文 | 偏科声，作者 | 李小年
一家造火箭的公司交出了上市以来最凶的一份财报。
10 月 20 日盘后，SpaceX 发布 2026 年第三季度业绩：营收 145.75 亿美元，同比增长 161%；净利润 44 亿美元，去年同期还亏约 10.5 亿。这是公司今年 6 月上市以来的第二份季报，也是第一次单季盈利。
真正的炸点在收入结构里。
SpaceX 的 AI 业务“超级智能”（Super Intelligence，SI）单季度收入 75.6 亿美元，同比增长 830%，占总营收 52%——首次超过星链，成为这家公司最大的收入来源。星链做了这么多年，这个季度收入 52.5 亿美元，增长 69%，但在 AI 面前已经不在一个量级。
财报发布第二天，SpaceX 股价大涨，市值站上 2 万亿美元，超过特斯拉、Neuralink 和 The Boring Company 三家估值之和。
一家以火箭和卫星为人熟知的公司，用四个季度时间，把自己变成了一家算力公司。
这件事值得拆开看：它的算力是怎么建起来的，钱是怎么赚的？
01 建超算这件事，别人按年算，马斯克按天来
孟菲斯南郊有一座废弃的伊莱克斯冰箱工厂，7 万多平方米的空厂房，2023 年底被一家地产投资公司买下后一直闲置——里面没有精密空调，没有冗余供电，连光纤主干道都没有。
2024 年 4 月，马斯克的团队进场，把它改造成了 Colossus 超算集群。后来的 Colossus II 也建在孟菲斯，部分场地扩到了紧邻的密西西比州 Southaven——三处场地连成一片，构成了目前全球规模最大的 AI 训练设施群之一。
这个集群原本属于 xAI，2026 年 2 月并入 SpaceX，跟着一起上了市，成了公司财报里最重的一块收入。
而这个集群跑出的速度，在行业里找不到第二个案例。
Colossus 一期，10 万张英伟达 H100，122 天上线。从机架运抵现场到开始训练，只用了 19 天。
Colossus II 更快：第一块集群，约 11 万张 GB200 加 210 兆瓦电力，91 天；第二块，11 万张 GB300 加 220 兆瓦，64 天。
这个速度是什么水平？
黄仁勋在播客节目里给过对照：“10 万张 GPU 的超算，行业常规是三年规划，再加一年集成调试，整整四年。他们只用了 19 天。”
这位见过全球所有大算力项目的芯片公司掌门人，用的形容词是“超人类的壮举”，补了一句：“据我所知，世界上只有一个人能做到。”
放到吉瓦级园区的尺度上，差距同样清晰。
研究机构 Epoch AI 用卫星图像和许可文件追踪了美国主要 AI 数据中心：从动工到达到 1GW 电力容量，xAI 的 Colossus 2 用了约 12 个月；OpenAI 的星门 Abilene 项目约 2.1 年，微软 Fairwater 约 2.1 年，亚马逊 Ridgeland 约 1.9 年。
头部云厂商基本要花 xAI 两倍的时间。
支撑速度的是一套火箭工程方法论：马斯克把全球发射经验最丰富的火箭工程团队拉来建数据中心，多个工序并行推进，冷却系统调试的同时机架在通电，他本人每周拿出约四天、日夜督战集群建设。
他在电话会上说，SpaceX 的部署速度“可能超过其他所有玩家的总和”。
这个打法已经开始改写行业的游戏规则——跟进最快的，是黄仁勋。
就在财报发布次日，10 月 21 日，黄仁勋在 GTC 柏林大会上发布了 DSX POD 参考设计：把 7 个 Vera Rubin 机架连同线缆、液冷系统在工厂里预装进集装箱，运到现场即插即用，配套电力就绪后，部署周期从约 6 个月压缩到最快 14 天。
卖芯片的公司亲自下场做“标准化提速”，等于承认了一件事——部署速度，已经取代芯片本身，成为 AI 基础设施竞争的核心变量。
马斯克的速度直接转化成了规模。截至 9 月 30 日，SpaceX 名义算力容量达到 1.7GW，高于 6 月底的 1.4GW 和去年同期的 0.6GW。
当前推进节奏是每月上线约 20 万块 GB300 GPU。公司计划年底做到 2.5 至 3GW，2027 年新增 10GW。
02 建好租出去，把对手变客户
集群建起来，怎么用？马斯克给出的答案颇有戏剧性：租给 Grok 的竞争对手。
2026 年 5 月，Anthropic 与 SpaceX 签下算力协议，租用 Colossus 的数据中心容量，合同总额最高达 845 亿美元，按每月 12.5 亿美元支付，期限到 2029 年 5 月。
6 月，Google 跟进。根据 SpaceX 递交给美国证监会的文件，Google 以每月 9.2 亿美元的价格租下约 11 万张 GPU 的容量，合同期从 2026 年 10 月到 2029 年 6 月。
路透社报道，客户名单里还有 ReflectionAI 等 AI 公司。
换句话说，Claude 和 Gemini 的训练任务，正在给 Grok 的母公司交租金。
别人建集群是成本中心，马斯克建集群是印钞机，连竞争对手都是客户。
更“会赚钱”的是定价方式。SpaceX 不签长约，倾向按接近现货市场的价格签短约，内部使用与对外租赁大致按 50:50 动态分配。
马斯克在电话会上说过一句话：“如果答应所有求租请求，算力可以租出去好几倍。”
这句话反过来读更有意思：求租的需求远大于产能，算力根本不够卖。所以 SpaceX 不签长约——产能攥在自己手里，客户排队、续约涨价，主动权全在他这边。
而且短约的续约价格正在走高，CFO Bret Johnsen 确认，当前定价已略高于上一季度给出的每瓦 30 至 50 美元区间的上沿。
算力这门生意成不成立，最终看回本周期。
Johnsen 给出的数字是：新部署算力的回本周期不足一年。马斯克补了一句，2027 年有可能缩短到 6 个月。
而传统云厂商的 AI 基础设施，通常按五到六年的折旧周期摊销——换句话说，别人按五年规划收回成本，他一年就翻倍了。
按他的说法，算力业务“启动后几乎可以自我供血”，资本开支的性质更接近销货成本，而不是传统意义上的长周期基建。
截至 9 月 30 日，SpaceX 在手订单总额 630 亿美元，其中约 420 亿美元预计将在未来一年内确认为收入。
SI 板块当季经营利润 4.31 亿美元，上一季度还是亏损 12.57 亿美元——这个数字还包含 Cursor 并购带来的约 12 亿美元股权激励费用，剔除后实际约 16 亿美元。
对，Cursor。今年 8 月，SpaceX 以隐含股权估值 600 亿美元完成了对 AI 编程工具 Cursor 的收购，GAAP 口径购买对价 425 亿美元。
Cursor 本季度并表贡献收入 6.08 亿美元，即便把它完全剔除，SI 的解决方案与基础设施收入环比仍增长 192%，靠的是新云服务合同的激活。
算力出租、模型、应用、编程工具，四层业务被马斯克装进了同一个上市公司。
而他的下一台印钞机已经排上了。
10 月，SpaceX 宣布与英伟达合作开发 Starmind AI1 卫星——把整套 AI 计算机架装进卫星，发射到近地轨道，用太阳能供电、靠真空散热，再把成千上万颗卫星用星链激光链路连成一张太空算力网。
首批发射提前到了 2027 年四季度，FCC 文件显示 SpaceX 申请的发射授权多达 100 万颗。JPMorgan 测算，到 2031 年底轨道算力约 75GW。
地面数据中心受土地、供电、水资源挤压，而太空的太阳能和真空散热没有这个问题。
这个活，全世界只有一家有星舰的公司能干。
高速扩张的另一面是钱。
第三季度资本开支 195.9 亿美元，其中 164.4 亿投向算力基础设施，单季投入超过当季全部营收，第四季度指引还要涨到 350 亿至 400 亿美元。
标普全球市场财智的市场一致预期显示，SpaceX 总债务将从 2025 年的 507 亿美元升至 2029 年的 3500 亿美元；摩根士丹利的 Adam Jonas 测算，2027 至 2029 年间公司需要新增约 2300 亿美元债务融资。
算力这门生意本是重资产、长周期、慢慢回本的基建账，硬是被马斯克做成短约、高周转、越卖越贵的现金流生意。
至于这场盛宴能持续多久，债务的不确定性还挂在半空，无人知晓。但至少这个季度，市场已经用 2 万亿美元的市值投了票。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Man, you've got to be kidding me.
A rocket-making company handed over one of the worst financial accounts since its listing.
After 20 October, SpaceX publish performed in the third quarter of 2026: received $14,57.5 billion, an increase of 161 per cent over the same period; net profit of $4.4 billion, and a shortfall of approximately $105 million over the same period last year. This is the second quarterly edition of the company since it came on the market in June this year, and the first single-season profit.
The real blast is in the revenue structure.
SpaceX's AI Business SuperIntelligence (Super Intelligence, SI) single-quarter revenue of $7.56 billion, an increase of 830 per cent over the same period, representing 52 per cent of total sales — for the first time exceeding the chain of stars — became the largest source of revenue for the company. Over the years, the chain has earned $5.25 billion this quarter, an increase of 69 per cent, but it's not on a scale in front of AI.
The next day, the financial paper publish, SpaceX stock prices rose sharply, with $2 trillion at the market value station, exceeding the sum of Tesla, Neuralink and The Boring Company valuations.
A company known as a rocket and satellite turned itself into a computer company over four quarters.
It's worth looking at: how does it work and how does it make money?
We're gonna build this thing on a year-by-year basis.
In the southern suburbs of Memphis there was an abandoned Ilex refrigerator, an empty plant of over 70,000 square metres, which had been idle since the end of 2023 when it was bought by a real estate investment company — there was no precision air conditioning, no redundant power supply, no fibre-optic mainway.
In April 2024, the team of Mask came in and turned it into a Colossus super-group. Later, Colossus II was also built in Memphis, and part of the site spread to Southeven, Mississippi, the immediate vicinity. The three sites are linked together and constitute one of the largest AI training facilities in the world.
This cluster, originally belonging to XAI, was incorporated into SpaceX in February 2026, and then entered the market, making it the largest piece of revenue in the company ' s financial statements.
And the speed at which this cluster is running is such that no second case can be found in the industry.
Colossus Issue, 100,000 Yvette H100, 122 Skyline. It took only 19 days to get to the scene and start training.
Colossus II Faster: first cluster, approximately 110,000 GB200 plus 210 MW electricity, 91 days; second block, 11,000 GB300 plus 220 MW, 64 days.
What level is this speed?
Wong In-hoon contrasts with the podcast program: "A hundred thousand GPUs are over-calculated, the industry routine is a three-year plan, plus one year of integration calibration for four years. They only took 19 days."
This man, who has met with the head of the chip company of all the big calculus projects in the world, uses the adjective “the superhuman feat”, adding: “As far as I know, only one person in the world can do it.”
The gap is also clear on the scale of the Giwa campus.
The research institute Epoch AI tracked the main AI data centre in the United States with satellite imagery and licence documents: from start-up to reach 1 GW power capacity, colossus 2 of x AI took approximately 12 months; Open AI's Stargate Ablene project was about 2.1 years, Microsoft Fairwater was about 2.1 years and Amazon Ridgeland was about 1.9 years.
It's almost twice as long for head cloud manufacturers.
The speed is supported by a rocket engineering methodology: Mask pulls the most experienced rocket engineering team from around the world to build a data centre, multiple processes go hand in hand, cooling systems are debugged while the plane is powered, and he himself takes out about four days a week of day and night surveillance cluster construction.
He said at the conference that SpaceX deployment was “perhaps faster than all other players combined”.
The game has begun to change the rules of the industry — the fastest follow-up is Wong In-hoon.
On the next day of the newspaper publish, 21 October, Hoang In-hoon had publish at the GTC Berlin Congress, and DSX POD reference design: pre-loading seven Vera Rubin racks with cable, liquid cooling systems in the factory, immediately on-site, plugged in, and equipped with electricity, the deployment cycle was reduced from about six months to the nearest 14 days.
The company selling the chip went down to do a “standardized speed boost” in its own right, recognizing one thing — the speed of deployment, which has replaced the chip itself as the core variable of the AI infrastructure competition.
Mask's speed is going straight to scale. As of 30 September, SpaceX had a nominal capacity of 1.7 GW, above 1.4 GW at the end of June and 0.6 GW at the same time last year.
The current propulsion rhythm is about 200,000 GB300 GPUs per month. The company plans to end the year with 2.5 to 3 GW, adding 10 GW in 2027.
Zero-two, build a lease and turn your opponent into a customer.
How do you use clustering? Musk's answer is dramatic: it's for Göok's competitors.
In May 2026, Anthropic signed a computing agreement with SpaceX to rent the data centre capacity of Colossus, with a total contract value of $84.5 billion, paid at $1.25 billion per month for the period up to May 2029.
June, Google follow up. According to SpaceX documents submitted to the US SEC, Google leased the capacity of a contract of 110,000 GPUs at $920 million per month for the period from October 2026 to June 2029.
Reuters reports that the customer list also includes the Reflection AI and other AI companies.
In other words, Claude and Gemini ' s training missions are paying rent to the Grok parent company.
Others build clusters as cost centres, and Mask build clusters as money-printers and even competitors as customers.
More “profits” are priced. SpaceX does not sign long contracts, prefers short contracts at prices close to the spot market, and internal usage and external leases are distributed around 50:50 dynamics.
Mask said at the conference: “If all rent requests are granted, they can be rented several times over.”
The sentence, in turn, is more interesting: rent demand is much greater than capacity and is not sold at all. So SpaceX doesn't sign long contracts -- it's in his hands, it's in line, it's up to him.
Moreover, the price of the short contract renewal is rising, and CFO Bret Johnsen confirmed that the current pricing is slightly higher than the upwards of the 30 to 50 dollars per watt given in the previous quarter.
It's a business that doesn't work, and it finally comes back to the cycle.
The figure given by Johnsen is that the new deployment calculator returns less than a year. Mask added that 2027 could be reduced to six months.
For traditional cloud manufacturers, the AI infrastructure is usually amortized on a five- to six-year depreciation cycle - in other words, he doubles in a year when someone else recovers costs according to a five-year plan.
According to him, capital expenditure is more similar in nature to the cost of selling, rather than the traditional long-term infrastructure, as the power business is “almost self-financing after start-up”.
As at 30 September, SpaceX was in possession of orders totalling $63 billion, of which approximately $42 billion was expected to be recognized as income in the coming year.
The SI Board operated profits of $431 million in the current season and was still in deficit in the previous quarter of $1,257 million — a figure that also included some $1.2 billion in equity incentive costs arising from Cursor's merger and acquisitions, which actually amounted to $1.6 billion.
Yeah, Cursor. In August of this year, SpaceX completed the acquisition of the AI programming tool Cursor with an implicit equity valuation of $60 billion, and the GAAP calibrated purchase against $42.5 billion.
Cursor contributed $608 million in revenue this quarter, even though it was completely eliminated, the SI solution and infrastructure revenue ring increased by 192 per cent, thanks to the activation of the New Clouds Service Contract.
Computation rents, models, applications, programming tools and four layers of operations were installed by Mask into the same listed company.
And his next machine is already in line.
In October, SpaceX announced the development of Starmind AI1 satellites in partnership with Wedda -- a complete set of AI computers installed in satellites, launched into near-Earth orbit, powered by solar power, spreads the heat in a vacuum, and connected thousands of satellites with a satellite chain laser link into a space computing network.
The first launch was advanced to the fourth quarter of 2027, and the FCC file shows that SpaceX applied for up to 1 million launch authorizations. JP Morgan calculates that by the end of 2031 the orbit calculates about 75 GW.
Ground-based data centres are squeezed by land, electricity and water resources, while solar and vacuum heat in space is not a problem.
There's only one company in the world with a starship.
The other side of the rapid expansion is money.
Capital expenditure in the third quarter amounted to $19.99 billion, of which $16.44 billion went to computing infrastructure, with single-season investments exceeding the full revenue received in the current season, and the fourth quarter guidelines rising to $35 billion to $40 billion.
According to consistent global market expectations, SpaceX's total debt will rise from $50.7 billion in 2025 to $350 billion in 2029; Adam Jonas of Morgan Stanley estimates that companies will need additional debt financing of about $230 billion between 2027 and 2029.
Calculator is an asset-intensive, long-cycled, slow-backed capital-based business, made by Mask into a short-term, high-swipe, more expensive cash-flow business.
As to how long this feast will last, the uncertainty of debt is still on the table, and nobody knows. But at least this quarter, the market has voted at $2 trillion in market value.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.