曾经，我有一些问题：/tech/887899/spacex-ipo-risks-ai 关于为什么 SpaceX，埃隆·马斯克最健康的公司，会收购 xAI：/tech/872619/elon-musk-merges-spacex-with-xai-and-x，他最虚弱的公司。现在我有一些问题，为什么我们把整个事情都叫做 SpaceX。
看，根据 SpaceX 作为上市公司的第一份季度收益声明：/science/975335/spacex-made-more-money-as-a-neocloud，从收入来看，我们这里主要是一家电信公司和一家计算租赁公司。本季度，公司太空业务收入未突破十亿美元，仅贡献了公司收入的略高于 10%。SpaceX 仍是自身最大的客户。没有足够的其他人想要它的火箭。而且，我不知道，也许如果火箭是公司主要关注点，SpaceX 就不会面临用太空垃圾在月球上炸出新坑的危险：https://www.theguardian.com/science/2026/aug/05/spacex-rocket-moon-crash-impact-falcon-9。
然而，大部分支出（以及炒作）仍然集中在被礼貌地称为“人工智能”的领域。
电信部分——SpaceX 称之为“连接服务”——是 Starlink 卫星互联网服务，收入为 42 亿美元，是 SpaceX 唯一没有运营亏损的部门。在电话会议上，Gwynne Shotwell 描述了一个计划，打算推出与 AT&T、Verizon 和 T-Mobile 竞争的电话服务，但大部分支出（以及炒作）仍然集中在被礼貌地称为“人工智能”的领域，以体现 Grok 可能做出的贡献。分析师 Alexander Potter 预计，明年用于新云业务的支出将攀升至 650 亿美元，比他之前估计的多 170 亿美元，彭博社指出：https://www.bloomberg.com/news/articles/2026-08-04/spacex-exceeds-revenue-estimates-in-first-earnings-since-ipo。
很难忽视租赁数据中心空间对SpaceX的重要性。不仅它比SpaceX的火箭赚得更多，还推动了更多支出——仅第二季度在AI上的花费就达158亿美元。（相比之下，航天和互联网络领域的支出各自略超过十亿美元。）该业务使它成为像CoreWeave和Nebius这样的新云公司的竞争对手。
只是那原本不是计划。
马斯克为Grok建造了SpaceX的Colossus 1数据中心，该AI偶尔也自称为MechaHitler，并有在未获同意的情况下剥衣女性和儿童的倾向。但xAI难以运行这个复杂系统，于是决定将其出租。此外，由于延迟问题很难训练内部模型，该中心使用了新旧混合芯片，造成瓶颈。在财报电话会上，马斯克表示SpaceX建造的算力中只有10%会用于Grok。
收入不是利润，建造数据中心成本很高
SpaceX目前与谷歌、Anthropic、Reflection AI以及马斯克最终选择收购的AI公司Cursor达成了合作。在第二季度财报电话会上，SpaceX首席财务官Bret Johnsen表示，这些交易使公司“走上轨道，包括Cursor的贡献，以达到1000亿美元的ARR（年度化经常性收入）水平”，ARR是一种基于短期估算来预测公司全年收入的财务指标。马斯克更为乐观，他表示，“12月达到1000亿美元的ARR不是疑问”，实际ARR可能更高。
这听起来很不错。但收入不是利润，建造数据中心成本很高。
涉足裸机租赁计算业务会有一些问题：/ai-artificial-intelligence/848988/nvidia-chip-loans-coreweave-gpu-debt-ai-neocloud ——具体来说，就是不可避免的过时、建设过程的不确定性，以及计算基本上是一种商品，这意味着公司在成本上竞争。建设的数据中心越多，可用的计算能力越多。可用的计算能力越多，公司就无法对其芯片收取那么高的价格。
马斯克曾声称，他将 SpaceX 上市是因为他想在太空中建设数据中心。现在，我可能会认为，如果我在地面运行数据中心时遇到困难，我会先在地面提高能力，再尝试做难度高一个数量级的事情：/ai-artificial-intelligence/845453/space-data-centers-astronomers ——甚至是前所未有的事情。但我想，这并不是马斯克思考问题的方式。
这一切听起来很昂贵，不是吗？
SpaceX 甚至向联邦通信委员会提出了一个轨道数据中心的计划：https://fccprod.servicenowservices.com/icfs?id=ibfs_application_summary&number=SAT-LOA-20260108-00016，由多达一百万颗卫星组成。该申请的技术细节很少：https://spacenews.com/spacex-files-plans-for-million-satellite-orbital-data-center-constellation/ ——例如，我们对卫星的尺寸或部署时间表一无所知 —— 这让我觉得这更多的是为公关，而不是其他目的。马斯克还公布了一些关于卫星的细节——或者至少是一些卫星的图纸：https://spacenews.com/spacex-offers-details-on-orbital-data-center-satellites/。在这个设想中，马斯克拥有的芯片制造商 Terafab 每年将生产一太瓦的芯片。十亿台 Optimus 机器人将完成这些工作，至少在马斯克弄清楚手的使用方法之后。
这一切听起来很昂贵，不是吗？马斯克声称，最终目标是在月球上建造一个大型加速器：https://www.businessinsider.com/spacex-first-earnings-report-spcx-stock-lockup-period-expiration-2026-8?utm_source=copy-link&utm_medium=referral&utm_content=topbar#musk-knows-manufacturing-on-the-moon-sounds-super-sci-fi-and-totally-nuts。
我并不太在意这种事情，就像我并不太在意马斯克在财报电话中所说的 Starlink 将提供“世界大部分互联网”一样。这都是马斯克的一贯习惯：https://aeon.co/essays/silicon-valley-has-a-science-fiction-problem，谈论科幻未来：https://www.businessinsider.com/spacex-first-earnings-report-spcx-stock-lockup-period-expiration-2026-8?utm_source=copy-link&utm_medium=referral&utm_content=topbar#musk-knows-manufacturing-on-the-moon-sounds-super-sci-fi-and-totally-nuts，由一群20世纪中叶的作家所幻想的。当然，有很多科学家认为马斯克提议的太空数据中心是一个坏主意：https://www.nbcnews.com/science/space/spacex-data-centers-space-ai-scientist-backlash-environment-rcna588930。你甚至可以讨论太空数据中心的可行性：https://www.verysane.ai/p/should-we-put-gpus-in-space。提醒我，超级高铁曾经建成过吗：https://www.fresnobee.com/opinion/editorials/article264451076.html？跟我一起说：在出货之前，这一切都是假想软件，亲爱的。
如果去掉伪科学的大言不惭，我们剩下的是一家主要为自己发射火箭的公司，拥有相当成功的卫星互联网业务，并且基本上在运营风险高、资本密集型的裸金属业务：/ai-artificial-intelligence/848988/nvidia-chip-loans-coreweave-gpu-debt-ai-neocloud，这是我们从一堆新兴云供应商那里看到的结果，都是因为它内部 AI 的失败。这可不像六年内到火星那么梦幻：https://www.theguardian.com/technology/2016/sep/27/elon-musk-spacex-mars-colony，对吧？
SpaceX 内部人员禁售期将于8月6日到期，也就是明天。
不过，它确实为特斯拉提供了一个重要客户，该客户购买了价值2.95亿美元的特斯拉 Megapack 储能电池：https://techcrunch.com/2026/08/04/spacex-has-bought-329m-worth-of-tesla-megapacks-so-far-this-year/。SpaceX 还购买了很多 Cybertruck，马斯克史上的失败之作：https://www.bloomberg.com/news/features/2026-07-22/tesla-cybertruck-risks-becoming-the-ford-edsel-of-the-ev-era。撰写本文时，特斯拉股票自今年1月以来已下跌25%。
马斯克也许可以因为他的政治关系在计算服务上收取高额费用：/elon-musk/676800/elon-musk-doge-leaving-politics-trump — 我认为他不会无缘无故在中期选举上花费一亿美金：https://www.nytimes.com/2026/07/30/us/politics/elon-musk-midterms-republicans.html。也许那些政治关系会让马斯克建设地面数据中心变得更容易—尽管这些在整个政治光谱中似乎都非常不受欢迎：/policy/969667/humans-first-data-center-protest-hernando-county-florida-republicans。谁知道呢，也许他的计划就是，“你那儿的AI业务不错。如果出点事就太可惜了。也许你应该买我的计算资源，这样就不会有事发生。”
但也许我想得太长远了。毕竟，SpaceX的内部人员锁定期将于8月6日到期，也就是明天。如果那些内部人员卖出——正如空头预期的那样：https://www.cnbc.com/2026/08/03/spacex-short-interest-overtakes-tesla-ahead-of-earnings-lockup-expiration.html——本已下跌的股价将更加糟糕。所以提醒大家月球上的大型加速器也许是为了鼓舞士气。幸运的是纳斯达克更改了针对SpaceX的规则，所以那些损失将影响持有指数基金的所有人：/business/968257/spacex-in-your-index-fund-explained!
实际上，我对这个有一个科幻的想法：https://en.wikipedia.org/wiki/Slapstick_%28novel%29。嘿，埃隆！你为什么不对着一个旋转的甜甜圈飞来一发？你为什么不对着月亮飞来一发？
Once, I had some questions：/tech/887899/spacex-ipo-risks-ai about why SpaceX, Elon Musk’s healthiest company, acquired xAI：/tech/872619/elon-musk-merges-spacex-with-xai-and-x, his sickliest one. Now I have some questions about why we’re calling the whole thing SpaceX.
Look, what we have here, by revenue, is primarily a telecom company and a company that rents compute, according to SpaceX’s first quarterly earnings statement as a public company：/science/975335/spacex-made-more-money-as-a-neocloud. The space sector of the business didn’t break a billion dollars this quarter and contributed only a touch over 10 percent of the company’s revenue. SpaceX remains its own biggest customer. There just aren’t enough other people who want its rockets. And, I dunno, maybe if the rockets were a primary focus of the company, SpaceX wouldn’t be in danger of blasting a new crater into the Moon：https://www.theguardian.com/science/2026/aug/05/spacex-rocket-moon-crash-impact-falcon-9 with its space trash.
The lion’s share of the spending (and the hype) is nonetheless around what is graciously being referred to as “AI”
The telecom part — SpaceX calls it “connectivity” — is Starlink, the satellite internet service, which had $4.2 billion in revenue, and was the only part of SpaceX that did not have a loss from operations. And while Gwynne Shotwell, on the call, outlined plans for a phone service meant to compete with AT&T, Verizon, and T-Mobile, the lion’s share of the spending (and the hype) is nonetheless around what is graciously being referred to as “AI” to capture whatever contributions Grok may be making. Analyst Alexander Potter expects spending on the neocloud business, which leases data center capacity to AI companies, to climb to $65 billion next year, $17 billion more than he’d previously estimated, Bloomberg notes：https://www.bloomberg.com/news/articles/2026-08-04/spacex-exceeds-revenue-estimates-in-first-earnings-since-ipo.
It was hard to miss how important leasing data center space was to SpaceX. Not only did it make more money than SpaceX’s rockets, it also drove more spending — $15.8 billion on AI alone in the second quarter. (By comparison, spending on the space and connectivity sectors was a touch over a billion each.) That business makes it a competitor with neocloud companies such as CoreWeave：/ai-artificial-intelligence/822011/coreweave-debt-data-center-ai and Nebius.
Except that wasn’t originally the plan.
Musk built SpaceX’s Colossus 1 data center in Memphis for Grok, his also-ran AI that also occasionally refers to itself as MechaHitler and has a penchant for undressing women and children without their consent. But xAI had trouble running the complex：https://www.bloomberg.com/news/articles/2026-06-12/spacex-rented-out-computing-after-own-teams-had-trouble-using-it, and decided to rent it out instead. In addition to encountering latency that made it hard to train in-house models, the center employed a mix of newer and older chips that created bottlenecks. On the earnings call, Musk said that only 10 percent of the compute SpaceX builds will go to Grok：https://www.businessinsider.com/spacex-first-earnings-report-spcx-stock-lockup-period-expiration-2026-8?utm_source=copy-link&utm_medium=referral&utm_content=topbar#musk-estimates-10-of-space-xs-compute-will-feed-grok.
Revenue isn’t profit, and building data centers is expensive
SpaceX now has deals with Google, Anthropic, Reflection AI, and Cursor, an AI company Musk eventually chose to acquire. On the second quarter earnings call, SpaceX chief financial officer Bret Johnsen said those deals put the company “on a trajectory, including contribution from Cursor, to reach $100 billion of ARR, or annualized revenue run rate,” a financial measurement used to estimate how much money a company will receive in a year, based on an estimate from a shorter period of time. Musk was even more bullish：https://techcrunch.com/2026/08/04/elon-musk-repeatedly-one-upped-his-execs-on-spacexs-first-earnings-call/, saying that “the $100 billion ARR in December is not a question mark,” and that the actual ARR might be higher.
That’s very nice. But revenue isn’t profit, and building data centers is expensive.
There are some problems with being in the bare-metal business of renting out compute：/ai-artificial-intelligence/848988/nvidia-chip-loans-coreweave-gpu-debt-ai-neocloud — specifically, inevitable obsolescence, the vagaries of construction, and the fact that compute is basically a commodity, which means that companies compete on cost. The more data centers that get built, the more compute is available. With more compute available, companies can’t charge as much for their chips.
Musk has claimed that he took SpaceX public because he wanted to build data centers in space. Now, I might think that if I had trouble running a data center on the ground, I’d get better at it before I tried to do something an order of magnitude harder：/ai-artificial-intelligence/845453/space-data-centers-astronomers — something that had never been done before, even. But that is not how Musk thinks, I suppose.
This all sounds expensive, doesn’t it?
SpaceX has even proposed an orbital data center：https://fccprod.servicenowservices.com/icfs?id=ibfs_application_summary&number=SAT-LOA-20260108-00016 consisting of as many as 1 million satellites to the Federal Communications Commission. The application is light on technical details：https://spacenews.com/spacex-files-plans-for-million-satellite-orbital-data-center-constellation/ — we have no idea about satellite size or deployment schedule, for instance — which makes me think that it’s more for PR than anything else. Musk has also released a few details about the satellites — or, at least, some drawings of them：https://spacenews.com/spacex-offers-details-on-orbital-data-center-satellites/. In this vision, a Musk-owned chip producer called Terafab will produce one terawatt of chips every year. A billion Optimus robots will do the work, at least once Musk figures out hands.
This all sounds expensive, doesn’t it? The end goal, Musk claims, is to build a mass accelerator on the Moon：https://www.businessinsider.com/spacex-first-earnings-report-spcx-stock-lockup-period-expiration-2026-8?utm_source=copy-link&utm_medium=referral&utm_content=topbar#musk-knows-manufacturing-on-the-moon-sounds-super-sci-fi-and-totally-nuts.
I don’t put much stock in this kind of thing, any more than I put much stock in Musk’s claims that Starlink will deliver “a majority of the world’s internet,” as he said on the earnings call. It’s all part and parcel of Musk’s general habit：https://aeon.co/essays/silicon-valley-has-a-science-fiction-problem of talking about a sci-fi future：https://www.businessinsider.com/spacex-first-earnings-report-spcx-stock-lockup-period-expiration-2026-8?utm_source=copy-link&utm_medium=referral&utm_content=topbar#musk-knows-manufacturing-on-the-moon-sounds-super-sci-fi-and-totally-nuts as imagined by a bunch of mid-20th-century writers. Sure, there are a bunch of scientists suggesting Musk’s proposed space data centers are a bad idea：https://www.nbcnews.com/science/space/spacex-data-centers-space-ai-scientist-backlash-environment-rcna588930. You can even argue about the feasibility of space data centers if you want：https://www.verysane.ai/p/should-we-put-gpus-in-space. Remind me, did the Hyperloop ever get built：https://www.fresnobee.com/opinion/editorials/article264451076.html? Say it with me now: It’s all vaporware till it ships, babe.
If we take away the pseudoscientific big talk, we are left with a company that launches rockets mostly for itself, has a fairly successful satellite internet business, and is basically running the risky and capital-intensive bare-metal business：/ai-artificial-intelligence/848988/nvidia-chip-loans-coreweave-gpu-debt-ai-neocloud we’ve seen from a bunch of the neoclouds, as a result of its failures of its in-house AI. That’s not quite as dreamy as getting to Mars in six years：https://www.theguardian.com/technology/2016/sep/27/elon-musk-spacex-mars-colony, is it?
SpaceX’s insider lockups start expiring on August 6th, which is tomorrow
What it does do, however, is provide Tesla with an important customer, one that buys $295 million in Tesla Megapack battery storage：https://techcrunch.com/2026/08/04/spacex-has-bought-329m-worth-of-tesla-megapacks-so-far-this-year/. SpaceX has also been buying a bunch of Cybertrucks, Musk’s historic flop：https://www.bloomberg.com/news/features/2026-07-22/tesla-cybertruck-risks-becoming-the-ford-edsel-of-the-ev-era of an automobile. As of this writing, Tesla stock is down 25 percent since January of this year.
Musk can maybe charge a premium on his compute because of his political connections：/elon-musk/676800/elon-musk-doge-leaving-politics-trump — I don’t think he’s planning on splashing out $100 million on the midterms for nothing ：https://www.nytimes.com/2026/07/30/us/politics/elon-musk-midterms-republicans.html. Maybe those political connections will make it easier for Musk to build terrestrial data centers — though those seem to be wildly unpopular：/policy/969667/humans-first-data-center-protest-hernando-county-florida-republicans across the entire political spectrum. Who knows, maybe the game plan is just, That’s a nice AI business you have there. Would be a real shame if something happened to it. Perhaps you should buy my compute, so nothing happens.
But maybe I’m just thinking too long-term. After all, SpaceX’s insider lockups start expiring on August 6th, which is tomorrow. If those insiders sell — as short-sellers expect them to：https://www.cnbc.com/2026/08/03/spacex-short-interest-overtakes-tesla-ahead-of-earnings-lockup-expiration.html — the already slumping stock will be even worse off. So reminding everyone of the mass accelerator on the Moon is maybe meant to rally the troops. Good thing Nasdaq changed its rules for SpaceX, so those losses will hit everyone with an index fund：/business/968257/spacex-in-your-index-fund-explained!
Actually, I have a science fiction thought about that：https://en.wikipedia.org/wiki/Slapstick_%28novel%29. Hey Elon! Why don’t you take a flying fuck at a rolling donut? Why don’t you take a flying fuck at the Mooooooon ?
扫码直达文章详情