Google × SpaceX 太空数据中心：当 AI 把地球电网吃光了，硅谷决定上天
Google × SpaceX Orbital Data Centers: When AI Ate the Power Grid, Silicon Valley Decided to Go to Space
> 📌 TL;DR
> Google 正在和 SpaceX 谈判，计划将 AI 数据中心送上太空轨道。代号"Project Suncatcher"的项目将于 2027 年初发射两颗搭载 TPU 的原型卫星。长期愿景是部署 81 颗卫星组成的计算集群，利用太空中近乎无限的太阳能和天然散热来运行 AI 模型。但现实很骨感：目前发射成本是经济可行性门槛的 10 倍以上，要到 2035 年才可能真正划算。
AI 吃光了地球的电，硅谷决定上天
2026 年 5 月 12 日，《华尔街日报》率先报道：Google 正在与 SpaceX 进行深入谈判，探索将 AI 数据中心部署到近地轨道。
这不是科幻小说的脑洞——这是两家市值合计超过 3 万亿美元的公司，在认真讨论的商业计划。
为什么？因为 AI 正在把地球的电网逼到极限。
我们之前分析过 AI 的能源危机——数据中心的电力需求正以指数级增长，电网扩容的速度远远跟不上。至少 11 个美国州已经提出了限制数据中心建设的法案。土地、水资源、电力审批，每一项都在制约 AI 算力的扩张。
Google 的答案是：既然地球上的资源快不够了，那就去太空找。
Project Suncatcher：从论文到原型
Google 的太空数据中心计划代号 Project Suncatcher（"捕日者"），最早在 2025 年 11 月公布研究论文，2026 年 5 月正式进入工程落地阶段。
核心设计
| 参数 | 规格 |
|------|------|
| 计算芯片 | Google Trillium TPU（已通过粒子加速器辐射测试）|
| 卫星集群 | 81 颗卫星，间距 100-200 米，集群半径 1 公里 |
| 轨道类型 | 太阳同步轨道，高度约 640 公里（400 英里）|
| 星间链路 | 自由空间光通信（FSOL），带宽 10+ Tbps |
| 供电 | 太阳能板——轨道上的太阳能效率是地面的 8 倍，且几乎不间断 |
| 散热 | 太空辐射散热器——不需要水冷、不需要空调 |
2027：两颗原型上天
Google 已经与卫星公司 Planet Labs 达成合作，计划在 2027 年初发射两颗原型卫星。这两颗卫星的任务很明确：
1. 验证 TPU 在太空环境下的性能——辐射测试虽然通过了，但真实轨道环境更复杂
2. 测试高带宽星间光通信链路——分布式 AI 计算对网络延迟极度敏感
3. 验证散热方案——如何把 TPU 产生的热量高效排放到太空
Google CEO Sundar Pichai 的原话："毫无疑问，十年后，太空数据中心会成为一种更常规的建设方式。"
经济账：理想很丰满，现实很骨感
太空数据中心的吸引力显而易见——无限太阳能、免费散热、没有土地审批。但成本呢？
发射成本是最大障碍
| 指标 | 数据（2026 年 5 月）|
|------|------|
| 当前发射成本 | $1,500-$2,900/公斤 |
| SpaceX 标准拼车价 | $7,000/公斤 |
| Google 的盈亏平衡点 | $200/公斤 |
| Starship 终极目标 | < $100/公斤 |
| 预计达到盈亏平衡 | 2035 年 |
换句话说：目前轨道数据中心的成本大约是地面的 3 倍。要等 Starship 火箭实现每年 180 次发射的稳定节奏，把成本压到 $200/公斤以下，太空数据中心才在经济上站得住脚。
技术挑战同样不小
1. 高带宽内存（HBM）的辐射敏感性：Google 的测试发现 HBM 子系统对累积辐射最敏感，在推理任务中错误率"尚可接受"，但对训练任务的影响"需要进一步研究"
2. 凯斯勒综合症风险：数百万颗卫星在轨运行可能引发碎片碰撞的连锁反应，FCC 已经开始将此作为政策问题处理
3. 维护和升级：地面数据中心坏了换个硬件就行，太空里呢？
SpaceX 的宏大野心：$1.75 万亿 IPO 的关键叙事
这笔交易对 SpaceX 来说同样意义非凡。
SpaceX 已于 2026 年 4 月 1 日秘密提交了 IPO 申请，目标估值 $1.75 万亿美元，计划 6 月上市。如果成功，将成为史上最大 IPO。
而"轨道 AI 计算"正是支撑这个天价估值的核心叙事之一。
SpaceX 已经向 FCC 提交了部署 100 万颗太阳能计算卫星 的文件。马斯克提出的愿景是：每年从轨道上部署 100 吉瓦 的 AI 计算能力。
SpaceX 的 S-1 招股书引用了一个 $28.5 万亿 的总可寻址市场（TAM），其中 AI 企业服务占大头。
但 S-1 也坦率地警告投资者："我们开发轨道 AI 计算的计划处于早期阶段，涉及重大技术复杂性和未经验证的技术，可能永远无法实现商业化。"
Starlink 的订阅引擎在 2025 年创造了 $44.2 亿的运营利润（翻倍增长），但 AI 部门在 2025 年亏损了 $64 亿，消耗了公司 61% 的资本支出。
竞争格局：不只是 Google
Google 并非唯一对太空算力感兴趣的科技巨头。
Anthropic（Claude 的母公司）上周与 SpaceX 达成协议，使用 SpaceX 收购的 xAI 在田纳西州孟菲斯的数据中心的算力资源，双方还在探索未来在轨道计算方面的合作。
这意味着 SpaceX 正在同时与 AI 行业的两大玩家进行太空计算的谈判——Google 和 Anthropic。
这对普通人意味着什么？
短期（2026-2030）：几乎没有直接影响。太空数据中心还在原型验证阶段，你的 ChatGPT 和 Claude 还是跑在地面机房里。
中期（2030-2035）：如果 Starship 成功把发射成本压到 $200/公斤以下，我们可能会看到第一批商业化的轨道计算集群上线，主要用于推理（inference）工作负载。
长期（2035+）：如果一切顺利，太空可能真的成为 AI 算力的重要补充。想象一下：你问 AI 一个问题，回答你的计算资源可能同时分布在弗吉尼亚的数据中心和近地轨道的卫星集群上。
但也有可能：这个项目像很多太空商业计划一样，在工程现实面前逐渐缩小规模，最终变成一个有趣但小众的应用。
清醒的判断
太空数据中心是 AI 行业最大胆的赌注之一。它解决了一个真实的问题（AI 算力对地球资源的巨大需求），但面临的挑战也同样真实（成本、辐射、维护、碎片风险）。
看多的理由：AI 算力需求的增长速度确实可能超过地球基础设施的扩容速度，太空是为数不多的能提供"无限"资源的地方。
看空的理由：SpaceX 自己都在 S-1 里承认这可能永远无法商业化。$200/公斤的发射成本目标需要 Starship 的飞行节奏达到一个前所未有的水平。
我的判断：2027 年的原型测试是关键拐点。 如果 TPU 在真实轨道环境中表现符合预期，星间链路带宽达到设计指标，这个赛道就会从"疯狂的想法"变成"激进但可行的方案"。反之，它可能会安静地回到论文阶段。
无论最终结果如何，这件事本身就标志着 AI 行业已经进入了一个新阶段——算力的瓶颈不再是芯片和算法，而是物理世界的资源约束。 当硅谷开始认真讨论"上天"的时候，你就知道地球上的 AI 军备竞赛有多疯狂了。
> ✨ 当 AI 的野心大到地球已经装不下，人类历史上最大胆的基础设施赌注就诞生了——不是修更多电厂，而是把计算机送上太空。
---
> 🔄 2026-05-18 更新
>
> SpaceX S-1 已公开：SpaceX 于 5 月 20 日正式向 SEC 提交并公开了 S-1 招股书。关键财务数据首次曝光：2025 年营收 $186.7 亿，净亏损 $49 亿；2026 年 Q1 营收 $46.9 亿，亏损 $42.7 亿。Musk 持有约 42% 股权和 85% 投票权（双层股权结构）。公司承诺 30% 散户配售比例。
>
> IPO 时间线更新：多方报道指向 6 月 12 日为潜在上市日期，将在纳斯达克以"SPCX"代码交易。目标估值区间扩大至 $1.75 万亿至 $2 万亿，计划融资 $750 亿（将超过沙特阿美的 $294 亿纪录）。承销团包括美银、花旗、高盛、摩根大通和摩根士丹利。SpaceX 承诺 30% 的散户配售比例（标准大型 IPO 通常仅 10%）。
>
> 最后更新：2026-05-18
> 🔄 2026-05-25 更新
>
> SpaceX S-1 已正式公开（5 月 20 日）：2025 年全年营收 $186.7 亿，净亏损 $49 亿。Musk 持 42% 股权 / 85% 投票权。路演 6 月 8 日启动，预计 6 月 12 日上市（纳斯达克 SPCX），估值区间 $1.75 万亿至 $2 万亿。Starship 预计 2026 年下半年开始向轨道运送有效载荷，轨道 AI 计算卫星最早 2028 年部署。
>
> 最后更新：2026-05-25
> 📌 TL;DR
> Google is in active talks with SpaceX to build AI data centers in orbit. Their "Project Suncatcher" will launch two prototype satellites with TPUs in early 2027. The long-term vision: 81-satellite compute clusters powered by unlimited solar energy with natural space cooling. The catch? Current launch costs are 10x above the economic viability threshold, and break-even isn't expected until 2035.
AI Ate the Power Grid, So Silicon Valley Is Going to Space
On May 12, 2026, The Wall Street Journal broke the story: Google is in deep discussions with SpaceX to deploy AI data centers into low Earth orbit.
This isn't science fiction. Two companies worth a combined $3+ trillion are having serious conversations about a real commercial plan.
Why? Because AI is pushing Earth's power grid to its limits.
We previously analyzed AI's energy crisis — data center power demand is growing exponentially, and grid expansion can't keep pace. At least 11 U.S. states have proposed legislation restricting data center construction. Land, water, power permits — every bottleneck is constraining AI compute expansion.
Google's answer: If Earth is running out of resources, go find them in space.
Project Suncatcher: From Paper to Prototype
Google's orbital data center initiative, codenamed Project Suncatcher, was first published as a research paper in November 2025 and entered the engineering phase in May 2026.
Core Design
| Parameter | Specification |
|-----------|--------------|
| Compute Chips | Google Trillium TPUs (passed radiation testing in particle accelerator) |
| Satellite Cluster | 81 satellites, 100-200m apart, 1km cluster radius |
| Orbit Type | Sun-synchronous orbit, ~640 km (400 miles) altitude |
| Inter-Satellite Links | Free-space optical communication (FSOL), 10+ Tbps bandwidth |
| Power | Solar panels — 8x more efficient in orbit, near-continuous generation |
| Cooling | Space radiators — no water cooling, no air conditioning needed |
2027: Two Prototypes Launch
Google has partnered with satellite company Planet Labs to launch two prototype satellites in early 2027. Their mission is clear:
1. Validate TPU performance in the real space environment — lab radiation tests passed, but actual orbital conditions are more complex
2. Test high-bandwidth optical inter-satellite links — distributed AI compute is extremely latency-sensitive
3. Verify thermal management — efficiently dissipating TPU heat into space
Google CEO Sundar Pichai: "There's no doubt to me that, a decade or so away, we'll be viewing it as a more normal way to build data centers."
The Economics: Beautiful Vision, Brutal Math
The appeal of space data centers is obvious — unlimited solar energy, free cooling, no land permits. But what about costs?
Launch Costs Are the Biggest Hurdle
| Metric | Data (May 2026) |
|--------|-----------------|
| Current launch cost | $1,500–$2,900/kg |
| SpaceX standard rideshare | $7,000/kg |
| Google's break-even point | $200/kg |
| Starship ultimate target | < $100/kg |
| Projected break-even year | 2035 |
In other words: orbital data centers currently cost roughly 3x their terrestrial equivalents. Only when SpaceX's Starship achieves a steady cadence of 180+ flights per year, driving costs below $200/kg, will space data centers become economically viable.
Technical Challenges Are Equally Significant
1. HBM radiation sensitivity: Google's tests found High Bandwidth Memory subsystems are most sensitive to cumulative radiation. Error rates are "likely acceptable for inference" but the impact on training "requires further study"
2. Kessler Syndrome risk: Millions of satellites in orbit could trigger cascading debris collisions — the FCC is beginning to treat this as a policy issue
3. Maintenance and upgrades: Swap a broken component in a ground data center? Easy. In orbit? Not so much
SpaceX's Grand Ambition: A Key Narrative for the $1.75T IPO
This deal is equally significant for SpaceX.
SpaceX confidentially filed for an IPO on April 1, 2026, targeting a $1.75 trillion valuation with a June listing. If successful, it would be the largest IPO in history.
"Orbital AI compute" is a central pillar of the narrative supporting that valuation.
SpaceX has filed FCC paperwork for deploying 1 million solar-powered compute satellites. Musk's stated vision: deploying 100 gigawatts of AI computing capacity from orbit per year.
SpaceX's S-1 cites a $28.5 trillion total addressable market, heavily weighted toward enterprise AI.
But the S-1 also candidly warns investors: "Our initiatives to develop orbital AI compute... are in early stages, involve significant technical complexity and unproven technologies, and may not achieve commercial viability."
Starlink's subscription engine generated $4.42 billion in operating income in 2025 (doubling year-over-year), but the AI division lost $6.4 billion and consumed 61% of consolidated capex.
Competitive Landscape: It's Not Just Google
Google isn't the only tech giant interested in space compute.
Anthropic (the company behind Claude) reached an agreement with SpaceX last week to use computing resources from xAI's data center in Memphis, Tennessee — the company SpaceX acquired in February. The two are also exploring future orbital computing collaboration.
This means SpaceX is simultaneously negotiating space compute deals with two of the AI industry's biggest players — Google and Anthropic.
What This Means for Regular People
Short term (2026–2030): Almost no direct impact. Space data centers are still in prototype testing. Your ChatGPT and Claude sessions still run in ground-based server farms.
Medium term (2030–2035): If Starship successfully drives launch costs below $200/kg, we could see the first commercial orbital compute clusters come online, primarily for inference workloads.
Long term (2035+): If everything works out, space could become a meaningful supplement to AI compute. Imagine asking an AI a question where the computation is simultaneously distributed across a Virginia data center and a satellite cluster in low Earth orbit.
But it's equally possible this project quietly scales back in the face of engineering reality, becoming an interesting but niche application.
A Sober Assessment
Space data centers represent one of the boldest bets in the AI industry. They address a real problem (AI's massive demand on Earth's resources), but the challenges are equally real (cost, radiation, maintenance, debris risk).
The bull case: AI compute demand growth may genuinely outpace Earth's infrastructure expansion, and space is one of the few places offering "unlimited" resources.
The bear case: SpaceX itself admits in its S-1 this may never be commercially viable. The $200/kg launch cost target requires Starship to achieve an unprecedented flight cadence.
My assessment: The 2027 prototype test is the critical inflection point. If TPUs perform as expected in real orbital conditions and inter-satellite link bandwidth hits design targets, this track moves from "crazy idea" to "aggressive but feasible." Otherwise, it may quietly return to the research paper stage.
Regardless of the outcome, this development signals that the AI industry has entered a new phase — the bottleneck is no longer chips and algorithms, but physical-world resource constraints. When Silicon Valley starts seriously discussing "going to space," you know just how intense the AI arms race on Earth has become.
> ✨ When AI's ambition outgrows Earth, humanity's boldest infrastructure bet is born — not building more power plants, but sending computers to space.
---
> 🔄 2026-05-18 Update
>
> SpaceX S-1 Filing Imminent: After SpaceX confidentially filed for its IPO on April 1, SEC rules require the registration statement to go public at least 15 days before the roadshow. With SpaceX targeting the week of June 8 for roadshow kickoff, the public S-1 is expected this week (May 18-22), providing the first detailed look at SpaceX's audited financials and offering terms.
>
> IPO Timeline Update: Multiple reports point to June 12 as the potential listing date on Nasdaq under ticker "SPCX." The target valuation range has expanded to $1.75-$2 trillion, aiming to raise $75 billion (more than doubling Saudi Aramco's $29.4B record). Underwriters include Bank of America, Citigroup, Goldman Sachs, JPMorgan, and Morgan Stanley. SpaceX has committed to a 30% retail allocation (3× the standard ~10% for mega-IPOs).
>
> Last updated: 2026-05-18
> 🔄 2026-05-25 Update
>
> SpaceX S-1 now public (May 20): 2025 full-year revenue of $18.67B, net loss of $4.9B. Musk holds 42% equity / 85% voting power. Roadshow starts June 8, expected listing June 12 on Nasdaq (SPCX), valuation range $1.75T-$2T. Starship payload delivery to orbit expected H2 2026, orbital AI compute satellites as early as 2028.
>
> Last updated: 2026-05-25