xAI 把算力租给了 Anthropic：马斯克的算力帝国开始漏风

marsbitPublicado a 2026-05-07Actualizado a 2026-05-07

撰文:小饼,深潮 TechFlow

如果三个月前你告诉一个硅谷投资人,马斯克会把 xAI 最大的训练集群 Colossus 1 全部租给 Anthropic,他大概会笑出声。

毕竟在二月份,马斯克还在 X 上骂 Anthropic"恨西方文明",在三月份给 Anthropic 起了个外号叫"misanthropic"(厌人的)。这家公司在马斯克眼里几乎是政治正确 AI 的代名词,是和 OpenAI 一样必须被打倒的对手。

然后 5 月 6 日,Anthropic 和 SpaceX 联合宣布:Anthropic 将获得 Colossus 1 的全部算力,超过 22 万张英伟达 GPU,300 兆瓦的电力容量,一个月内交付完毕。Anthropic 明确表示,这批算力将直接用于改善 Claude Pro 和 Claude Max 订阅用户的服务体验。

马斯克在 X 上发了一条让所有人愣住的帖子:他说他过去一周和 Anthropic 的高管接触很深,"印象深刻","他们都很有能力,也都在认真做正确的事"。他甚至说 Claude"大概率会是好的(probably be good)"。

同一天,他宣布 xAI 将作为独立公司解散,改名为 SpaceXAI。

这是一次产能让渡

主流英文媒体把这件事写成"AI 算力共享的标志性事件",但他们漏掉了一个关键事实:

Colossus 1 是 xAI 最核心的训练设施,不是什么"备用产能"。

回顾一下时间线。Colossus 1 于 2024 年 9 月在孟菲斯落成,从动土到通电只用了 122 天,是数据中心建设史上的奇迹。这是 xAI 用来训练 Grok 3、Grok 4 的主力集群,也是马斯克"算力即权力"叙事的物理载体。它配置了超过 22 万张 GPU,包括 H100、H200 和最新的 GB200,集群规模在 2025 年底曾经是全球前三。

把这种规模的训练集群整体让给一家直接竞争对手,等价于台积电把 5 纳米产线的全部产能租给三星。这种事情在半导体行业从来没有发生过。任何一个看过周期的人都知道,这种动作只在一种情况下出现:自己用不完。

而 SpaceXAI 官方的说法是:Anthropic 的算力将"直接惠及 Claude Pro 和 Claude Max 订阅用户"。也就是说,Anthropic 拿这部分算力做的是推理(inference),给 Claude 的付费用户跑模型,给马斯克最讨厌的那个 AI 跑用户请求。

用"客户合作"四个字概括这件事是不准确的,Colossus 1 的实际控制权,某种意义上易主了。

Grok 的故事撑不起 Colossus 的体量

为什么会"用不完"?

最直接的答案藏在 Grok 的用户数据里。

根据 Similarweb 4 月发布的数据,Grok 在全球移动应用的日活(DAU)从 3 月的 1390 万,跌到 4 月的 1220 万,环比下滑 12.5%。在美国市场跌得更狠,从 140 万跌到 110 万,环比 -15.6%。一年前还是仅次于 ChatGPT 的全球第二大 AI 应用,4 月份已经跌到第五,被 Claude、Gemini、DeepSeek 依次超过。

同期 Claude 的 DAU 从 1600 万涨到 2300 万,环比 +44%。

这是一个非常残酷的对照:在 AI 应用普遍高速增长的 2026 年,Grok 是少数几个用户在流失的头部产品。原因不复杂,Grok 的核心场景一直绑定在 X(原 Twitter)平台内部,作为"实时搜索+辛辣点评"的工具型存在。但它在独立 App 和 Web 端始终没有形成像 Claude 那样的"工作流粘性"。Reddit 上不少用户抱怨 Grok 把图片和视频生成功能逐步移到付费墙后面,加上多个国家的监管调查、Apple 的封禁威胁,让它的增长引擎几乎熄火。

更要命的是 xAI 内部。

根据《Fast Company》4 月的报道,xAI 过去几个月有 80 多名员工离职,包括几位联合创始人。《金融时报》2 月的报道则提到马斯克持续向团队施加"不合理的技术指标压力",试图追赶竞争对手,这是典型的产品节节败退期的领导者反应。

把这两件事放在一起看,Colossus 1 为什么会有富余产能,答案就出来了:它当初是为一个比现在大得多的 Grok 准备的。

SpaceXAI 真正的难题在估值叙事

"Grok 需求不足"只是表层。

更深一层的逻辑是:马斯克需要一个新的故事来支撑 SpaceXAI 的 1.25 万亿美元估值。

回忆一下今年 2 月发生了什么。SpaceX 以全股票方式收购了 xAI,合并后实体估值 1.25 万亿美元,这是史上最大的并购案。在并购前,xAI 最近一轮融资是 1 月份的 Series E,200 亿美元、估值 2300 亿美元。把 xAI 塞进 SpaceX,本质上是用 SpaceX 火箭业务的现金流,去给 xAI 那个仍在每季度亏损 14.6 亿美元的烧钱黑洞续命。

但即便有 SpaceX 输血,SpaceXAI 仍然要面对一个尖锐的问题:它凭什么值这个钱?

OpenAI 最近一轮估值 8520 亿美元,ARR 约 240-250 亿美元,估值/收入比约 35 倍。Anthropic 正在谈的 9000 亿美元估值,ARR 是 300 亿美元,估值/收入比 30 倍。

xAI 呢?2025 年第三季度收入 1.07 亿美元,净亏损 14.6 亿美元。即便给 Grok 2026 年的预期收入按 20 亿美元乐观估计,SpaceXAI 那部分对应的估值/收入比也要远高于 OpenAI 和 Anthropic。换句话说,马斯克迫切需要给 SpaceXAI 讲一个新的现金流故事。靠 Grok 的用户增长讲不出来,靠企业 API 收入也讲不出来。

把 Colossus 租给 Anthropic 就是那个故事的开头。

它瞬间把 SpaceXAI 从一个"模型公司"重新定位成"AI 云基础设施提供商",一个有点像 CoreWeave、但规模和电力供给都更大的玩家。在叙事估值的世界里,云厂商比模型公司值钱。云厂商可以拿出长期合同和可预测的现金流,这是纯模型公司很难提供的东西。

加上 Anthropic 和 SpaceX 那个含糊其辞的"轨道算力中心"备忘录,双方同意"探索"在太空部署多吉瓦级别的 AI 数据中心,你就明白这一切的指向了。这是为 SpaceX 的 IPO 准备的一份新的资产负债表。火箭、星链、地面数据中心、轨道算力,全部打包成一个巨型基础设施故事。Grok 本身已经无关紧要,重要的是马斯克手里的 GPU、电力、发射工位。

马斯克态度 180 度大转弯的真正含义

在这个框架下,马斯克对 Anthropic 态度的剧变就有了另一种解释。

它是一笔交易。

Anthropic 给 SpaceXAI 的还有租金之外的东西,信用背书。让 Anthropic 公开背书 Colossus 1 的可用性、扩展性、运维质量,等于给 SpaceXAI 颁发了一张进入"算力基础设施俱乐部"的入场券。这个俱乐部的成员包括 AWS、GCP、Azure、CoreWeave。在这之前,xAI 在云服务市场的口碑约等于零,它过去只用算力训练自己的模型,从未真正做过对外的运营商业务。

而对 Anthropic 来说,这笔交易也极为划算。它正在以 9000 亿美元的估值融资,可能在 10 月份 IPO。它公开披露的需求是 5 吉瓦的训练算力,SpaceX 给的 300 兆瓦看起来不算大,但价值在于"立刻交付":一个月内通电,直接缓解 Claude 当下的推理压力。Anthropic 在 4 月份公开承认,Claude 因为"基础设施压力"在高峰时段"可靠性和性能"受到了影响。300 兆瓦的紧急产能,价值远超账面数字。

这是一笔双向的叙事交易。Anthropic 换到的是服务稳定性,SpaceXAI 换到的是估值故事。

谁让步了?马斯克自己确实低头了,他和老对手做了生意,还说了对方的好话。但在更深的层面上,让步的是 Grok。Grok 作为产品、作为模型公司、作为马斯克对抗 OpenAI/Anthropic 的旗舰武器,正在被悄悄地降级为 SpaceXAI 业务组合中的一项普通业务。Colossus 这种核心战略资产被腾出来给客户用,意味着马斯克已经不再把"自研模型"当成主战场。

在这个意义上,5 月 6 日是 Grok 作为"前沿模型公司"时代的终结。

AI 行业的拐点信号:产能开始向少数玩家集中

把视角再拉远一点,这件事的行业意义可能比我们当下能看到的更大。

整个 2024 和 2025 年,AI 算力市场是一个"全行业疯抢"的状态。OpenAI 抢、Anthropic 抢、xAI 抢、Mistral 抢、各国主权基金抢。GPU 是硬通货,数据中心选址是地缘政治议题,电力供给是国家战略问题。在这种全员紧缺的环境下,没有人会把自己的训练集群租给竞争对手,因为你今天租出去的每一个 GPU 小时,都可能是明天追赶对手时的关键算力。

而现在,xAI 做了。

这意味着 AI 算力市场正在出现第一道分化:头部模型公司(OpenAI、Anthropic、Google DeepMind)的算力需求仍在指数级增长,而二线及以下的模型公司开始出现产能松动。这种分化在所有产能扩张周期的中后期都会出现,从太阳能电池到电动车电池到比特币矿机,剧本几乎一模一样。早期所有人都缺,中期产能开始溢出二线玩家,末期一线玩家整合上下游,二线玩家要么转型基础设施服务商,要么被收购,要么死掉。

CoreWeave 是最好的对照。它原本是一家以太坊矿场,在 2018 年抓住 GPU 过剩的窗口转型 AI 云,2024 年 IPO 时市值 600 亿美元。它的存在本身就证明了"模型不行就转算力"这条路是可以走通的。SpaceXAI 正在重走这条路,版本更激进,除了卖地面算力,马斯克还要把算力卖到太空去。

AI 泡沫顶点的真正信号,可能就是二线模型公司纷纷转型云服务商。当一个行业的核心叙事从"我有最好的模型"切换到"我有最多的 GPU",通常意味着差异化竞争已经走到尽头。

值得记下的一个细节是:Colossus 1 所在的孟菲斯,xAI 在建设时为了赶进度,部署了几十台天然气燃气轮机用于供电,声称"临时使用"无需联邦许可。当地居民因为空气污染问题持续抗议,这件事直到现在仍未平息。

而现在,这些靠燃气轮机供电的 GPU,接下来会用来跑 Anthropic 的 Claude,一个在 AI 伦理和气候问题上立场最严谨的实验室之一。

更荒诞的是,Anthropic 和 SpaceX 在公告里"表达了兴趣",要在轨道上部署多吉瓦级别的 AI 算力。马斯克的逻辑是:地球上的电力和散热终将不够,AI 的未来在太空。

孟菲斯的燃气轮机和马斯克 PPT 里的轨道太阳能板之间,横着的是一个巨大的估值预期。Colossus 1 租给 Anthropic,是马斯克为这个估值预期讲述的第一个新故事。

xAI 从一个对抗者变成一个供应商,只用了三个月。下一个被重新定价的会是谁?

Lecturas Relacionadas

Jensen Huang's CMU Speech: In the AI Era, Don't Just Watch, Build

Jensen Huang, CEO of NVIDIA and a first-generation immigrant, delivered the commencement address to Carnegie Mellon University's class of 2026. He shared his personal journey from a humble background to founding NVIDIA, emphasizing resilience, learning from failure, and the responsibility that comes with leadership. Huang framed the present moment as the dawn of the AI revolution, a shift he believes is more profound than previous computing waves. He described AI as fundamentally resetting computing—moving from human-written software to machines that understand, reason, and use tools. This will create a new industry for generating intelligence and transform every sector. While acknowledging AI's potential to automate tasks and displace some jobs, Huang distinguished between the *tasks* of a job and its core *purpose*. He argued AI will augment human capability, not replace humans. The real risk, he stated, is not AI itself, but people being left behind by those who effectively use AI. He presented AI as a generational opportunity for massive infrastructure investment—in chip factories, data centers, energy grids, and advanced manufacturing—that could re-industrialize nations like the U.S. and bridge the digital divide by making computing and intelligent tools accessible to all. Huang called for a balanced approach: advancing AI safely and responsibly, establishing prudent policies, ensuring broad access, and encouraging universal participation. He urged the graduates not to fear the future but to engage with optimism and ambition, reminding them of CMU's motto, "My heart is in the work." His core message was clear: this is their moment to actively build and shape the AI-powered future, not merely observe it.

marsbitHace 14 min(s)

Jensen Huang's CMU Speech: In the AI Era, Don't Just Watch, Build

marsbitHace 14 min(s)

The Era Has Arrived Where Human Writers Must Prove They Are Not Machines

The article describes an era where AI-generated content is flooding the market, forcing human authors to prove they are not machines. It begins with the example of dozens of AI-written, error-ridden biographies of Henry Kissinger appearing on Amazon within hours of his death, a pattern repeated for other deceased celebrities and even living experts who find fraudulent books under their names. This spam content has exploded, with monthly new book releases on platforms like Amazon reaching 300,000 by late 2025. The issue spans genres, from suspiciously high proportions of AI-written teen romance and self-help books to dangerous, AI-generated foraging guides containing lethal advice. The platforms' automated review systems, designed to catch plagiarism and banned words, are ill-equipped to detect AI-generated text that avoids these pitfalls while being nonsensical or fraudulent. The problem has infiltrated traditional publishing. A major publisher, Hachette, had to recall a bestselling horror novel after AI detection tools suggested 78% of its content was machine-generated. An acclaimed European philosophy book was later revealed to be entirely written by AI under a fake author persona. In response, authors are fighting back. At the 2026 London Book Fair, 10,000 writers published a blank book titled "Don't Steal This Book" containing only their signatures—using emptiness as a protest weapon in an age of AI overproduction. Initiatives like the "Human Author Certification" program have emerged, ironically placing the burden on humans to prove their work is not machine-made. The article warns of a vicious cycle: AI-generated low-quality books pollute the data used to train future AI models, leading to "model collapse" and an ever-worsening flood of digital waste, eroding trust in publishing and devaluing human creativity.

marsbitHace 40 min(s)

The Era Has Arrived Where Human Writers Must Prove They Are Not Machines

marsbitHace 40 min(s)

Strategy CEO Highlights Scenarios Where Company Would Sell Bitcoin — Report

Strategy CEO Phong Le outlined potential scenarios for the company to sell portions of its massive Bitcoin holdings, a departure from its "Never Sell" philosophy. In a CNBC interview, Le stated sales could occur to finance dividend payments for its Perpetual Preferred Stock if doing so increases shareholder value, defined as a rise in "Bitcoin per share." He emphasized a data-driven approach, prioritizing math over ideology. Other reasons include when the company's book value is below market value or to capture tax benefits. As the largest corporate holder with 818,334 BTC (~$65 billion), Strategy has significant annual dividend obligations (~$1.5 billion). Le believes the Bitcoin market's high liquidity (~$60B daily volume) can absorb such sales without major price impact. Bitcoin currently trades around $80,840.

bitcoinistHace 51 min(s)

Strategy CEO Highlights Scenarios Where Company Would Sell Bitcoin — Report

bitcoinistHace 51 min(s)

Crypto Funds See Six Consecutive Weeks of Net Inflows, CLARITY Act Drives $858 Million Influx

Global cryptocurrency investment products recorded a net inflow of $857.9 million last week, marking the sixth consecutive week of positive flows and the largest single-week inflow since April 24, according to CoinShares' latest report. The primary catalyst was legislative progress on the U.S. CLARITY Act, specifically the final compromise text for stablecoin yield provisions, alongside Bitcoin's price breaking above $80,000. U.S.-listed products dominated with $776.6 million in inflows, a significant jump from the prior week's $47.5 million. European markets also saw increased participation. Bitcoin products led with $706.1 million in inflows, while short-Bitcoin products experienced their largest weekly outflow of 2026, indicating unwinding bearish positions. Major altcoins like Ethereum, Solana, and XRP all saw renewed positive flows, with Ethereum reversing its prior week's outflow to attract $77.1 million. Total assets under management (AUM) climbed to $160 billion.

marsbitHace 54 min(s)

Crypto Funds See Six Consecutive Weeks of Net Inflows, CLARITY Act Drives $858 Million Influx

marsbitHace 54 min(s)

Silicon Valley Wang Chuan: How Can You Not Feel Anxious When the Neighbor Lao Wang Earned 30x by Investing in Storage Stocks?

The author explores anxiety, attributing it to the amygdala's stress response, and suggests writing down worries to shift control to the rational prefrontal cortex, easing anxiety more effectively than mere reassurance. The piece then examines a source of financial anxiety in May 2026: a 150% surge in the Philadelphia Semiconductor Index, with storage stock SNDK skyrocketing 38x. The author analyzes the underlying logic behind this boom, tracing it to soaring valuations of AI startups like Anthropic (Anth). Anth's reported $400 billion ARR is scrutinized, revealing that cumulative revenue since founding is only about $108 billion, with the company still losing billions monthly and relying on continuous funding (having raised $723 billion). The author contrasts Anth's $1+ trillion valuation speculation with Berkshire Hathaway's fundamentals ($3.7 trillion revenue, $669 billion profit, $4 trillion cash, $1 trillion market cap), questioning the belief that a loss-making, funding-dependent startup could be worth more than the established giant. The narrative concludes by linking the massive valuations of Anth and OpenAI to a surge in AI industry capital expenditure (capex), which in turn fueled the storage sector's prosperity, setting the stage for a further discussion.

marsbitHace 1 hora(s)

Silicon Valley Wang Chuan: How Can You Not Feel Anxious When the Neighbor Lao Wang Earned 30x by Investing in Storage Stocks?

marsbitHace 1 hora(s)

Trading

Spot
Futuros
活动图片