析认为，出租算力是Meta解决峰值冗余的必然选择，市场的核心误判在于将一家企业的算力富余等同于全行业需求见顶。

Jay Yoon, a commentator, pointed out that computing power is still severely insufficient, which represents a computing power allocation issue. He indicated that partial specific players, who have built up idle chips due to over-construction based on their own model requirements, do not signify that the industry as a whole is in surplus. Meta’s entry into the field may instead, through competition, lower computing power prices and expand the accessibility of AI applications, thereby enlarging the computing power pie from a total perspective rather than reducing it.

国泰海通证券指出，从产业趋势看，头部科技企业通过外部租赁实现GPU集群利用率提升，平滑折旧压力并验证算力现金流属性，反而说明AI算力正在成为可计量、可调度、可交易的基础设施。

So, after this round of pain subsides, what remains as the next core catalyst propelling the upward breakthrough of large-scale AI infrastructure?

胡延平表示，AI大基建的驱动力已经发生了阶段性位移：早期主要由训练拉动，去年和今年则由推理服务全面驱动。而接下来资本开支继续向上突破的动能，将呈现出多点迸发而不是单点引爆的生态化态势，主要来自多模态和物理AI，包括智能驾驶、具身智能、泛在智能设备等。

谭剑梳理了AI基建“下一幕”的具体落地节奏：第一层是Agent的规模化，算力开支就从营销费变成人力成本；第二层是电力和国家级算力协同基建，第三层是如玻璃基板这一类的先进封装技术期权。他补充道：“新一代芯片架构是下一轮的燃料，不是这一轮的火柴，看方向可以有信念，看节奏得有耐心。”

来源：巨头开始卖算力，打开了什么魔盒 | 中国能源网