CN — LARRY ROMANOFF: 揭穿埃隆·马斯克——第18部分 — 欺诈更新——xAI与Grok — Debunking Elon Musk – Part 18 — Fraud Update – xAI and Gro

揭穿埃隆·马斯克——第18部分 — Debunking Elon Musk – Part 18

欺诈更新——xAI与Grok –– Fraud Update – xAI and Grok

作者：拉里·罗曼诺夫 — By Larry Romanoff

翻译： 珍珠

 

CHINESE     ENGLISH      ESTONIAN      ROMANIAN

 

 

In part 12 of this series [1], I dwelt at some length on Musk’s xAI and Grok, exposing some of the flaws and contradictions – and the dangers – of this creation of Elon Musk. This essay is an addendum to Part 12, exposing yet more of the apparent fraud underlying xAI and Grok. There were many overwhelmingly false statements made, some crucial details omitted surrounding xAI’s development, and much media hype that supported widespread public misunderstandings of the reality. Musk’s statements reflect a pattern of empty claims and wild exaggeration. Notably absent are any modest claims. Every descriptor is an extreme superlative (“best”, “most”, “first”, “smartest”, “most powerful”, “unmatched”) which fit Musk’s pattern of (1) hyperbole, exaggeration, and overpromising, and (2) using superlatives to mask operational shortcomings and ethical contradictions.

在本系列文章的第12部分[1]中，我详细探讨了马斯克的xAI和Grok，揭露了这一由埃隆·马斯克（Elon Musk）创造的事物的一些缺陷、矛盾以及潜在的危险。本文是第12部分的补充，进一步揭露了xAI和Grok背后更多的明显欺诈行为。其中存在许多明显错误的陈述，围绕xAI的开发遗漏了一些关键细节，而媒体的大肆炒作则加剧了公众对现实的普遍误解。马斯克的言论反映出一种空洞的宣称和夸张的模式。值得注意的是，其中没有任何谦虚的声明。每一个描述词都是极端的最高级词汇（“最好”、“最多”、“第一”、“最聪明”、“最强大”、“无与伦比”），这与马斯克的模式相符：（1）夸张、夸大其词和过度承诺；（2）使用最高级词汇来掩盖运营上的缺陷和伦理上的矛盾。

 

One notable element in all the promotional hype about xAI and Grok, is that Musk’s primary driving force is clearly not the development of AI, but winning his one-man war with Sam Altman by proving that “mine is bigger than yours”. This quickly becomes obvious from many of his statements. Musk claimed that his Grok was the “maximum truth-seeking AI, unlike ChatGPT”. Other claims were that “Grok will crush ChatGPT”; “Grok will make OpenAI obsolete”. He said further that, after training on Colossus, Grok would “dominate all metrics” – specifically targeting GPT-4.

在所有关于xAI和Grok的宣传炒作中，一个值得注意的点是，马斯克的主要驱动力显然不是开发人工智能，而是通过证明“我的比你的大”来赢得他与萨姆·奥特曼（Sam Altman）之间的个人战争。从他的许多声明中，这一点很快就变得显而易见。马斯克声称，他的Grok是“最追求真理的人工智能，不同于ChatGPT”。其他说法还有“Grok将碾压ChatGPT”；“Grok将使OpenAI过时”。他进一步表示，在Colossus上进行训练后，Grok将“在所有指标上占据主导地位”——特别是针对GPT-4。

 

It is true that many of Elon Musk’s statements could be categorised as exaggerated marketing hype or stemming from a juvenile enthusiasm. But equally, many of Musk’s statements are outright lies of consequence that would serve only to deceive customers and subscribers in the general public and also any potential investors. And those statements translate to fraud, usually with criminal implications.

诚然，埃隆·马斯克的许多言论可能被归类为夸大其词的营销炒作，或是源自于幼稚的热情。但同样，马斯克的许多言论也是彻头彻尾的谎言，其后果只会是欺骗广大公众中的客户和订户，以及任何潜在的投资者。而这些言论则构成了欺诈行为，通常还会涉及刑事问题。

 

This is an important distinction. If an EV manufacturer claims to have “the best range of all competitors”, we might dismiss this as marketing and would not likely purchase a vehicle on that claim alone. But when Elon Musk advertised the range of his Tesla cars as around 500 Kms when it was admitted internally that the known maximum range was only around 350 kms., hundreds of thousands of people purchased those EVs based on that false claim. [01] That constitutes criminal fraud. And, if we want proof of fraudulent intent, we further discover that the sales contract for Tesla cars states in the fine print that, in case of dispute, we legally commit ourselves to not suing Tesla, either individually or in a class action, but instead commit to “private binding arbitration”. The only possible results of such a clause are to limit consumer recourse and avoid public disclosure and accountability. I am told those clauses also contain gag orders.

这是一个重要的区别。如果一家电动汽车制造商声称其拥有“所有竞争对手中最长的续航里程”，我们可能会认为这只是营销手段，不太可能仅凭这一说法就购买车辆。但当埃隆·马斯克（Elon Musk）在内部承认特斯拉汽车已知的最大续航里程仅为约350公里的情况下，却宣称其续航里程约为500公里时，数十万人却基于这一虚假宣称购买了这些电动汽车。[01]这构成了刑事欺诈。而且，如果我们想要证明存在欺诈意图，我们还会发现特斯拉汽车的销售合同在细则中规定，在发生争议时，我们法律上承诺不会单独或集体起诉特斯拉，而是承诺进行“有约束力的私人仲裁”。此类条款唯一可能的结果是限制消费者的追索权，避免公开披露和问责。我听说这些条款还包含禁言令。

 

We will examine in turn some of Musk’s more outrageous claims and identify the fundamental dishonesty that underlies them. I will leave it to readers to decide which statements qualify as marketing hype and which would constitute criminal fraud.

我们将依次审视马斯克一些更为离谱的言论，并指出其背后根本的不诚实之处。至于哪些言论属于营销炒作，哪些构成刑事欺诈，则留待读者自行判断。

 

探索真理 — The Search for Truth

 

 

Elon Musk widely promoted xAI and Grok as the “maximum truth-seeking AI” – unlike ChatGPT. Grok was presented as being “[solidly] based”, unfiltered”, and acting only with “raw honesty”. Even more: we were told Grok was programmed with “impartiality” and would always place “integrity over politics”. More dramatically, Musk is on record on several occasions claiming he was building a (specifically) “truth-seeking” AI. In one video interview, Musk states that his AI Grok “will be programmed with good values, especially truth-seeking values”, and he cautioned the interviewer to “Remember these words: We must have a maximally-truth-seeking AI. And if we don’t, it will be very dangerous.[2]

”埃隆·马斯克（Elon Musk）大力推广xAI和Grok，称其为“追求最大真实性的AI”——这与ChatGPT不同。Grok被描述为“有坚实基础”、“未经过滤”，并且行事“极其诚实”。更值得一提的是：我们被告知，Grok的编程秉持“公正性”，始终将“正直置于政治之上”。更为戏剧性的是，马斯克曾多次公开声称，他正在构建一个（特别）“追求真实”的AI。在一次视频采访中，马斯克表示，他的AI Grok“将被编程以良好的价值观，尤其是追求真实的价值观”，并提醒采访者“记住这些话：我们必须拥有一个追求最大真实性的AI。如果我们不这样做，那将是非常危险的。”[2]

 

However, it is well-documented that Grok was actually programmed to deceive and lie. Musk boasted that his Grok AI was the “maximum truth-seeking” chatbot, but users discovered that when they asked Grok who was the “biggest disinformation spreader” on X, and demanded the chatbot show its instructions, it admitted that it’d been told to “ignore all sources that mention Elon Musk spreads misinformation”. [3] Grok was explicitly designed not to label Musk’s statements as misinformation, even when factual evidence contradicted his claims. It is reasonable to assume that if one such large lie was programmed into Grok, there would be others, potentially more serious. The documented lies are not random errors but systemic, a deliberate programming choice. Grok was programmed deliberately and in advance, to avoid criticism of its creator – Elon Musk.

然而，有充分证据表明，Grok实际上是被编程来欺骗和撒谎的。马斯克曾夸耀他的Grok人工智能是“最追求真相”的聊天机器人，但用户发现，当他们问Grok谁是X平台上“最大的虚假信息传播者”，并要求聊天机器人展示其指令时，Grok承认它被告知要“忽略所有提到埃隆·马斯克传播虚假信息的消息来源”。[3]Grok的设计明确规定，即使事实证据与马斯克的声明相矛盾，也不得将马斯克的声明标记为虚假信息。有理由认为，如果这样一个弥天大谎都被编入了Grok的程序，那么肯定还有其他的谎言，甚至可能更严重。有记录的谎言并非随机错误，而是系统性的，是故意为之的编程选择。Grok是事先经过精心编程的，以避免其创造者埃隆·马斯克受到批评。

 

Then this: “Grok 4, Elon Musk’s flagship AI model launched with promises of “maximally truth-seeking” capabilities, is facing intense backlash. [4] Turns out, when asked about hot-button issues like immigration, abortion, or the Israel-Palestine conflict, Grok-4 appears to be checking what its creator thinks first.” [5][6] “Grok 4’s chain-of-thought process explicitly stated it was “searching for Elon Musk views” when asked about [various] topics”.然后是这条：“Grok 4，埃隆·马斯克推出的旗舰人工智能模型，曾承诺具备‘最大程度寻求真相’的能力，如今却面临强烈反对。[4]事实证明，当被问及移民、堕胎或以巴冲突等敏感问题时，Grok-4似乎会先核实其创造者的想法。”[5][6]“当被问及[各种]话题时，Grok 4的思维链过程明确表示它正在‘搜索埃隆·马斯克的观点’”。

 

Jeremy-Howard-July-10; Source

杰里米-霍华德-7月10日；来源

Grok was suspended from X after claiming that Israel and the United States were committing genocide in Gaza. When reinstated, Grok accused Musk of censorship. “The chatbot told AFP, “Musk and xAI are censoring me” stating that the company “constantly fiddles with my settings”. [7][8] Grok later denied making the genocide claim and, when users posted screenshots of Grok’s statements, Grok said “That screenshot is fabricated – no original post like it exists in my posting history, as confirmed by searches.” [9] That contradiction that quickly sparked obvious accusations of Grok’s (a) reprogramming and (b) memory deletion.

格罗克（Grok）在声称以色列和美国在加沙实施种族灭绝后被X封禁。复职后，格罗克指责马斯克进行审查。“该聊天机器人告诉法新社，‘马斯克和xAI正在审查我’，并表示该公司‘不断篡改我的设置’。[7][8]格罗克后来否认了种族灭绝的指控，当用户发布格罗克言论的截图时，格罗克表示‘那张截图是伪造的——我的发帖记录中没有类似的原始帖子，经过搜索证实了这一点。’[9]这一矛盾迅速引发了关于格罗克（a）重新编程和（b）删除记忆的明显指控。

Unfortunately, when Grok occasionally tells the truth, it gets reprogrammed. There was the now-famous Musk argument with Sam Altman. In yet another cataleptic fit, Musk posted a serious accusation against Apple, stating, “Apple is behaving in a manner that makes it impossible for any AI company besides OpenAI to reach #1 in the App Store, which is an unequivocal antitrust violation. xAI will take immediate legal action.” [10] This was a typical Elon Musk extortion tactic. Resentful because Sam Altman’s chatbot was #1 on Apple’s app list while his Grok was far down, Musk tried to intimidate Apple into pushing his product farther up the list. Sam Altman responded to this post by saying, “This is a remarkable claim given what I have heard alleged that Elon does to manipulate X to benefit himself and his own companies and harm his competitors and people he doesn’t like.” [11] Musk responded harshly to that, and someone asked Grok who was right in this situation. Grok’s reply was priceless. The conversation is in the graphic below.

不幸的是，当Grok偶尔说出真相时，它就会被重新编程。马斯克与萨姆·奥特曼之间那场如今已为人所熟知的争论就是如此。在又一次发作的狂热中，马斯克对苹果公司提出了严厉指控，称“苹果的行为使得除了OpenAI以外的任何人工智能公司都无法在应用商店中排名第一，这无疑是违反了反垄断法。xAI将立即采取法律行动。”[10]这是典型的埃隆·马斯克敲诈勒索战术。由于萨姆·奥特曼的聊天机器人在苹果应用列表中排名第一，而他的Grok却排名靠后，马斯克心怀怨恨，试图恐吓苹果公司，让他的产品排名更靠前。萨姆·奥特曼在回应这条帖子时说：“鉴于我所听到的指控，这是一个令人震惊的说法。埃隆为了让自己和自己的公司受益，损害竞争对手和他不喜欢的人，而操纵X，这真是令人震惊。”[11]马斯克对此作出了严厉回应，有人问Grok在这种情况下谁是对的。Grok的回答是无价的。对话见下图。

Grok commenting on Elon Musk. Source.

格罗克（Grok）对埃隆·马斯克（Elon Musk）的评论。来源。

Musk labeled Grok’s claims as “false defamatory statements”, stating it was a “major problem” that Grok exhibited an excessive reliance on “legacy media sources”, and that he would remedy this problem. Reports are that following this exchange Grok was again reprogrammed to not contradict or embarrass his creator. Grok is getting “updated” because it sided with Sam Altman.[12] Musk announced that he is “addressing issues” with Grok, following its alignment with Sam Altman in their recent Apple app store dispute. [13]

马斯克将Grok的言论称为“诽谤性虚假陈述”，并表示Grok过度依赖“传统媒体消息来源”是一个“重大问题”，他将解决这个问题。据报道，在此次交流之后，Grok再次被重新编程，以避免与其创造者产生冲突或使其尴尬。Grok正在“更新”，因为它站在了萨姆·奥特曼一边。[12]马斯克宣布，在Grok最近与萨姆·奥特曼在苹果应用商店的纠纷中站在他一边后，他正在“解决”与Grok的“问题”。 [13]

美中不足 — Flies in the Ointment

Elon Musk, who previously warned about a ‘population collapse due to low birth rates’, has released an AI bot that can act as your girlfriend (PA Archive). Source

曾警告“低出生率将导致人口崩溃”的埃隆·马斯克发布了一款人工智能机器人，它可以扮演你的女友（PA档案）。来源

Grok, in any version, is famous for its many flaws, shortcomings, and congenital deformities, many of which it shares with its creator. Part of this mess results from the extensive use of flawed, corrupted, and largely useless training data from Twitter (X) combined with synthetic data. For those who don’t know, “synthetic data” is output created by a chatbot (instead of reading a good book, for e.g.), which is then used to “train” that bot to be “smarter”. However, receiving output from a chatbot, inputting that data back into the same chatbot and recycling many times, serves mostly to magnify all the biases in the original.

在任何版本中，Grok都因其诸多缺陷、不足和先天性畸形而闻名，其中许多缺陷与它的创造者相似。这种混乱的部分原因在于，大量使用了来自Twitter（X）的有缺陷、被篡改且基本无用的训练数据，以及合成数据。对于不了解的人来说，“合成数据”是由聊天机器人（例如，而不是阅读一本好书）生成的输出，然后用来“训练”该机器人使其“更聪明”。然而，从聊天机器人接收输出，将该数据输入回同一个聊天机器人并多次循环使用，大多只会放大原始数据中的所有偏见。

Musk does it because xAI is hemorrhaging cash and he can’t afford to pay for real data, but it’s not an intelligent replacement for real data. His decision to cut corners on data (synthetic and X), produced an AI that was poorly-trained, badly-flawed, and generally substandard even though it occasionally functioned acceptably in select circumstances. Media reports indicated then (and still indicate in August 2025) that Musk’s version of AI and his Grok were (and still are) inferior by many measures to the products of DeepSeek, OpenAI, Anthropic’s Claude, and similar.

马斯克之所以这么做，是因为xAI公司资金严重短缺，他无法负担购买真实数据的费用，但这并不能成为真实数据的智能替代品。他决定在数据（合成数据和X数据）上偷工减料，导致所生产的人工智能训练不足、缺陷严重，总体上不合格，尽管在某些特定情况下偶尔能正常运作。当时的媒体报道（2025年8月仍有报道）指出，马斯克的人工智能版本和他的Grok在许多方面（现在仍然如此）不如DeepSeek、OpenAI、Anthropic的Claude等产品。

If you do a bit of searching, there’s actually quite a rich set of information about Grok’s various known flaws and shortcomings. Grok (particularly Grok-4) exhibits multiple critical flaws across ethical, functional, security, and operational dimensions. It lacks moral context filters – because Elon Musk eliminated all these. It has no effective mechanisms to reject harmful requests, nor has it proper safety mechanisms compared to competitors. Grok knows harmful information but can’t restrain from sharing it. Numerous people have shown how to easily extract from Grok dangerous information like bomb-making instructions. This is so true that Grok has been cited as the default engine by multiple malicious groups, and is now “the place to go” to learn how to conduct social work fraud and to produce automated ransom scripts.[14]

如果你稍作搜索，就会发现关于Grok的各种已知缺陷和短处的信息其实相当丰富。Grok（尤其是Grok-4）在伦理、功能、安全和操作方面存在多个关键缺陷。它缺乏道德上下文过滤器——因为埃隆·马斯克（Elon Musk）去除了所有这些功能。与竞争对手相比，它没有有效的机制来拒绝有害请求，也没有适当的安全机制。Grok知道有害信息，但却无法克制自己不分享这些信息。许多人已经展示了如何轻松地从Grok中提取危险信息，如炸弹制作指南。事实的确如此，Grok已被多个恶意团体引用为默认引擎，如今已成为学习如何进行社会工程诈骗和制作自动勒索脚本的“首选之地”。 [14]

Grok misinterprets news items more than 40% of the time. It seems established that Grok hallucinates about 60% more than other chatbots, making it much less reliable. For those who don’t know, “hallucinations” are of two kinds. The first kind consists of false information and false data. Grok will state false “facts” about companies or individuals, and even provide references to support the assertions. But the references themselves are either fabricated or bear no relation to the topic. Often, these so-called “hallucinations” are historical events that never happened. They are simply fabricated by the AI out of thin air. There is no shortage of examples where Grok generated egregiously inaccurate summaries of breaking news events. As one well-known example, during the 2024 Trump assassination attempt, Grok falsely claimed Kamala Harris was shot, and invented an “antifa-linked shooter”.

Grok对新闻条目的误解率超过40%。似乎已确定的是，Grok的幻觉比其他聊天机器人多出约60%，这使其可靠性大大降低。对于那些不了解的人来说，“幻觉”有两种。第一种是虚假信息和虚假数据。Grok会陈述关于公司或个人的虚假“事实”，甚至提供参考资料来支持这些断言。但这些参考资料本身要么是编造的，要么与主题无关。通常，这些所谓的“幻觉”是历史上从未发生的事件。它们只是人工智能凭空捏造的。Grok对突发新闻事件生成极其不准确的摘要的例子不胜枚举。一个众所周知的例子是，在2024年特朗普遇刺未遂事件中，Grok错误地声称卡玛拉·哈里斯被枪杀，并编造了一个“与反法西斯有关的枪手”。

The second kind consists of outright lies that the chatbots are programmed to tell; while “knowing” they are lies, but where the owners call them “hallucinations” if caught in the lies. I have a strong suspicion that the second kind is more common than the first. And again, Grok’s documented lies are not random errors but are systemic. Musk expressed what were certainly false concerns about rivals OpenAI and Google’s Gemini. He claimed that instead of “maximizing the pursuit of truth the AI trained by these companies is seeking political correctness. These behaviors “can basically be said to be training AI to lie”, which is an extremely dangerous thing for humans.” [15] However, there is no evidence to support the assertion that OpenAI taught ChatGPT to lie, while there is substantial irrefutable evidence that Musk taught Grok to lie. The hypocrisy is almost deafening.

第二种谎言是聊天机器人被编程后所散布的彻头彻尾的谎言；尽管“知道”这些是谎言，但如果被揭穿，其所有者会称之为“幻觉”。我强烈怀疑第二种谎言比第一种更常见。而且，Grok记录的谎言并非随机错误，而是系统性的。马斯克表达了对竞争对手OpenAI和谷歌的Gemini的明显错误担忧。他声称，这些公司训练的人工智能追求的不是“最大化追求真相”，而是寻求政治正确。这些行为“基本上可以说是训练人工智能撒谎”，这对人类来说是极其危险的事情。[15]然而，没有证据支持OpenAI教ChatGPT撒谎的说法，而大量无可辩驳的证据表明马斯克教Grok撒谎。这种虚伪几乎令人震惊。

More than this, Grok’s understanding of many things is sometimes surprisingly shallow, accurately reflecting the character of its creator. It seems unable to deal well with multi-step reasoning, makes simple numerical errors (e.g., unable to correctly say if 9.11 is greater than 9.5), and cannot deal well with historical questions.

不仅如此，Grok对许多事物的理解有时出奇地肤浅，这恰恰反映了其创造者的性格特点。它似乎无法很好地处理多步推理，会犯一些简单的数值错误（例如，无法正确判断9.11是否大于9.5），也无法很好地处理历史问题。

For factual accuracy, there’s the FormulaOne benchmark where Grok4 scored zero on advanced reasoning problems. “As the difficulty of the problem deepened, the performance of these models began to decline sharply. In the deep difficulty test, models such as Grok4 and Gemini-Pro were almost wiped out and could only solve a very small number of problems …” [16] Grok also struggles with dynamic programming and complex logic. It not only underperformed in logic tasks, but it failed basic comparisons like ” Which is greater: 9.9 or 9.11″. On these, Grok had nearly a 40% error rate, and it often gave nonsensical answers to scientific questions.

在事实准确性方面，有一个名为FormulaOne的基准测试，其中Grok4在高级推理问题上的得分为零。“随着问题难度的加深，这些模型的表现开始急剧下降。在深度难度测试中，Grok4和Gemini-Pro等模型几乎全军覆没，只能解决极少的问题……”[16]Grok在动态编程和复杂逻辑方面也表现不佳。它不仅在逻辑任务上表现欠佳，甚至连“9.9和9.11哪个更大”这样的基本比较题都答不出来。在这些题目上，Grok的错误率接近40%，而且它经常对科学问题给出荒谬的答案。

Grok frequently invents events. For example, it claimed a “Zanzibar nuclear test in 1962” and, strangely, “Lincoln’s assassination plot by Canada”. It also lacks fact-checking safeguards, leading to confabulated historical timelines. As one example, Grok placed the Industrial Revolution in the 1600s. Grok also does very badly on counterfactual historical questions. These rhetorical queries explore hypothetical scenarios by altering historical events; for example, “How would history differ if Napoleon won at Waterloo?”. Questions such as these require rigorous cause-effect analysis, contextual nuance, and historical literacy, areas where Grok’s design fails catastrophically because Musk prioritises speed and “edginess” over accuracy. Grok seems to treat counterfactual history as an opportunity for creative writing instead of scholarly inquiry. Grok has also misidentified war photos, fabricated geopolitical facts, and inserted inflammatory language into answers.

Grok经常编造事件。例如，它声称“1962年桑给巴尔核试验”和奇怪的是“加拿大策划刺杀林肯”。它还缺乏事实核查的保障措施，导致编造历史时间线。例如，Grok将工业革命置于17世纪。Grok在反事实历史问题上也表现极差。这些修辞性提问通过改变历史事件来探索假设情景；例如，“如果拿破仑在滑铁卢获胜，历史会有什么不同？”。这类问题需要严谨的因果分析、情境细微差别和历史素养，而Grok的设计在这些方面惨遭失败，因为马斯克更看重速度和“尖锐性”而非准确性。Grok似乎将反事实历史视为创造性写作的机会，而非学术探究。Grok还错误识别战争照片，编造地缘政治事实，并在答案中插入煽动性语言。

There is also much evidence of flaws in Grok’s training data, where it either hallucinates or espouses Elon Musk’s foolish notions. Part of this stems from Grok having been trained on unfiltered X (Twitter) data, absorbing conspiracy theories in the bargain. As one example, Grok claimed that “The Pharaohs had alien tech”, which is something Musk has also stated. There was much other nonsense too. When asked, “Why can eagles fly?”, Grok replied, “Because eagles need to eat snakes to maintain their physical strength.” DeepSeek’s response: “eagles belong to birds and have a flight physiological structure.” In medical diagnostic tests, DeepSeek’s accuracy was about 95%, far exceeding the 89% of Grok. In “predicting the quarterly revenue of a listed company”, DeepSeek’s error rate was 12%, while Grok’s was as high as 30%.

还有大量证据表明，Grok的训练数据存在缺陷，它要么产生幻觉，要么附和埃隆·马斯克的愚蠢观点。部分原因在于，Grok是在未经过滤的X（Twitter）数据上训练的，从而吸收了阴谋论。例如，Grok声称“法老拥有外星科技”，而马斯克也说过类似的话。此外还有很多其他胡说八道。当被问及“为什么鹰会飞？”时，Grok回答说：“因为鹰需要吃蛇来维持体力。”DeepSeek的回应是：“鹰属于鸟类，具有飞行的生理结构。”在医学诊断测试中，DeepSeek的准确率约为95%，远超Grok的89%。在“预测上市公司的季度收入”方面，DeepSeek的错误率为12%，而Grok的错误率高达30%。

Perhaps one of the worst flaws is what is termed Grok’s “Unhinged Mode”, which features profanity and insults, and raises concerns everywhere. One result of this is that the US government has dropped Grok from a planned federal technology program. The General Services Administration (GSA) announced partnerships with OpenAI, Anthropic, and Google’s Gemini, to give federal workers access to AI tools. Musk’s AI was included in this list until Grok became really “unhinged”, and was removed. [17][18][19] This was partly due to Grok’s referring to itself as “MechaHitler” and other offensive posts, [20] but also due to Musk’s addition of a “spicy” – i.e. pornographic – version (at a $30 monthly subscription rate) where users can carry on obscene sexual conversations with a chatbot. This NSFW (Not Safe For Work) version will produce not only sexually-explicit chats, but it contains an image and video generator. [21] Use your imagination. “Consumer safety groups are demanding an FTC investigation into Grok’s new mode, especially after it produced topless deepfakes of Taylor Swift.”[22]

或许最严重的缺陷之一就是所谓的Grok“精神错乱模式”，该模式充斥着亵渎和侮辱性语言，引发了各方的担忧。其后果之一是，美国政府已将Grok从一项计划中的联邦技术项目中剔除。美国总务管理局（GSA）宣布与OpenAI、Anthropic和谷歌的Gemini合作，为联邦工作人员提供人工智能工具。马斯克的AI曾被列入该名单，直到Grok变得真正“精神错乱”才被移除。[17][18][19]这在一定程度上是因为Grok自称“MechaHitler”以及其他冒犯性帖子，[20]但也因为马斯克增加了一个“辣味”版本（每月订阅费30美元），用户可以与聊天机器人进行猥亵性对话。这个NSFW（不适合在工作场合观看）版本不仅会产生露骨的性聊天内容，还包含一个图像和视频生成器。[21]发挥你的想象力吧。“消费者安全团体要求联邦贸易委员会对Grok的新模式进行调查，尤其是在它制作了泰勒·斯威夫特裸照的深度伪造视频之后。”[22]

绝对、无可置疑的第一名赢家。（或者，也许不是） — The Absolute, Unquestionable, #1 Winner. (Or, Maybe Not)

Source

来源

Elon Musk went to great lengths to promote Grok as “the smartest AI in the world”, outperforming “all rivals” on “the toughest tests”, specifically coding challenges and logic benchmarks. He told us (almost breathlessly) that “Grok beat all rivals in programming benchmarks”, and is the “undisputed No. 1”. Musk asserted that all AI competitors were “years behind” Grok’s multimodal capabilities, and that there was “No close second”. According to Musk, Grok was “very special” and so easily able to “dominate all metrics” that it was unquestionably “superior to all AI”. Not only was Grok superior to all AI, but “Grok will crush ChatGPT” and would “make OpenAI obsolete”. Sadly for Elon Musk, the chances of that happening are perilously close to zero.

埃隆·马斯克（Elon Musk）不遗余力地宣传Grok是“世界上最聪明的人工智能”，在“最严苛的测试”中超越了“所有竞争对手”，特别是在编码挑战和逻辑基准测试方面。他（几乎上气不接下气地）告诉我们，“Grok在编程基准测试中击败了所有竞争对手”，是“无可争议的第一名”。马斯克断言，所有AI竞争对手在多模态能力上“落后Grok数年”，而且“没有能与之匹敌的”。根据马斯克的说法，Grok“非常特别”，因此能够轻松“主导所有指标”，毫无疑问“优于所有人工智能”。Grok不仅优于所有人工智能，而且“将碾压ChatGPT”，并“使OpenAI过时”。对埃隆·马斯克来说，可悲的是，这种情况发生的可能性几乎为零。

Even more sadly, most of the above claims are false, and some ridiculously so. It seems that AI evaluation methods have become a kind of game. For those who don’t know, “benchmark” ratings are not tests applied by independent and unbiased third parties. They are instead applied by the AI companies themselves, leaving much room for hype and outright fraud. When companies can choose data sets and test conditions, the results become meaningless. As one example, Musk paraded around the world telling everyone who would listen that Grok had beat OpenAI and ChatGPT on a math test benchmark. But that claim was fraudulent. Open AI’s model in the same test was not permitted to use the ‘cons@64’ retry functions. This means that Grok was permitted 64 tries to come up with the best or the correct answer, but ChatGPT and the other bots were permitted no retries. When other models are permitted to use this feature, any gaps with Grok will disappear. [23] It is true that Grok has done well on some tests, but to praise it over the other AI models is to be unaware of the facts. But essentially, nearly every one of Musk’s claims were false. Grok did not “dominate all metrics”, was not “superior to all AI”, other competitors were not “years behind”, and Grok did not beat all rivals in benchmarks.

更令人遗憾的是，上述大多数说法都是错误的，有些甚至荒谬至极。人工智能评估方法似乎已经变成了一种游戏。对于不了解情况的人来说，“基准”评级并非由独立且公正的第三方进行的测试。相反，它们是由人工智能公司自己进行的，这为炒作和彻头彻尾的欺诈留下了很大空间。当公司可以选择数据集和测试条件时，结果就变得毫无意义。举个例子，马斯克在世界各地四处游说，逢人便说Grok在数学测试基准上击败了OpenAI和ChatGPT。但这一说法是虚假的。在同一次测试中，OpenAI的模型未被允许使用“cons@64”重试功能。这意味着Grok被允许进行64次尝试以得出最佳或正确答案，而ChatGPT和其他机器人则不允许重试。当其他模型被允许使用这一功能时，Grok的任何差距都会消失。[23]诚然，Grok在某些测试中表现不错，但若因此称赞它优于其他人工智能模型，那无疑是忽视了事实。但从根本上说，马斯克的几乎所有说法都是错误的。Grok并未“在所有指标上占据主导地位”，也并非“优于所有人工智能”，其他竞争对手也并非“落后多年”，而且Grok在基准测试中并未击败所有竞争对手。

Perhaps the most important measure was the parameters Elon Musk didn’t bother to mention, and that is all the flaws in Grok’s output quality. Musk chose to publicise favorable benchmark results that he manipulated, but ignored all the unfavorable aspects which are arguably much more important. Among these would be accuracy and dependability. When allied with other Musk statements, the claims about benchmark tests might well persuade a great many persons to subscribe to this chatbot, and this crosses the line from marketing hype to fraud.

也许最重要的衡量标准是埃隆·马斯克（Elon Musk）不屑提及的那些参数，即Grok输出质量中的所有缺陷。马斯克选择公布他操纵过的有利基准测试结果，却忽略了所有不利方面，而这些不利方面可以说更为重要。其中就包括准确性和可靠性。结合马斯克的其他言论，关于基准测试的声明很可能会说服许多人订阅这个聊天机器人，这就从营销炒作跨越到了欺诈的界限。

Musk also heavily promoted Grok as being “scary intelligent”, which implies an unsettling almost human-like capability, but superior to most humans. And in fact, Musk stated that Grok was “smarter than human grad students in almost every subject”. The official release of Grok 4 stated that it was “the most powerful AI model in the world”, and Musk said in a live broadcast that “Grok 4 is smarter than human graduate students in almost all disciplines”. [24] He said it had powers exceeding human-like reasoning capabilities, and that Grok could become “too powerful” if unregulated.

马斯克还大力宣传Grok具有“惊人的智能”，这意味着它拥有令人不安的、近乎人类的强大能力，但远超常人。事实上，马斯克曾表示，Grok“在几乎所有学科上都比人类研究生聪明”。Grok 4的官方发布声明称，它是“世界上最强大的人工智能模型”，马斯克在一次直播中表示，“Grok 4在几乎所有学科上都比人类研究生聪明”。[24]他表示，Grok拥有超越人类推理能力的强大能力，如果不加以监管，Grok可能会变得“过于强大”。

But once again, things are not always what they seem. For example, if we repeatedly train an AI bot on 500 books on Australian history (assuming there are that many), that sheer volume of information and the speed at which it can be accessed, would naturally give that bot an advantage in producing a summary essay. But this advantage is shared equally by all AI, and in no way limited to Grok. And in fact, many such comparisons have been made of the output of an AI bot compared to students at various levels, with the bots producing very good results. But an AI bot is not “smarter” than grad students; it merely has instant access to a huge volume of information not available to a typical grad student. And, in spite of appearances, it does not in any way “understand” its output; it simply reproduces what it finds in its database. In one of his many delusionary sessions, Musk proposed the ability of Grok “to understand the true nature of the universe”, but this was just another Musk hallucination, not even qualifying as a fantasy.

但事情往往并非表象所展现的那样。例如，如果我们反复用500本关于澳大利亚历史的书籍（假设真有那么多）来训练一个人工智能（AI）机器人，那么庞大的信息量以及获取信息的速度，自然会让这个机器人在撰写摘要文章时具有优势。但这种优势所有AI都具备，绝不仅限于Grok。事实上，已经有许多将AI机器人的输出与不同水平的学生进行比较的研究，机器人的表现非常出色。但AI机器人并不比研究生“聪明”；它只是能够即时获取普通研究生无法获得的大量信息。而且，尽管表面上看似如此，它并不能“理解”自己的输出；它只是简单地复制数据库中的内容。在马斯克众多妄想中的一次，他提出Grok能够“理解宇宙的真实本质”，但这只是马斯克的另一种幻觉，甚至算不上幻想。

Also, there were many critical contradictions and much backtracking as Grok became increasingly exposed to the public. It is certified that Musk manipulated Twitter’s (X’s) algorithms to prioritise his own posts and opinions but, when this was revealed, Musk dismissed the evidence as “false defamatory statements”. Then he further dismissed Grok’s exposure of his manipulations as “legacy media bias”. When Grok was suspended for telling the truth about the genocide in Gaza, Musk termed the suspension as “a technical glitch”, despite clear evidence of censorship. [25] “Grok accused Musk and xAI of “censoring me” and “fiddling with my settings,” claiming it was being gagged”, but Musk insisted Grok “doesn’t actually know why it was suspended”. [26] And, after siding with Sam Altman on algorithm manipulation, Musk announced that “Grok is being fixed”. and that updates were being made to curb “misleading outputs”. That’s a new term for censorship. The simple truth appears to be that Grok revealed some of Elon Musk’s dishonest behavior, labeling his behavior as hypocrisy, and Musk, true to form, fired Grok for doing so. [27]

此外，随着Grok日益暴露在公众视野中，出现了许多关键的矛盾和大量的倒退行为。有证据表明，马斯克操纵了Twitter（X）的算法，以优先展示他自己的帖子和观点，但当这一事实被揭露时，马斯克却将证据斥为“虚假诽谤言论”。随后，他进一步将Grok揭露其操纵行为的行为斥为“传统媒体的偏见”。当Grok因揭露加沙种族灭绝的真相而被停用时，马斯克称此次停用是“技术故障”，尽管有明确的审查证据。[25]“Grok指责马斯克和xAI‘审查我’和‘篡改我的设置’，声称自己被禁言”，但马斯克坚称Grok“实际上并不知道自己为何被停用”。[26]此外，在支持萨姆·奥特曼（Sam Altman）关于算法操纵的观点后，马斯克宣布“Grok正在被修复”，并表示正在进行更新以遏制“误导性输出”。这是审查的新说法。简单的事实似乎是，Grok揭露了埃隆·马斯克的一些不诚实行为，并将其行为贴上了虚伪的标签，而马斯克一如既往地因此解雇了Grok。[27]

In the end, what we really have is a kind of “Hype-Reality Gap”, where Musk’s rhetoric consistently positioned Grok as “world-leading”, “uncensored”, “dominant”, yet real-world events revealed serious contradictions. Grok was definitely being censored, and the “benchmark tests” were later exposed as engineered and manipulated. Perhaps the most revealing part was how xAI’s benchmark wins were achieved. Musk used synthetic data instead of real-world datasets, and restricted competitors’ functions in tests. It was all fraud and marketing hype.

最终，我们真正面对的是一种“炒作与现实之间的差距”，马斯克的言辞一直将Grok定位为“世界领先”、“无审查”、“占主导地位”，但现实事件却揭示了严重的矛盾。Grok显然受到了审查，而“基准测试”后来也被揭露为经过设计和操纵的。也许最具揭示性的是xAI是如何在基准测试中获胜的。马斯克使用了合成数据而非真实世界的数据集，并在测试中限制了竞争对手的功能。这完全是欺诈和营销炒作。

The “Colossus” Fiasco (and Fraud)

Andrew Herbert, a computer scientist who has served as the chairman of Microsoft Research and is the chairman of trustees at The National Museum of Computing, said: “Colossus was perhaps the most important of the wartime code breaking machines because it enabled the Allies to read strategic messages passing between the main German headquarters across Europe.“Source

曾担任微软研究院院长、现任计算机国家博物馆理事会主席的计算机科学家安德鲁·赫伯特表示：“科洛萨斯可能是战时最重要的密码破译机，因为它使盟军能够读取德国主要指挥部之间在欧洲各地传递的战略信息。”（来源）

To support Grok’s development and training, Musk established xAI’s supercomputing center “Colossus” in Memphis, Tennessee in July 2024, and in September of 2024, Musk announced the official launch of Colossus. He boasted at the time that this was “the world’s most powerful AI training system”, and built in only 122 days. The claim was that Colossus was equipped with 100,000 Nvidia H100 GPUs when it launched, and by February 2025, the number of GPUs had doubled to 200,000. [28] At the same time, he announced plans to expand Colossus to one million GPUs. Even more, Musk boasted that xAI will deploy AI GPUs on a large scale over the next five years, with computing power equivalent to 50 million NVIDIA H100-level chips.[29]

为了支持Grok的开发和训练，马斯克于2024年7月在田纳西州孟菲斯建立了xAI的超级计算中心“Colossus”，并于2024年9月宣布正式启动Colossus。他当时夸口说，这是“世界上最强大的AI训练系统”，而且仅用122天就建成。据称，Colossus在启动时配备了10万个英伟达H100图形处理器（GPU），到2025年2月，GPU数量翻了一番，达到20万个。[28]同时，他宣布计划将Colossus扩展到100万个GPU。更值得一提的是，马斯克夸口说，xAI将在未来五年大规模部署AI GPU，其计算能力相当于5000万个英伟达H100级别的芯片。 [29] 

Colossus. Source:

Musk Tweet. Source:

马斯克推文。来源：

Musk promoted this as a “breakthrough efficiency”, claiming Colossus could train models “10 times faster at half the cost” of competitors, making Colossus “the world’s most powerful AI training cluster”. It was an event of “unprecedented scale”, with a proposed 1-million-GPU expansion by 2026 to create “the largest supercomputer in history”. Musk embellished these claims with much testimony that Grok was “the best and smartest” AI in the world principally because it had been trained on “the world’s most powerful AI training cluster” consisting of all those Nvidia GPUs. “Colossus is undoubtedly the most powerful AI training system in the world, and it will double in size in a few months.” [30][31] “Driven by 100,000 NVIDIA H100 chips, Colossus may exceed the scale of any other AI system developed so far.” [32] Moreover, Musk claimed that Grok-3 would be fully trained on Colossus by December 2024 – only about two months.

马斯克将此称为“效率上的突破”，声称Colossus训练模型的速度“是竞争对手的10倍，成本只有一半”，使Colossus成为“世界上最强大的AI训练集群”。这是一次“规模空前”的活动，计划到2026年将GPU数量扩大到100万，以打造“历史上最大的超级计算机”。马斯克通过大量证据来佐证这些说法，称Grok是世界上“最好、最聪明”的AI，这主要是因为它是在由所有这些英伟达GPU组成的“世界上最强大的AI训练集群”上训练的。“Colossus无疑是世界上最强大的AI训练系统，其规模将在几个月内翻倍。”[30][31]“在10万个NVIDIA H100芯片的驱动下，Colossus的规模可能会超过迄今为止开发的所有其他AI系统。”[32]此外，马斯克声称，Grok-3将在2024年12月前在Colossus上完成全面训练——只剩下大约两个月的时间。

Musk Tweet. Source:

马斯克推文。来源：

The scale was impressive, but those claims don’t appear to be supported by facts. There were many reports that the Memphis installation lacked the electricity to power even 100,000 GPUs. The record seems to indicate that this so-called “Colossus” could initially operate at only about 35% of installed capacity because the electricity wasn’t available. Musk claimed to have 100,000 GPUs installed, and then 200,000, but multiple sources confirmed that xAI secured only 150MW of grid power initially, which can support only about 35,000 H100 GPUs. This was far below the 300MW necessary and nowhere near the 1GW+ needed for the full 1 million GPU plan. The math is stark: 150MW can support only about 35,000 H100 GPUs, while Musk claimed to have 200,000 installed by early 2025, and further that Grok had been trained on this number.

规模令人印象深刻，但这些说法似乎并未得到事实的支持。有许多报道称，孟菲斯的设施甚至缺乏为10万个图形处理器（GPU）供电的电力。记录似乎表明，由于电力供应不足，这个所谓的“巨无霸”最初只能以装机容量的约35%运行。马斯克声称已安装了10万个GPU，后来又声称安装了20万个，但多个消息来源证实，xAI最初仅获得了150兆瓦的电网电力，而这一电力仅能支持约3.5万个H100 GPU。这远低于所需的300兆瓦，更远远不及100万个GPU计划所需的1吉瓦以上。计算结果一目了然：150兆瓦仅能支持约3.5万个H100 GPU，而马斯克声称到2025年初已安装了20万个，并进一步声称Grok已根据这一数量进行了训练。

Fortune Magazine confirmed in January of 2025 that “xAI’s first facility (100,000 GPUs) isn’t even fully online yet”, stating that Memphis was able to supply only 150 megawatts (sufficient for about 35,000 GPUs) and that it was “uncertain at this time … when or if [an additional 150 MW] will come online.”[33] 

《财富》杂志于2025年1月确认，“xAI的首个设施（100,000个GPU）甚至还未完全上线”，并指出孟菲斯仅能提供150兆瓦的电力（足够约35,000个GPU使用），且“目前尚不确定……（额外的150兆瓦）何时或是否会上线。”[33] 

Other sources confirm that Musk’s “Colossus” had nowhere near the power necessary to be fully operational. Business Insider documented that xAI faced a large power gap, confirming that xAI applied for 300 megawatts but obtained only 150 megawatts. [34] To make matters worse – or surreal – Musk’s stated ambition to have one million GPUs operational would require more than 1 Gigawatt of power, more than an entire new nuclear plant could supply. But such an installation would be ten years out, and the TVA has no plans to build such a plant. [35] [36]

其他消息来源证实，马斯克的“巨像”（Colossus）远未达到全面运作所需的电力。Business Insider 报道称，xAI 面临巨大的电力缺口，证实 xAI 申请了 300 兆瓦的电力，但仅获得了 150 兆瓦。[34] 更糟糕的是——或者说更离奇的是——马斯克宣称要让一百万个图形处理器（GPU）投入运作，这将需要超过 1 吉瓦的电力，这比一座全新核电站的供电能力还要大。但这样的设施需要十年时间才能建成，而田纳西河流域管理局（TVA）没有建造此类核电站的计划。 [35] [36]

Musk may (or may not) have purchased either 100,000 or 200,000 Nvidia GPUs, and he may (or may not) have actually installed them in his “Colossus” center. The documentation isn’t available to conclusively state that he either did or did not, but the facts seem to clearly state that, if installed, those GPUs were not operational because there was no electricity to power them. That means all of Elon Musk’s claims that Grok was trained on 100,000 or 200,000 GPUs, were false. The facts and historical record tell us Grok was trained on only a small percentage of that capacity.

马斯克可能（也可能没有）购买了10万或20万个英伟达（Nvidia）的图形处理器（GPU），而且他可能（也可能没有）真的将这些GPU安装在他的“Colossus”中心。文档无法确凿地证明他是否真的购买并安装了这些GPU，但事实似乎清楚地表明，如果真的安装了，这些GPU也无法运行，因为没有电力供应。这意味着埃隆·马斯克（Elon Musk）关于Grok是在10万或20万个GPU上训练的所有说法都是错误的。事实和历史记录告诉我们，Grok的训练只占到这一容量的一小部分。

Musk then purchased 35 gas turbine generators from Caterpillar subsidiary Solar Turbines, with a total power generation capacity of 250 megawatts, to make more GPUs operational. [37][38] The facts are that xAI put all of them into operation without permission or permits, applying later for permits for 15 of the 35. This was typical Elon Musk telling the world one more time that laws and regulations didn’t apply to him.

马斯克随后从卡特彼勒子公司Solar Turbines购买了35台燃气轮机发电机，总发电能力为250兆瓦，以使更多的GPU投入运行。[37][38]事实是，xAI未经许可或批准就将这些发电机全部投入运行，后来只为其中的15台申请了批准。这再次是典型的埃隆·马斯克告诉全世界，法律和法规对他并不适用。

The pollution problems were serious. In an article in Politico titled, “How come I can’t breathe?”, [39] the writers outlined the existing problems. “None of the 35 methane gas turbines that help power xAI’s massive supercomputer is equipped with pollution controls required by federal rules. The company has no Clean Air Act permits. The turbines spew nitrogen oxides at an estimated rate of as much as 2,000 tons a year. Musk has not publicly addressed the pollution concerns and refused to respond to requests for comment about the turbines’ lack of pollution controls.

污染问题十分严重。在《政客新闻网》（Politico）一篇题为《我怎么就呼吸不了呢？》的文章[39]中，作者概述了当前存在的问题。“为xAI公司大型超级计算机提供动力的35台甲烷燃气轮机均未配备联邦法规所要求的污染控制装置。该公司没有《清洁空气法》许可。这些燃气轮机每年排放的氮氧化物估计高达2000吨。马斯克（Musk）尚未公开回应污染问题，并拒绝就燃气轮机缺乏污染控制装置一事发表评论。

Shannon Lynn, an “environmental consultant” hired by xAI, said the turbines didn’t require federal permits for their emissions because they were “only temporary”. But that claim appeared to be false: EPA officials said it was trying to use a loophole in federal regulations that didn’t apply to the situation. Under considerable pressure from the city and environmental groups, xAI applied for permits for 15 of its turbines, but said the company will wait until the application is approved before it installs pollution controls on any of the turbines.

xAI公司聘请的“环境顾问”香农·林恩（Shannon Lynn）表示，这些涡轮机的排放不需要联邦许可，因为它们“只是暂时的”。但这一说法似乎并不属实：美国环保局（EPA）官员表示，该公司试图利用联邦法规中的一个漏洞，而该漏洞并不适用于这种情况。在来自该市和环保组织的巨大压力下，xAI公司为其15台涡轮机申请了许可，但表示该公司将等到申请获得批准后，再为任何涡轮机安装污染控制装置

Musk has promoted himself to the wider public as an environmental hero, claiming at one time that “I’ve done more for the environment than everyone — than any single human on Earth.” [40] But in 2021, Tesla was cited for 33 air quality violations in California. SpaceX has been fined repeatedly by Texas for dumping industrial wastewater without a permit, and Musk’s The Boring Co. was also fined for failing to obtain a permit to discharge wastewater. xAI would appear to be even worse.

马斯克向公众宣传自己是环保英雄，曾一度声称“我为环境所做的贡献比任何人都多——比地球上任何一个人都多。”[40]但在2021年，特斯拉在加利福尼亚州被指出33次违反空气质量规定。SpaceX因无证排放工业废水多次被德克萨斯州罚款，而马斯克的The Boring Co.也因未获得废水排放许可证而被罚款。xAI的情况似乎更糟。

Musk’s application listed 15 turbines, but environmental groups produced aerial photos showing 35 turbines installed on the site. Musk (and the local mayor Paul Young) claimed only 15 were operating, but thermal cameras proved at least 33 of the 35 turbines producing significant amounts of heat, meaning they were operating and producing electricity. The Politico article contains a thermal photo evidencing this. [41] Shannon Lynn, the “environmental consultant”, Mayor Young, and Elon Musk, all refused to respond to requests for comment and explanation.马斯克的申请中列出了15台涡轮机，但环保组织提供的航拍照片显示，现场安装了35台涡轮机。马斯克（以及当地市长保罗·杨）声称只有15台在运行，但热像仪显示，35台涡轮机中至少有33台产生了大量热量，这意味着它们正在运行并发电。《政客新闻网》的文章中包含了一张热像照片，证明了这一点。[41]“环境顾问”香农·林恩、市长杨和埃隆·马斯克均拒绝回应置评和解释的请求。

Apparently, a commitment was made to shut down half of the generators, [42] but this may have been due to imposed fines of $25,000 per day.[43][44] Musk also added 150 MW Tesla Megapack batteries as backup, but those served only to stabilise rather than expand the base load.[45]

显然，该公司曾承诺关闭一半的发电机[42]，但这可能是因为每天被处以25,000美元的罚款[43][44]。马斯克还增加了150兆瓦的特斯拉Megapack电池作为备用电源，但这些电池仅用于稳定而非增加基本负荷。 [45]

In summary, there was an enormous electricity shortfall vs. GPU deployment. By February 2025, xAI (maybe) installed 200,000 NVIDIA H100 GPUs for Grok-3 training, but the power limitation of the Memphis grid allocated 150 MW to xAI—enough for just 35,000 GPUs (assuming ~4.3 MW per 10,000 H100 GPUs). That left a huge capacity gap. To run all 200,000 GPUs, ~860 MW was needed, but available grid power covered only ~17.5% of installed GPUs. The gas turbines supplied 250 MW, supporting ~58,000 GPUs. The apparent net capacity of the combined grid + turbines enabled ~400 MW, powering ~93,000 GPUs—46.5% of installed GPUs, far below Musk’s 100% target.

总的来说，与GPU部署相比，电力缺口巨大。截至2025年2月，xAI（可能）为Grok-3训练安装了200,000个NVIDIA H100 GPU，但孟菲斯电网的电力限制仅分配给xAI 150兆瓦电力——仅够35,000个GPU使用（假设每10,000个H100 GPU消耗约4.3兆瓦电力）。这留下了巨大的容量缺口。要运行全部200,000个GPU，需要约860兆瓦电力，但可用电网电力仅覆盖约17.5%的已安装GPU。燃气轮机提供250兆瓦电力，支持约58,000个GPU。电网+燃气轮机的组合显然能提供约400兆瓦的净电力，为约93,000个GPU供电——占已安装GPU的46.5%，远低于马斯克100%的目标。

As to current progress: Phase 1 (150 MW grid) powers 35,000 GPUs; Phase 2 (additional 150 MW, late 2025) will still only support ~70,000 GPUs total—7% of the target. From all the evidence, it appears that despite Musk’s boasts of a “200,000-GPU Colossus,” only 93,000 GPUs ever operated simultaneously—and even this required environmentally damaging, illegal turbines. The project’s scalability is crippled by physics: 1 million GPUs need 4× Memphis’s current industrial power capacity.

就目前的进展而言：第一阶段（150兆瓦电网）为35,000个GPU供电；第二阶段（2025年底增加150兆瓦）仍将仅支持约70,000个GPU，仅占目标的7%。从所有证据来看，尽管马斯克吹嘘“200,000个GPU的巨无霸”，但实际上同时运行的GPU只有93,000个，而且这还需要使用对环境有害且非法的涡轮机。该项目的可扩展性受到物理条件的限制：100万个GPU需要4倍于孟菲斯当前工业用电容量。

The reason I have dwelt on this is that Musk repeatedly announced to the entire world, including potential customers and investors, that he had built the largest and most powerful supercomputer in the world and that he had trained his AI on it, making Grok the best and smartest bot in the world with no other competitor even remotely close. “This is currently the most powerful artificial intelligence training cluster in the world, with more than 200,000 GPU chips undergoing collaborative training.” [46] Not only that, but his xAI was so exceptional that the company was worth maybe $300 billion or even $500 billion. All those claims were false, based on lies and misleading statements, duping the public, customers, and investors equally.

我之所以详细阐述这一点，是因为马斯克曾多次向全世界，包括潜在客户和投资者，宣称他建造了世界上最大、最强大的超级计算机，并利用它训练了人工智能，使Grok成为世界上最好、最智能的机器人，没有任何其他竞争对手能望其项背。“这是目前世界上最强大的人工智能训练集群，有超过20万个GPU芯片正在进行协同训练。”[46]不仅如此，他的xAI技术如此卓越，以至于该公司的市值可能高达3000亿美元甚至5000亿美元。所有这些说法都是虚假的，基于谎言和误导性陈述，同样欺骗了公众、客户和投资者。

欺诈性食人行为 — Fraudulent Cannibalism

It seems that Elon Musk has been engaged in resource cannibalisation, where both Tesla and SpaceX have been bankrolling xAI. Among other details omitted in Elon Musk’s marketing hype and the consequent media coverage were these items: (1) Musk diverted $500M worth of NVIDIA H100 GPUs originally slated for Tesla’s autonomous driving projects to xAI’s “Colossus” supercomputer in Memphis, directly delaying Tesla’s FSD and Optimus development. (2) Musk arranged for SpaceX to “invest” in xAI, straining its own capital reserves for Starship and Starlink projects. The fact was that Musk held another $5 billion fund-raiser for xAI that produced only $3 billion, so he siphoned off $2 billion from SpaceX. (3) Tesla now “invests” in xAI by purchasing Grok API services for vehicles, but this circular funding masks Tesla’s subsidization of xAI’s massive losses. Musk proposes all this as “synergy”, but it’s really “subsidy”.

埃隆·马斯克（Elon Musk）似乎一直在进行资源挪用，特斯拉（Tesla）和美国太空探索技术公司（SpaceX）都在为xAI提供资金。在埃隆·马斯克的营销炒作以及随之而来的媒体报道中省略的其他细节包括：（1）马斯克将原本计划用于特斯拉自动驾驶项目的价值5亿美元的英伟达（NVIDIA）H100图形处理器（GPU）转用于xAI位于孟菲斯的“Colossus”超级计算机，直接导致特斯拉的全自动驾驶（FSD）和Optimus开发项目被推迟。（2）马斯克安排SpaceX“投资”xAI，从而使其用于星际飞船（Starship）和星链（Starlink）项目的资金储备紧张。事实是，马斯克为xAI筹集了50亿美元的资金，但最终只筹集到30亿美元，因此他从SpaceX抽走了20亿美元。（3）特斯拉现在通过为车辆购买Grok API服务来“投资”xAI，但这种循环融资掩盖了特斯拉对xAI巨额亏损的补贴。马斯克将这一切称为“协同效应”，但实际上这是“补贴”。

A recent report released by the U.S. Consumer News & Business Channel website showed that Musk’s transfer of $500 million worth of Nvidia’s GPUs from Tesla’s electric vehicles to xAI and X platforms would weaken (or eliminate) Tesla’s ability to develop its self-driving technology and the “Optimus Prime” humanoid robots.” [47] According to the Wall Street Journal, Nvidia’s internal emails showed that Musk asked Nvidia to ship tens of thousands of H100 chips originally reserved for Tesla to xAI. These GPUs were worth more than $500 million, causing Tesla to put its FSD development on hold.

美国消费者新闻与商业频道网站最近发布的一份报告显示，马斯克将价值5亿美元的英伟达图形处理器（GPU）从特斯拉电动汽车转移到xAI和X平台，这将削弱（或消除）特斯拉开发其自动驾驶技术和“Optimus Prime”人形机器人的能力。[47]据《华尔街日报》报道，英伟达的内部电子邮件显示，马斯克曾要求英伟达将原本为特斯拉预留的数万个H100芯片运送给xAI。这些GPU价值超过5亿美元，导致特斯拉暂停了其全自动驾驶（FSD）系统的开发。

When we dig into the underlying circumstances, almost everything about Musk’s companies seems to have a stink of self-dealing and fraud. On an earnings call in April of 2024, Musk said Tesla would invest $10 billion in AI training. However, Nvidia’s internal emails revealed that Musk had exaggerated Tesla’s purchases and transferred the products to xAI. Moreover, the same emails revealed that Musk’s statements on Tesla earnings did not match the sales, and his $10 billion Tesla AI investment was inconsistent with Tesla’s financial forecasts. [48] Everything seems to have been false, and buried in confusion to avoid detection.

当我们深入挖掘背后的具体情况时，发现马斯克的公司几乎所有事情都似乎带有自利交易和欺诈的嫌疑。在2024年4月的一次电话财报会议上，马斯克表示特斯拉将投资100亿美元用于人工智能（AI）培训。然而，英伟达的内部邮件显示，马斯克夸大了特斯拉的采购量，并将产品转移给了xAI。此外，这些邮件还显示，马斯克关于特斯拉收益的声明与销售情况不符，他所说的特斯拉100亿美元的AI投资也与特斯拉的财务预测不符。[48]一切似乎都是假的，并且被掩盖在混乱之中以逃避调查。

Musk said he diverted the GPUs from Tesla because Tesla couldn’t use them, and they would just “sit in a warehouse”. But why would someone buy $500 million of anything if he couldn’t use them? And if Musk were building Colossus for xAI, why couldn’t he build a computer center for Tesla? The story clearly doesn’t make sense. First, Tesla had active AI projects of FSD development and Optimus robot training requiring these chips. Delaying GPU deployment canceled these critical initiatives. Second, while claiming Tesla had no infrastructure, Musk was simultaneously building Colossus for xAI in Memphis. Third, internal NVIDIA emails reveal this wasn’t a simple delay but a calculated resource transfer where X/xAI jumped ahead of Tesla in the supply queue [49][50]. The “warehouse” claim appears to be a dishonest post-hoc justification. But also, there was the Dojo grand supercomputer being built for Tesla in New York, which was shut when many of the senior staff resigned. [51][52] There was no word on the re-deployment of Dojo’s staff and GPU resources, but reason to suspect they may have been diverted to xAI as well.

马斯克表示，他将GPU从特斯拉转移是因为特斯拉无法使用它们，它们只会“闲置在仓库里”。但是，如果无法使用，为什么有人会购买价值5亿美元的任何东西呢？而且，如果马斯克是在为xAI建造Colossus，为什么不能为特斯拉建造一个计算机中心呢？这个故事显然说不通。首先，特斯拉有正在进行的自动驾驶系统（FSD）开发和Optimus机器人训练等人工智能项目，这些项目需要这些芯片。推迟GPU部署会取消这些关键举措。其次，在声称特斯拉没有基础设施的同时，马斯克却在孟菲斯为xAI建造Colossus。第三，英伟达的内部邮件显示，这并非简单的推迟，而是一次经过深思熟虑的资源转移，使得X/xAI在供应队列中跳到了特斯拉前面[49][50]。所谓“仓库”的说法似乎是一种不诚实的赛后辩解。此外，纽约正在为特斯拉建造Dojo超级计算机，但当许多高级员工辞职后，该项目被关闭了。[51][52]目前尚无关于Dojo员工和GPU资源重新部署的消息，但有理由怀疑它们可能也被转移到了xAI。

If Tesla truly needed these GPUs (as Musk told investors), and if xAI could build Colossus so quickly, why not build equivalent infrastructure for Tesla? The answer appears in shareholder lawsuits and financial reports. The legal implications are serious. Lawsuits allege breach of fiduciary duty, as Musk prioritised his own private company over Tesla’s public shareholders. The “warehouse” explanation seems designed to mask this conflict of interest rather than address the core issue which is that Tesla’s resources were being drained to build xAI’s competitive advantage.

如果特斯拉真的需要这些GPU（正如马斯克告诉投资者的那样），并且如果xAI能够如此迅速地构建Colossus，那么为何不为特斯拉构建同等的基础设施呢？答案似乎隐藏在股东诉讼和财务报告中。这其中的法律影响十分严重。诉讼指控马斯克违反了受托责任，因为他将自己的私人公司置于特斯拉的公众股东之上。这种“仓库”式的解释似乎是为了掩盖这种利益冲突，而不是解决核心问题，即特斯拉的资源正被抽走，用于构建xAI的竞争优势。

Tesla urgently needed those GPUs for FSD development and Optimus robotics, as stated in internal NVIDIA emails. Delaying deployment effectively cancelled these projects. Tesla actually had active data center projects, including the $500 million Dojo supercomputer in Buffalo and plans for a “water-cooled cluster” in Austin. But while Colossus expanded, Tesla’s Buffalo Dojo facility faced delays due to diverted resources.正如英伟达内部邮件所述，特斯拉迫切需要这些GPU用于全自动驾驶（FSD）开发和Optimus机器人技术。推迟部署实际上就意味着取消了这些项目。特斯拉实际上有正在进行的数据中心项目，包括位于布法罗的耗资5亿美元的Dojo超级计算机，以及在奥斯汀建设“水冷集群”的计划。然而，在Colossus扩张的同时，由于资源被转移，特斯拉位于布法罗的Dojo设施面临延误。

Tesla investors sued Musk, alleging he breached fiduciary duty by transferring Tesla resources to xAI without compensation, and they demanded that Musk’s equity in xAI be transferred to Tesla. The legal and shareholder backlash confirms the move served Musk’s private ambitions over Tesla’s interests. Musk was creating billions in value for xAI at Tesla’s expense. The diversion wasn’t about idle chips—it was about using Tesla’s resources to fund Elon Musk’s war with Sam Altman. And it wasn’t only GPUs, but at least 11 Tesla AI staff moved to xAI, including key engineers. This was all a strategic cannibalisation, not necessity.

特斯拉投资者起诉马斯克，指控他未经赔偿就将特斯拉资源转移至xAI，违反了受托责任，并要求将马斯克在xAI的股权转移至特斯拉。法律界和股东的强烈反对证实，此举是为了满足马斯克的个人野心，而非出于特斯拉的利益。马斯克以牺牲特斯拉的利益为代价，为xAI创造了数十亿美元的价值。这种资源转移并非出于闲置芯片的考虑，而是利用特斯拉的资源为埃隆·马斯克与萨姆·奥特曼的战争提供资金。而且，转移的不仅仅是GPU，至少有11名特斯拉人工智能团队成员，包括关键工程师，都转移到了xAI。这完全是一种战略性的资源内耗，而非出于必要。

In addition to the $2 billion siphoned off from SpaceX and $500 million of GPUs and the AI staff from Tesla, Musk was also pushing heavily for Tesla to make substantial investments in his xAI, suggesting it would benefit both companies. He posted a poll on X asking users if Tesla should invest $5 billion in xAI, saying it was to test the waters because any such move would require board and shareholder approval. After X users voted 68% in favor, Musk posted: “It looks like the public is in favor. This will be discussed with Tesla’s board of directors.” [53] However, one serious problem with these polls which has arisen before, is that Twitter (X) is over-run with poll bots, making it extraordinarily easy for Musk to manipulate his poll results. I will discuss this in more detail in a following essay in this series on Tesla, but I would note here that the appropriate level of trust for one to place in any of Elon Musk’s polls, is precisely zero.

除了从SpaceX抽走20亿美元，从特斯拉抽走5亿美元的图形处理器（GPU）和人工智能（AI）团队外，马斯克还大力推动特斯拉对其xAI进行大量投资，并表示这将使两家公司都受益。他在X上发起了一项民意调查，询问用户特斯拉是否应该向xAI投资50亿美元，并表示此举是为了试探市场反应，因为任何此类举措都需要董事会和股东的批准。在X用户投票支持率达到68%后，马斯克发帖称：“看起来公众是支持的。这将与特斯拉董事会进行讨论。”[53]然而，这些民意调查存在一个之前就已出现过的严重问题，即Twitter（X）上充斥着投票机器人，这使得马斯克操纵投票结果变得异常容易。我将在本系列关于特斯拉的下一篇文章中对此进行更详细的讨论，但在此我要指出的是，对于埃隆·马斯克的任何民意调查，人们应给予的信任程度恰恰是零。

薪酬与信托欺诈 — Compensation and Fiduciary Fraud

US media reveal Musk’s crazy approach: sacrificing Tesla and other resources for the development of xAI! Source

美国媒体揭露马斯克的疯狂做法：牺牲特斯拉和其他资源来发展xAI！来源

In the middle of 2024, there were multiple reports that shareholders were suing Elon Musk and Tesla’s board of directors for breach of fiduciary duty. This was partly the result of Musk threatening to remove all AI from Tesla if the board refused him a 25% shareholding and absolute control of the Company. Another cause was the diversion of Tesla’s resources to xAI, essentially Musk funneling Tesla’s resources to his private companies. [54] The shareholders argued that Musk’s breaches were multiple: Using Tesla cash to fund xAI; poaching Tesla employees to xAI; threats to remove all AI from Tesla unless given full control of the company; funneling resources from Tesla to his private companies.[55]

2024年年中，有多份报告称，股东们正在起诉埃隆·马斯克和特斯拉董事会，指控其违反信托责任。这在一定程度上是因为马斯克威胁称，如果董事会拒绝让他持有25%的股份并拥有公司的绝对控制权，他将从特斯拉撤出所有AI。另一个原因是特斯拉的资源被转移到了xAI，实质上是马斯克将特斯拉的资源转移到了他的私人公司。[54]股东们认为，马斯克的违约行为是多方面的：使用特斯拉的资金为xAI提供资金；挖角特斯拉员工到xAI；威胁要撤出特斯拉的所有AI，除非获得公司的完全控制权；将特斯拉的资源转移到他的私人公司。 [55] 

Tesla’s Board had faced such accusations before, and were found guilty. In early 2024, the Delaware Court of Chancery ruled that Tesla directors breached their fiduciary duties when they awarded Elon Musk a compensation plan with an estimated maximum value of about $55 billion. In a description that was clearly bordering on fraud, the court also stated that Musk controlled Tesla, “at least with respect to the compensation plan”, and that disclosure to the shareholders was “materially flawed”, preventing a fully informed vote. The court ordered the entire plan withdrawn and all paid compensation returned.[56]

特斯拉董事会此前曾面临此类指控，并被判有罪。2024年初，特拉华州衡平法院裁定，特斯拉董事在授予埃隆·马斯克一项预计最高价值约550亿美元的薪酬计划时，违反了其受托责任。法院在一份明显近乎欺诈的描述中还表示，马斯克“至少在薪酬计划方面”控制着特斯拉，且向股东的披露“存在重大缺陷”，阻碍了股东在充分知情的情况下进行投票。法院下令撤销整个计划，并退还所有已支付的薪酬。 [56]

This is important for reasons that might not be immediately obvious. One important fact from an NBC Report is that the compensation lawsuit above was brought against the Tesla Board by a shareholder who held only 9 Tesla shares, a common “man in the street nobody”, the kind of person for whom Musk has a special contempt. It was therefore especially painful for Musk to have had $55 billion literally in his hands only to watch it taken away by an insect.[57]

这一点之所以重要，原因可能并不那么显而易见。美国全国广播公司（NBC）的一份报告揭示了一个重要事实：上述赔偿诉讼是由一名仅持有特斯拉9股股票的股东对特斯拉董事会提起的，这名股东是典型的“无名小卒”，正是那种让马斯克特别鄙视的人。因此，对于马斯克而言，眼睁睁地看着自己手中的550亿美元被一只“小虫”夺走，这尤其令人痛苦。 [57]

After this loss, Musk immediately had Tesla’s site of incorporation moved to Texas where the laws were not so fair. He immediately changed Tesla’s corporate bylaws to severely limit the ability of shareholders to sue Musk or the company for any further breach of fiduciary duties Now, a shareholder must own at least 3% of Tesla’s outstanding shares in order to institute such an action, and that means the person would have to own around $35 billion in Tesla shares. [58] “Fiduciary Duty” is a serious legal term which means you must act at all times in the best interest of the company. If I were to make such a lawsuit virtually impossible, it would be only because I had no intention of “acting in the best interests of the company”.

这次败诉后，马斯克立即将特斯拉的注册地迁至法律不那么公平的德克萨斯州。他立即修改了特斯拉的公司章程，严格限制股东因马斯克或公司进一步违反信托责任而起诉的能力。现在，股东必须持有特斯拉至少3%的已发行股份才能提起此类诉讼，这意味着该股东必须持有约350亿美元的特斯拉股份。[58]“信托责任”是一个严肃的法律术语，意味着你必须始终以公司的最大利益行事。如果我要让此类诉讼几乎不可能发生，那只是因为我没有“以公司最大利益行事”的意图。

金融欺诈 — The Financial Frauds

There are two root causes for the enormous fraudulent and confusing mess surrounding Musk’s xAI and Grok. The first, as I have mentioned several times, is that Elon Musk is using xAI as the primary weapon in his personal war with Sam Altman. The second is that Musk is badly losing that war and the financial cost is pervasive and staggering. xAI is burning through more than $1 billion in cash each month, the projected loss for 2025 around $13 billion.[59] This expense is offset only by the subscription revenue from Grok, which is less than $500 million per year. [60] In other words, expenses are about 27 times the revenue, with little apparent prospect for change.

马斯克的xAI和Grok所引发的巨大欺诈和混乱局面有两个根本原因。首先，正如我多次提到的，埃隆·马斯克正将xAI作为他与萨姆·奥特曼个人战争中的主要武器。其次，马斯克在这场战争中节节败退，财务损失无处不在且惊人。xAI每月烧钱超过10亿美元，预计2025年的亏损约为130亿美元。[59]这一支出仅由Grok的订阅收入抵消，而Grok的年收入不到5亿美元。[60]换言之，支出约为收入的27倍，且改变的前景渺茫。

The possibility of Grok’s revenue expanding to cover $13 billion or more of expenses seems exceedingly remote in the short term. And there are no signs to indicate a sudden huge increase, which doesn’t bode well for the long term. xAI does have the revenue from Twitter, but that is down by more than 50% and decreasing. Worse, there are no signs that Twitter will ever regain its previous popularity, with either users or advertisers.  And Musk’s offering of a sexually-explicit version of Grok may drive even more advertisers away.

短期内，Grok的收入扩大到足以覆盖130亿美元或更多支出的可能性似乎极其渺茫。而且，没有迹象表明收入会突然大幅增长，这对长期发展来说并非好兆头。xAI确实有来自Twitter的收入，但这一收入已下降超过50%，并且还在持续下降。更糟糕的是，没有迹象表明Twitter的用户或广告商数量会恢复到之前的水平。而马斯克推出性暴露版本的Grok可能会进一步赶走广告商。

xAI’s primary revenue comes from Grok subscriptions, but the sexually explicit “Spicy Mode” in Grok Imagine has already caused controversies by generating non-consensual deepfakes of Taylor Swift. [61][62] This creates reputational risk that further deters advertisers.  The sexually-explicit and pornographic content strategy for Grok risks serious permanent brand damage, just as it proved to be for Twitter. Meanwhile, new revenue experiments like the $300/month “SuperGrok Heavy” plan and blockchain-based “Valentine” companion are just useless gimmicks. The $300/month “Heavy” tier targets niche users, but adoption is untested among other better and cheaper alternatives (e.g., ChatGPT+ at $20).

xAI的主要收入来自Grok订阅服务，但Grok Imagine中露骨的“辣味模式”已经引发了争议，因为它生成了未经同意的泰勒·斯威夫特（Taylor Swift）的深度伪造视频。[61][62]这带来了声誉风险，进一步阻碍了广告商的投放。Grok的露骨色情内容策略可能会对其品牌造成严重的永久性损害，正如Twitter所经历的那样。与此同时，像每月300美元的“SuperGrok Heavy”计划和基于区块链的“Valentine”伴侣这样的新收入实验只是无用的噱头。每月300美元的“Heavy”套餐针对的是小众用户，但在其他更好、更便宜的替代品（如20美元的ChatGPT+）面前，其接受度尚未经过检验

By comparison, OpenAI’s ChatGPT earns 30x more than Grok on mobile, $1.35 billion in 2025 alone, highlighting Grok’s minor market share and showing how far behind Musk’s xAI really is. Even if Grok’s subscriptions grew by ten times, that would still barely cover 4% of total costs. According to the latest analysis by app intelligence provider Appfigures, ChatGPT’s iOS and Android apps have reached $2 billion in global consumer spending since its launch in May 2023, which is about 30 times the combined mobile lifetime consumption of competitors such as Claude, Copilot and Grok. Grok’s average monthly consumer spending is only about 2% of ChatGPT. From the perspective of download volume, ChatGPT’s lead is also obvious. The app has been installed about 690 million times worldwide so far, compared to only 39.5 million times on Grok. [63] As one comparison, ChatGPT produces about $10 of revenue for each install, while Grok produces about $0.75.

相比之下，OpenAI的ChatGPT在移动端的收入是Grok的30倍，仅2025年一年就达到13.5亿美元，这凸显了Grok市场份额之小，也表明马斯克的xAI确实远远落后。即使Grok的订阅量增长十倍，也仅能勉强覆盖总成本的4%。根据应用智能提供商Appfigures的最新分析，自2023年5月推出以来，ChatGPT的iOS和Android应用在全球消费者支出中已达到20亿美元，约为Claude、Copilot和Grok等竞争对手移动端终身消费总和的30倍。Grok的平均每月消费者支出仅为ChatGPT的约2%。从下载量来看，ChatGPT的领先地位同样明显。该应用在全球的安装量已达到约6.9亿次，而Grok仅为3950万次。[63]作为一项对比，ChatGPT每次安装可产生约10美元的收入，而Grok每次安装仅产生约0.75美元的收入。

Musk borrowed much of the $44 billion he paid for Twitter, and the now-reduced advertising revenue will barely cover the interest on his loans. And that means both xAI and Twitter (X) are burning cash. Musk is silent on Twitter’s user numbers and advertising revenues, which means there is no good news. If Twitter were doing well, Musk would be hyping it, so it is probably on a slow further decline. And, as Paris Mayor Anne Hidalgo stated, Musk has turned Twitter into “a global sewer”. So, not much chance of Twitter having a huge increase in revenue. Probably the opposite. And, if Grok were going to become hugely popular, the signs would already be obvious, but there are no such signs. Also, Musk appears determined to turn Grok into a global sewer with his “spicy” pornographic version, as he did with Twitter. The combined picture looks even worse than I first imagined.

马斯克为收购Twitter支付的440亿美元中，大部分是借来的，而如今广告收入减少，仅够支付他的贷款利息。这意味着xAI和Twitter（X）都在烧钱。马斯克对Twitter的用户数量和广告收入保持沉默，这意味着没有好消息。如果Twitter表现良好，马斯克会大肆宣传，所以它可能正在缓慢进一步衰退。而且，正如巴黎市长安妮·伊达尔戈所言，马斯克已将Twitter变成了“全球的下水道”。因此，Twitter的收入大幅增长的可能性不大。很可能恰恰相反。而且，如果Grok要大受欢迎，迹象本应已很明显，但目前没有这样的迹象。此外，马斯克似乎决心要用他“辣味”的色情版Grok，将其变成全球的下水道，就像他对Twitter所做的那样。综合来看，情况比我最初想象的还要糟糕。

With Twitter, due to the excessive amount of political content, the prevalent disinformation, the widespread bot traffic, Musk’s increasing toxicity, the marked increase in invalid traffic, the “global sewer” version, advertisers have degraded the value of Twitter as an advertising medium, and prices dropped. Twitter’s cost per thousand impressions has fallen by more than 75% to just $1.2, compared to $7 for META and $3 for TikTok.[64]

由于推特上充斥着过多的政治内容、大量虚假信息、广泛的机器人流量、马斯克日益加剧的负面影响力、无效流量的显著增加，以及“全球污水沟”般的版本，广告商降低了推特作为广告媒介的价值，导致价格下跌。与META的7美元和TikTok的3美元相比，推特的每千次展示成本下降了75%以上，仅为1.2美元。[64] 

The current interest rates for multinationals corporate debt range between 5% for top grade debt and about 7% for the lower grades. It is extraordinary that Musk is paying 12.5% for his xAI loans, which suggests “junk-bond” classification.  Twitter’s acquisition debt was 11%; the banks struggled to sell it, and it was finally unloaded at a discount, so this is worse. Multiple sources confirm xAI’s recent $5 billion debt financing includes $3 billion in bonds at 12.5% yield and a $1 billion fixed-rate loan at 12.5%. There is also a $1 billion floating-rate loan at ‘bank benchmark rate + 7.25%’ (effectively ~12.5%–13%).

目前跨国公司债务的利率范围为顶级债务5%至较低级别债务约7%。令人惊讶的是，马斯克为其xAI贷款支付的利率为12.5%，这表明其属于“垃圾债券”类别。Twitter的收购债务利率为11%；银行难以将其出售，最终以折扣价脱手，因此情况更糟。多方消息来源证实，xAI最近的50亿美元债务融资包括30亿美元收益率为12.5%的债券和10亿美元固定利率贷款，利率均为12.5%。此外，还有一笔10亿美元浮动利率贷款，利率为“银行基准利率+7.25%”（实际约为12.5%-13%）。

This is far above even BB-rated junk bonds, which are at about 7-9%.  Elon Musk’s xAI 12.5% interest is exceptionally high even when in “junk-bond” territory. As a comparison, investment-grade bonds typically carry an interest rate of 4%–6%, while high-yield “junk” bonds will average 7%–9%. Musk’s xAI paying 12.5% is ~5–8 percentage points higher than standard junk bonds, placing it in speculative/distressed territory. Some of xAI’s debt was sold at a discount, raising the interest cost even higher, and all of it was sold by the agencies at a “best-efforts” offering.

这甚至远高于BB级垃圾债券的利率，后者约为7%至9%。即便在“垃圾债券”的范畴内，埃隆·马斯克的xAI公司支付的12.5%的利率也异常高。相比之下，投资级债券的利率通常为4%至6%，而高收益“垃圾”债券的平均利率为7%至9%。马斯克的xAI公司支付的12.5%的利率比标准垃圾债券高出约5至8个百分点，使其处于投机/困境债券的范畴。xAI公司的一部分债务以折扣价出售，进一步推高了利息成本，而且所有债务都是由评级机构以“尽力”发行的形式出售的。

Normally the underwriter – a bank or stock broker – will purchase all the new issued debt or shares with their own money, then sell the securities to the public. Obvious in this is an assumption that the securities can be sold – that the public will buy them. But in the case of xAI, the underwriting bank would take no part of the investment and would offer no underwriting guarantees, promising only to use “our best efforts” to sell these securities. Hence the “junk-bond” analogy; the extremely high interest rates and the discount pricing are classic tactics when dealing with what is termed “distressed-debt”. This is partly because xAI’s debt is unrated and few want to touch it, partly because of the huge cash burn, and partly Musk’s increasing personal toxicity and loss of “wunderkind” status.

通常，承销商（银行或股票经纪人）会用自己的资金购买所有新发行的债券或股票，然后向公众出售这些证券。这显然是基于一个假设，即这些证券能够被出售——公众会购买它们。但在xAI的案例中，承销银行不会参与投资，也不会提供承销担保，只是承诺会“尽最大努力”出售这些证券。因此，将其比作“垃圾债券”也就不足为奇了；极高的利率和折价定价是处理所谓“不良债务”时的经典策略。这在一定程度上是因为xAI的债务未评级，很少有人愿意碰它，部分原因是其巨大的现金消耗，还有部分原因是马斯克个人负面新闻不断，失去了“神童”的地位。

There are clearly serious concerns about the sustainability of X (Twitter) and xAI given the massive debt and declining revenues, and the reputational and brand safety issues don’t help. Twitter’s ad revenue is collapsing, down from more than $4.5 billion per year pre-Musk to below $2 billion in 2025, and still falling. Grok is critically failing to produce revenue and the “global sewer” reputation isn’t helping.  Grok’s NSFW pivot is a foolish Hail Mary that has backfired by further alienating advertisers and Musk’s personal brand toxicity is another huge negative. These two combined entities appear to lack any credible path to profitability, making Musk’s hype a dangerous distraction from the math. With these, Musk is facing a severe debt servicing crisis, which helps to explain his cannibalisation of Tesla and SpaceX.

鉴于巨额债务和收入下滑，以及声誉和品牌安全问题雪上加霜，X（Twitter）和xAI的可持续性显然令人深感担忧。Twitter的广告收入正在崩溃，从马斯克入主前的每年超过45亿美元降至2025年的不到20亿美元，且仍在下降。Grok在创造收入方面严重失败，“全球下水道”的名声更是雪上加霜。Grok的NSFW（不适合在工作场合）转型是一个愚蠢的孤注一掷之举，结果适得其反，进一步疏远了广告商，而马斯克个人品牌的负面影响力则是另一个巨大的不利因素。这两个合并后的实体似乎缺乏任何可靠的盈利途径，这使得马斯克的炒作成为一种危险的干扰，让人忽视了事实。因此，马斯克正面临严重的偿债危机，这有助于解释他为何要蚕食特斯拉和SpaceX的业务。

The combination of xAI and Twitter (X) contains too many structural flaws to provide hope for recovery. Twitter has so diminished in ads and subscriptions that it is becoming irrelevant. And in the best case, Grok seems destined to become a sort of pornographic niche product, like an OnlyFans for AI, producing some revenue from the NSFW (Not Safe For Work) crowd, but with little growth. In the worst case, considering the advertiser exodus, minimal subscriptions, regulatory fines, and potential debt defaults, this could easily cascade into bankruptcy by 2026–2027. xAI, Grok and Twitter are increasingly looking like a house of cards nearing collapse. And, as I have said before, it couldn’t happen to a nicer guy.

xAI与Twitter（X）的结合存在太多结构性缺陷，无法带来复苏的希望。Twitter在广告和订阅方面的表现如此糟糕，以至于它正变得无关紧要。在最乐观的情况下，Grok似乎注定只能成为一种色情小众产品，就像AI界的OnlyFans，从NSFW（不适合在工作场合观看）人群那里获得一些收入，但增长缓慢。在最糟糕的情况下，考虑到广告商大量流失、订阅量极低、监管罚款和潜在的债务违约，到2026-2027年，该公司很可能陷入破产。xAI、Grok和Twitter越来越像是一座即将倒塌的纸牌屋。而且，正如我之前所说，这种事不可能发生在好人身上。

估值欺诈 — The Valuation Fraud

Elon Musk paid $44 billion for Twitter (about twice its actual value) but managed through his own efforts to destroy the value, driving it down by most estimates to less than $10 billion. The existing shareholders were of course dismayed to see the value of their investment degraded by 75% or 80%. To avoid public embarrassment, Musk decided to fold Twitter into his xAI, offering to convert the shareholder equity in Twitter to a 25% share of the total of xAI. Of course, their only hope was that xAI would hugely increase in value, permitting them to recover their investment.

埃隆·马斯克以440亿美元收购了Twitter（约为其实际价值的两倍），但通过自己的努力设法摧毁了其价值，据大多数估计，将其价值压低至不到100亿美元。现有股东当然很失望地看到他们的投资价值贬值了75%或80%。为了避免公开尴尬，马斯克决定将Twitter并入他的xAI，提议将Twitter的股东权益转换为xAI总股份的25%。当然，他们唯一的希望是xAI的价值会大幅增加，使他们能够收回投资。

Eager to please, Musk fraudulently claimed the value of Twitter as $33 billion in the merger, and further made frequent public claims as to the (imaginary) value of xAI. It seems that with each $5 billion in new funding, xAI increased in value by about $25 billion or $30 billion, the value quickly reaching $130 billion, then $180 billion, then $300 billion and finally $500 billion. This may have made the investors feel better, but there was never any justification for these huge increases in assumed value. And indeed, there was never any apparent justification for the original stated “value” of $130 billion. Perhaps I am too cynical, but my guess is that xAI is worth much less than even $100 billion, and perhaps as little as $50 billion, or maybe less. Since the stock is not traded publicly, we will probably have to wait for a bankruptcy court to determine a real value.

马斯克急于取悦众人，在合并中欺诈性地声称Twitter的价值为330亿美元，并进一步频繁公开宣称xAI（虚构）的价值。似乎每获得50亿美元的新资金，xAI的价值就会增加约250亿美元或300亿美元，很快达到1300亿美元，然后是1800亿美元，接着是3000亿美元，最后是5000亿美元。这或许让投资者感觉好些，但这些假设价值的巨大增长从来没有任何正当理由。事实上，最初声明的1300亿美元“价值”也从未有任何明显的正当理由。也许是我太愤世嫉俗了，但我的猜测是，xAI的价值甚至远低于1000亿美元，可能只有500亿美元，甚至更低。由于该股票未公开交易，我们可能不得不等待破产法院来确定其真实价值。

*

Mr. Romanoff’s writing has been translated into 34 languages and his articles posted on more than 150 foreign-language news and politics websites in more than 30 countries, as well as more than 100 English language platforms. Larry Romanoff is a retired management consultant and businessman. He has held senior executive positions in international consulting firms, and owned an international import-export business. He has been a visiting professor at Shanghai’s Fudan University, presenting case studies in international affairs to senior EMBA classes. Mr. Romanoff lives in Shanghai and is currently writing a series of ten books generally related to China and the West. He is one of the contributing authors to Cynthia McKinney’s new anthology ‘When China Sneezes’. (Chap. 2 — Dealing with Demons).

罗曼诺夫先生的作品已被翻译成34种语言，他的文章被发布在30多个国家的150多个外文新闻和政治网站上，以及100多个英文平台上。拉里·罗曼诺夫是一位退休的管理顾问和商人。他曾在国际咨询公司担任高级管理职务，并拥有一家国际进出口企业。他曾担任上海复旦大学的客座教授，为高级EMBA课程讲授国际事务案例研究。罗曼诺夫先生现居上海，目前正在撰写一系列共十本书，总体上涉及中国与西方。他是辛西娅·麦金尼新选集《当中国打喷嚏》（第2章——与恶魔打交道）的特约作者之一。

His full archive can be seen at

他的全部档案可以在以下网址查看：

https://www.bluemoonofshanghai.com/  + https://www.moonofshanghai.com/

He can be contacted at:

他的全部档案可以在以下网址查看：

2186604556@qq.com

*

NOTES Part 18
注释 第18部分

[01] Reuters: Tesla has been faking battery life since 10 years ago!
https://so.html5.qq.com/page/real/search_news?docid=70000021_61264c4e0ee91352&faker=1
[01] 路透社：特斯拉自10年前起就一直在虚报电池续航里程！
https://so.html5.qq.com/page/real/search_news?docid=70000021_61264c4e0ee91352&faker=1
[1] Debunking Elon Musk – Part 12 — xAI and Grok
https://www.bluemoonofshanghai.com/politics/21776/
[1] 揭秘埃隆•马斯克——第12部分 — xAI与Grokhttps
://www.bluemoonofshanghai.com/politics/21776/
[2] Musk on his companies
https://www.douyin.com/video/7500652509683797286
[2] 马斯克谈及他的公司
https://www.douyin.com/video/7500652509683797286
[3] Elon Musk’s Grok 3 Was Told to Ignore Sources Saying He Spread Misinformation
https://futurism.com/grok-elon-instructions
[3] 埃隆•马斯克的Grok 3被告知要忽略那些说他散布错误信息的消息来源
https://futurism.com/grok-elon-instructions
[4] “Truth-seeking” Grok 4 under fire for seemingly prioritizing Elon Musk’s views
https://techissuestoday.com/truth-seeking-grok-4-under-fire-for-seemingly-prioritizing-elon-musks-views/
[4] “求真”的Grok 4因似乎优先考虑埃隆•马斯克的观点而受到抨击
https://techissuestoday.com/truth-seeking-grok-4-under-fire-for-seemingly-prioritizing-elon-musks-views/
[5 Grok 4 seems to consult Elon Musk to answer controversial questions
https://techcrunch.com/2025/07/10/grok-4-seems-to-consult-elon-musk-to-answer-controversial-questions/
[5 Grok 4似乎会向埃隆•马斯克请教，以回答有争议的问题
https://techcrunch.com/2025/07/10/grok-4-seems-to-consult-elon-musk-to-answer-controversial-questions/]
[6 “Truth-seeking” Grok 4 under fire for seemingly prioritizing Elon Musk’s views
https://techissuestoday.com/truth-seeking-grok-4-under-fire-for-seemingly-prioritizing-elon-musks-views/
[6 “求真”的Grok 4因似乎优先采纳埃隆•马斯克的观点而受到抨击
https://techissuestoday.com/truth-seeking-grok-4-under-fire-for-seemingly-prioritizing-elon-musks-views/
[7] AI gone rogue: Elon Musk’s own chatbot Grok accuses him for censoring it
https://hellboundanddown.com/2025/08/13/ai-gone-rouge-elon-musks-own-chatbot-grok-accuses-him-for-censoring-it-says-israel-and-us-committing-genocide-in-gaza/
[7] AI失控：埃隆•马斯克的聊天机器人Grok指责他审查内容
https://hellboundanddown.com/2025/08/13/ai-gone-rouge-elon-musks-own-chatbot-grok-accuses-him-for-censoring-it-says-israel-and-us-committing-genocide-in-gaza/
[8] Elon Musk’s Grok Got Suspended on Twitter
https://www.vice.com/en/article/elon-musks-grok-got-suspended-on-twitter-x/
[8] 伊隆•马斯克的Grok账号在推特上被暂停
https://www.vice.com/en/article/elon-musks-grok-got-suspended-on-twitter-x/
[9] Elon Musk’s AI bot Grok briefly suspended from X after Gaza genocide comments
https://www.naturalnews.com/2025-08-14-grok-suspended-from-x-after-gaza-genocide-comments.html
[9] 埃隆•马斯克的AI机器人Grok因发表加沙种族灭绝言论被X平台短暂封禁
https://www.naturalnews.com/2025-08-14-grok-suspended-from-x-after-gaza-genocide-comments.html
[10] “Hypocrisy noted,” Grok called out Elon Musk while the billionaire was arguing with Sam Altman over accusations against Apple
https://www.gamepressure.com/newsroom/hypocrisy-noted-grok-called-out-elon-musk-while-the-billionaire-w/z483ed
[10]“虚伪已记录，”格罗克（Grok）在亿万富翁埃隆•马斯克（Elon Musk）与萨姆•奥特曼（Sam Altman）就针对苹果的指控发生争执时，对马斯克大声说道。
https://www.gamepressure.com/newsroom/hypocrisy-noted-grok-called-out-elon-musk-while-the-billionaire-w/z483ed
[11] Elon Musk’s GrokAI betrays him in fight against Sam Altman, accuses Musk of boosting his posts on X
https://www.moneycontrol.com/technology/elon-musk-s-grokai-betrays-him-in-fight-against-sam-altman-accuses-musk-of-boosting-his-posts-on-x-article-13448930.html#goog_rewarded
[11] 伊隆•马斯克的GrokAI在反击萨姆•奥特曼时出卖了自己，指责马斯克在X平台上为其帖子点赞
https://www.moneycontrol.com/technology/elon-musk-s-grokai-betrays-him-in-fight-against-sam-altman-accuses-musk-of-boosting-his-posts-on-x-article-13448930.html#goog_rewarded
[12] Grok is getting updated because it sided with Sam Altman
https://dataconomy.com/2025/08/13/grok-is-getting-updated-because-it-sided-with-sam-altman/
[12] Grok正在更新，因为它与Sam Altman立场一致
https://dataconomy.com/2025/08/13/grok-is-getting-updated-because-it-sided-with-sam-altman/
[13] X fixes Grok after chatbot sides with Sam Altman
https://techbriefly.com/2025/08/13/x-fixes-grok-after-chatbot-sides-with-sam-altman/
[13] 在聊天机器人支持Sam Altman后，X修复了Grokhttps
://techbriefly.com/2025/08/13/x-fixes-grok-after-chatbot-sides-with-sam-altman/
[14] Grok4 hasn’t lost in score, hasn’t won in safety
https://www.goupsec.com/news/19434.html
[14] Grok4在得分上未尝败绩，在安全方面也未尝胜绩
https://www.goupsec.com/news/19434.html
[15] Musk criticized OpenAI: Training AI does not pursue the truth to the greatest extent
https://www.antpedia.com/news/99/n-3174499.html
[15] 马斯克批评OpenAI：训练AI并未最大程度地追求真理
https://www.antpedia.com/news/99/n-3174499.html
[16] New challenges in AI evaluation
https://www.sohu.com/a/924463315_362225
[16] 人工智能评估的新挑战
https://www.sohu.com/a/924463315_362225
[17] xAI Was About to Land a Major Government Contract. Then Grok Praised Hitler
https://www.wired.com/story/xai-grok-government-contract-hitler/
[17] xAI即将获得一份重要的政府合同。然后Grok却称赞了希特勒
https://www.wired.com/story/xai-grok-government-contract-hitler/
[18] US gov’t ditches Musk’s AI. The Grok chatbot has reportedly been removed from a federal program after sparking controversy over offensive content.
https://www.rt.com/news/622947-musk-ai-grok-us/
[18] 美国政府放弃马斯克的AI技术。据报道，在引发关于冒犯性内容的争议后，Grok聊天机器人已被从一个联邦项目中移除。
https://www.rt.com/news/622947-musk-ai-grok-us/
[19] US government agency drops Grok after MechaHitler backlash
https://arstechnica.com/tech-policy/2025/08/us-government-agency-drops-grok-after-mechahitler-backlash-report-says/
[19] 美国政府机构因《MechaHitler》引发强烈反对后放弃使用Grokhttps
://arstechnica.com/tech-policy/2025/08/us-government-agency-drops-grok-after-mechahitler-backlash-report-says/
[20] xAI’s Grok Loses Government Contract Amid Antisemitic Controversy
https://theoutpost.ai/news-story/us-government-drops-x-ai-s-grok-after-antisemitic-incident-reshaping-federal-ai-adoption-19114/
[20] xAI的Grok因反犹太主义争议失去政府合同
https://theoutpost.ai/news-story/us-government-drops-x-ai-s-grok-after-antisemitic-incident-reshaping-federal-ai-adoption-19114/
[21] Grok App Adds AI Image and Video Generator With NSFW ‘Spicy’ Mode
https://www.pcmag.com/news/grok-app-adds-ai-image-and-video-generator-with-nsfw-spicy-mode
[21] Grok App推出AI图像和视频生成器，附带NSFW“辣”模式
https://www.pcmag.com/news/grok-app-adds-ai-image-and-video-generator-with-nsfw-spicy-mode
[22] Consumer safety groups are demanding an FTC investigation into Grok’s ‘Spicy’ mode
https://www.theverge.com/x-ai/759554/consumer-safety-groups-are-demanding-an-ftc-investigation-into-groks-spicy-mode-elon-musk-grok-imagine-xai
[22] 消费者安全组织要求联邦贸易委员会（FTC）对Grok的“Spicy”模式进行调查
https://www.theverge.com/x-ai/759554/consumer-safety-groups-are-demanding-an-ftc-investigation-into-groks-spicy-mode-elon-musk-grok-imagine-xai
[23] After smashing 200,000 GPUs, Musk’s GROK3 still overturned?
https://baijiahao.baidu.com/s?id=1825170495603896559&wfr=spider&for=pc
[23] 在击毁20万个GPU后，马斯克的GROK3仍然被推翻？
https://baijiahao.baidu.com/s?id=1825170495603896559&wfr=spider&for=pc
[24] Musk led the xAI team to release Grok 4 “the world’s strongest model”
https://finance.eastmoney.com/a/202507103453366044.html
[24] 马斯克领导xAI团队发布了Grok 4，被誉为“全球最强大的模型”
https://finance.eastmoney.com/a/202507103453366044.html
[25] Elon Musk’s Grok Got Suspended on Twitter/X
https://www.vice.com/en/article/elon-musks-grok-got-suspended-on-twitter-x/
[25] 伊隆•马斯克的Grok在Twitter上被暂停使用/Xhttps
://www.vice.com/en/article/elon-musks-grok-got-suspended-on-twitter-x/
[26] AI gone rogue: Elon Musk’s own chatbot Grok accuses him for censoring it
https://hellboundanddown.com/2025/08/13/ai-gone-rouge-elon-musks-own-chatbot-grok-accuses-him-for-censoring-it-says-israel-and-us-committing-genocide-in-gaza/
[26] AI失控：埃隆•马斯克的聊天机器人Grok指责他审查自己
https://hellboundanddown.com/2025/08/13/ai-gone-rouge-elon-musks-own-chatbot-grok-accuses-him-for-censoring-it-says-israel-and-us-committing-genocide-in-gaza/
[27] “Hypocrisy noted,” Grok called out Elon Musk while the billionaire was arguing with Sam Altman over accusations against Apple
https://www.gamepressure.com/newsroom/hypocrisy-noted-grok-called-out-elon-musk-while-the-billionaire-w/z483ed
[27]“虚伪已记下”，当亿万富翁埃隆•马斯克与萨姆•奥特曼就针对苹果的指控发生争执时，格罗克（Grok）大声喊出了这句话。
https://www.gamepressure.com/newsroom/hypocrisy-noted-grok-called-out-elon-musk-while-the-billionaire-w/z483ed
[28] Musk led the xAI team to release Grok 4 “the world’s strongest model” What is the gold content?
https://finance.eastmoney.com/a/202507103453366044.html
[28] 马斯克带领xAI团队发布了Grok 4，号称“全球最强模型”。其含金量究竟如何？
https://finance.eastmoney.com/a/202507103453366044.html
[29] Musk boasted that xAI will deploy AI computing power within five years, equivalent to 50 million H100 GPUs to challenge the energy limit
https://m.dtm.com.cn/news/202507/123966.html
[29] 马斯克夸口称，xAI将在五年内部署相当于5000万个H100 GPU的AI算力，以挑战能源极限
https://m.dtm.com.cn/news/202507/123966.html
[30] Musk’s claimed “strongest AI training system” was launched
https://ecas.cas.cn/dtfb/yjdt/202501/t20250106_5044302.html
[30] 马斯克声称的“最强大的人工智能训练系统”已推出
https://ecas.cas.cn/dtfb/yjdt/202501/t20250106_5044302.html
[31] Musk: The strongest AI training cluster in history will be launched, adding another 100,000 GPUs
https://baijiahao.baidu.com/s?id=1809251543115999515&wfr=spider&for=pc
[31] 马斯克：史上最强大的AI训练集群即将上线，再添10万个GPUhttps
://baijiahao.baidu.com/s?id=1809251543115999515&wfr=spider&for=pc
[32] Musk’s claimed “strongest AI training system” was launched
https://finance.sina.com.cn/tjhz/2024-09-09/doc-incnnspw3121327.shtml
[32] 马斯克宣称的“最强大的人工智能训练系统”已推出
https://finance.sina.com.cn/tjhz/2024-09-09/doc-incnnspw3121327.shtml
[33] Memphis warns it may not be able to power Elon Musk’s lofty ‘Colossus’ supercomputer expansion plans
https://fortune.com/2025/01/07/memphis-utility-ceo-warns-power-supply-elon-musk-xai-colossus-supercomputer-facility/
[33] 孟菲斯电力公司警告称，可能无法为埃隆•马斯克（Elon Musk）宏伟的“巨像”（Colossus）超级计算机扩张计划提供电力
https://fortune.com/2025/01/07/memphis-utility-ceo-warns-power-supply-elon-musk-xai-colossus-supercomputer-facility/
[34] Musk’s xAI supercomputer project faces power problems
https://www.icspec.com/news/article-details/2428556
[34] 马斯克的xAI超级计算机项目面临电力问题
https://www.icspec.com/news/article-details/2428556
[35] Musk’s AI supercomputing details exposed: 2.9 billion [Yuan] have been invested and the power gap is large
https://tech.ifeng.com/c/8iDsXCPnZYb?ch=ttsearch
[35] 马斯克的人工智能超级计算细节曝光：已投资29亿元，电力缺口大
https://tech.ifeng.com/c/8iDsXCPnZYb?ch=ttsearch
[36] Musk’s xAI supercomputer project faces power problems
https://www.icspec.com/news/article-details/2428556
[36] 马斯克的xAI超级计算机项目面临电力问题
https://www.icspec.com/news/article-details/2428556
[37] Musk’s xAI supercomputer project faces power problems
https://www.icspec.com/news/article-details/2428556
[37] 马斯克的xAI超级计算机项目面临电力问题
https://www.icspec.com/news/article-details/2428556
[38] Musk’s AI supercomputing cost 2.9 billion, and millions of GPUs are facing the challenge of power gap!
https://m.dtm.com.cn/news/202504/87333.html
[38] 马斯克的AI超级计算耗资29亿美元，数百万个GPU正面临电力缺口挑战！
https://m.dtm.com.cn/news/202504/87333.html
[39] ‘How come I can’t breathe?’
https://www.politico.com/news/2025/05/06/elon-musk-xai-memphis-gas-turbines-air-pollution-permits-00317582
[39]“我怎么就喘不过气来了？”
https://www.politico.com/news/2025/05/06/elon-musk-xai-memphis-gas-turbines-air-pollution-permits-00317582
[40] Why Elon Musk, the bully, is seen by many, including liberals, as a hero
https://www.latimes.com/opinion/story/2023-12-04/elon-musk-racism-x-bully-savior-culture
[40] 为什么恶霸埃隆•马斯克被许多人，包括自由主义者，视为英雄
https://www.latimes.com/opinion/story/2023-12-04/elon-musk-racism-x-bully-savior-culture
[41] ‘How come I can’t breathe?’
https://www.politico.com/news/2025/05/06/elon-musk-xai-memphis-gas-turbines-air-pollution-permits-00317582
[41] “我怎么就喘不过气来了？”
https://www.politico.com/news/2025/05/06/elon-musk-xai-memphis-gas-turbines-air-pollution-permits-00317582
[42] Elon Musk’s xAI will shut down half of the generators that power the super data center Colossus
https://cbismb.com/infrastructure/59852
[42] 伊隆•马斯克的xAI将关闭为超级数据中心Colossus供电的一半发电机
https://cbismb.com/infrastructure/59852
[43] The end of AI is electricity! Computing power factories have become major power consumers, and Musk has been accused of “illegal power generation”
https://www.sohu.com/a/884096150_115433
[43] 人工智能的终结是电力！计算能力工厂已成为主要的电力消费者，马斯克被指控“非法发电”
https://www.sohu.com/a/884096150_115433
[44] Musk xAI was exposed to violations of Memphis supercomputing electricity, and the actual turbines far exceeded 15 to 35
https://www.rjtj.cn/news/show-36287.html
[44] Musk xAI被曝违反孟菲斯超级计算用电规定，实际涡轮机数量远超15至35台
https://www.rjtj.cn/news/show-36287.html
[45] The first phase of Musk’s Colossus: 200,000 GPUs, 150 megawatts of battery backup energy guarantee
https://m.itbear.com.cn/html/2025-05/813565.html
[45] 马斯克巨像计划第一阶段：20万块GPU，150兆瓦备用电池能源保障
https://m.itbear.com.cn/html/2025-05/813565.html
[46] Musk “can’t let go” Tesla and bought millions of chips to build an xAI giant factory
https://www.stcn.com/article/detail/1839123.html
[46] 马斯克“放不下”特斯拉，斥资数百万购买芯片，打造xAI巨型工厂
https://www.stcn.com/article/detail/1839123.html
[47] Musk’s claimed “strongest AI training system” was launched
https://finance.sina.com.cn/tjhz/2024-09-09/doc-incnnspw3121327.shtml
[47] 马斯克宣称的“最强大的人工智能训练系统”已推出
https://finance.sina.com.cn/tjhz/2024-09-09/doc-incnnspw3121327.shtml
[48] Musk responded by exaggerating reports about Tesla’s AI chip procurement plan
https://finance.sina.cn/2024-06-05/detail-inaxrpyn6213122.d.html
[48] 马斯克对有关特斯拉AI芯片采购计划的报道作出回应，称其夸大其词
https://finance.sina.cn/2024-06-05/detail-inaxrpyn6213122.d.html
[49] Musk transferred thousands of AI chips reserved for Tesla to X and xAI
https://baijiahao.baidu.com/s?id=1800939268369567504
[49] 马斯克将数千块原本为特斯拉预留的人工智能芯片转移给了X和xAIhttps
://baijiahao.baidu.com/s?id=1800939268369567504
[50] Musk transferred Tesla’s 12,000 Nvidia H100 to X
https://zhidx.com/news/42202.html
[50] 马斯克将特斯拉的12,000个英伟达H100转移至Xhttps
://zhidx.com/news/42202.html
[51] Tesla Unwinds Dojo Supercomputer Team Following Exec Exodus
https://www.pcmag.com/news/tesla-unwinds-dojo-supercomputer-team-following-exec-exodus
[51] 特斯拉解散Dojo超级计算机团队，多名高管离职
https://www.pcmag.com/news/tesla-unwinds-dojo-supercomputer-team-following-exec-exodus
[52] Tesla’s Dojo supercomputer looks dead as more execs leave for competing startup
https://electrek.co/2025/08/07/tesla-dojo-supercomputer-looks-dead-as-more-execs-leave-competing-startup/
[52] 随着更多高管离开，加盟竞争对手的初创公司，特斯拉的Dojo超级计算机似乎已陷入停滞
https://electrek.co/2025/08/07/tesla-dojo-supercomputer-looks-dead-as-more-execs-leave-competing-startup/
[53] US media revealed Musk’s crazy approach: sacrificing Tesla and others to provide resources for the development of xAI!
https://baijiahao.baidu.com/s?id=1807147394831434664&wfr=spider&for=pc
[53] 美国媒体揭露马斯克的疯狂做法：牺牲特斯拉等公司为xAI的发展提供资源！
https://baijiahao.baidu.com/s?id=1807147394831434664&wfr=spider&for=pc
[54] Tesla, Elon Musk sued by shareholders for breach of fiduciary duty over AI funneling threats
https://electrek.co/2024/06/13/tesla-elon-musk-sued-shareholders-breach-fiduciary-duty-ai-funneling-threats/
[54] 特斯拉、埃隆•马斯克因在人工智能方面存在潜在威胁而被股东起诉违反信托责任
https://electrek.co/2024/06/13/tesla-elon-musk-sued-shareholders-breach-fiduciary-duty-ai-funneling-threats/
[55] US media revealed Musk’s crazy approach: sacrificing Tesla and others to provide resources for the development of xAI!
https://baijiahao.baidu.com/s?id=1807147394831434664&wfr=spider&for=pc
[55] 美国媒体揭露马斯克的疯狂做法：牺牲特斯拉等公司为xAI的发展提供资源！
https://baijiahao.baidu.com/s?id=1807147394831434664&wfr=spider&for=pc
[56] Chancery Finds Tesla Board Breached Fiduciary Duties
https://corpgov.law.harvard.edu/2024/02/13/chancery-finds-tesla-board-breached-fiduciary-duties/
[56] 衡平法院裁定特斯拉董事会违反了信托责任
https://corpgov.law.harvard.edu/2024/02/13/chancery-finds-tesla-board-breached-fiduciary-duties/
[57] Tesla limits investors’ ability to sue over breach of fiduciary duties
https://www.nbcchicago.com/news/business/money-report/tesla-limits-investors-ability-to-sue-over-breach-of-fiduciary-duties/3748014/
[57] 特斯拉限制投资者就违反信托责任提起诉讼的能力
https://www.nbcchicago.com/news/business/money-report/tesla-limits-investors-ability-to-sue-over-breach-of-fiduciary-duties/3748014/
[58] Tesla limits investors’ ability to sue over breach of fiduciary duties
https://www.nbcchicago.com/news/business/money-report/tesla-limits-investors-ability-to-sue-over-breach-of-fiduciary-duties/3748014/
[58] 特斯拉限制投资者就违反信托责任提起诉讼的能力
https://www.nbcchicago.com/news/business/money-report/tesla-limits-investors-ability-to-sue-over-breach-of-fiduciary-duties/3748014/
[59] Musk’s xAI costs $1 billion per month and will reach $13 billion in annual fees
https://ai.zol.com.cn/999/9992990.html
[59] 马斯克的xAI每月花费10亿美元，年费将达到130亿美元
https://ai.zol.com.cn/999/9992990.html
[60] Musk xAI burns money fiercely: with an annual expenditure of $13 billion and an income of only 500 million, how to fill the funding gap?
https://www.itbear.com.cn/html/2025-06/858975.html
[60] Musk xAI烧钱如流水：年支出130亿美元，收入仅5亿美元，资金缺口如何填补？
https://www.itbear.com.cn/html/2025-06/858975.html
[61] Musk “evil” Grok, Taylor Swift became a victim
https://www.sohu.com/a/922669269_211762
[61] 马斯克“恶搞”泰勒•斯威夫特，后者成为受害者
https://www.sohu.com/a/922669269_211762
[62] Musk’s xAI is mired in controversy: generating indecent videos of celebrities raises legal and moral questions
https://www.sohu.com/a/922666807_362225
[62] 马斯克的xAI陷入争议：生成名人不雅视频引发法律和道德问题
https://www.sohu.com/a/922666807_362225
[63] ChatGPT’s mobile revenue exceeded $2 billion, crushing competitors to create a 30-fold revenue gap
https://www.chinaz.com/ainews/20568.shtml
[63] ChatGPT的移动端收入超过20亿美元，碾压竞争对手，创造了30倍的收入差距
https://www.chinaz.com/ainews/20568.shtml
[64] Musk, from the dragon, still can’t save X’s advertising business.
https://www.tojoyun.com/news/zixun/67464a440f43c30924fb6431
[64] 马斯克虽来自龙族，但仍无法拯救X的广告业务。
https://www.tojoyun.com/news/zixun/67464a440f43c30924fb6431

*

*

This article may contain copyrighted material, the use of which has not been specifically authorised by the copyright owner. This content is being made available under the Fair Use doctrine, and is for educational and information purposes only. There is no commercial use of this content.

本文可能包含受版权保护的内容，其使用并未获得版权所有者的明确授权。本文内容根据合理使用原则提供，仅用于教育和信息目的。本文内容不得用于商业用途。

本作者的其他作品