這家公司在做什麼,以及IPO的三層估值邏輯
用星鏈現金流支撐公司運轉,用星舰量產打開估值天花板,用AI重新定義自己——三層故事同時定價。
"I think the future of humanity is going to bifurcate in two directions: Either it's going to become multiplanetary, or it's going to remain confined to one planet and eventually there's going to be an extinction event. I'd like to die on Mars — just not on impact."
「我認為人類的未來將走向兩條截然不同的路:要麼成為多行星物種,要麼永遠被困在一顆星球上、最終面臨某種滅絕事件。我想死在火星上——但不是撞擊著陸的那種死法。」
Elon Musk · SpaceX Founder & CEO
公司介紹
1. SpaceX(Space Exploration Technologies Corp.)成立於2002年,是全球唯一能夠以可復用火箭大規模商業發射的私人公司,以猎鷹9號、猎鷹重型和正在試飛中的星舰為主力火箭,2025年完成170次軌道發射,包攬全球80%以上的入軌總質量,發射成本(約2,700美元/公斤)僅為傳統火箭的1/7。
2. 公司目前有三大業務:星鏈(Starlink,衛星寬帶訂閱服務,2026年Q1訂閱用戶突破1,030萬,覆蓋164個國家),火箭發射(猎鷹系列商業和政府發射),以及AI板塊(2026年2月合併後的xAI和X/推特)。三大業務2025年合計收入186.74億美元,年複合增速(2023-2025年)約34%。
3. SpaceX於2026年5月20日向美國SEC提交S-1招股說明書,計畫在納斯達克以SPCX為代碼上市,目標估值1.75-2兆美元。如果順利完成,這將是史上規模最大的IPO,預計募資超過300億美元,超越沙特阿美2019年約290億美元的歷史紀錄。
IPO的三層估值邏輯
要理解SpaceX的IPO,必須先理解一件事:這不是一次普通的科技公司上市,而是一次「三層故事同時定價」的複雜估值練習。
第一層:星鏈的真實現金流
這是SpaceX今天的收入引擎,也是整個公司估值中最「看得見摸得著」的部分。2025年星鏈收入113.87億美元(同比增長50%),EBITDA(稅前折舊攤銷前利潤,粗略代表現金創造能力)71.68億美元,利潤率接近63%。在全球164個國家擁有1,030萬寬帶訂閱用戶、740萬月活DTC(衛星直連手機)設備。這是一個已經跑通的訂閱商業模式,可以用30-40倍本銷比(PS倍數)去參照估值,對應大約4,000-6,000億美元的市值。
第二層:星舰的成本革命承諾
這是SpaceX未來的業務放大器。猎鷹9號把入軌成本從傳統火箭的18,500美元/公斤降至2,700美元/公斤,而星舰的目標是進一步降至100美元/公斤以下(降幅約99%)。如果這個目標能實現,星鏈的部署成本將急劇下降,每顆V3衛星的入軌成本可能低至數萬美元,使星鏈的護城河更深、利潤率更高。星舰也是SpaceX承接NASA阿爾忒弥斯登月任務的核心載具。2026年截至5月,星舰已完成12次試飛,V3版本已首飛,公司計畫2026年下半年正式開展軌道商業載荷發射。
第三層:太空AI基礎設施的宏大想象
這是估值上限最難計算、也最有爭議的部分。SpaceX的招股書顯示,公司估算的潛在市場規模(TAM)高達28.5兆美元,其中AI板塊佔26.5兆(93%)——包括AI基礎設施2.4兆、AI消費者訂閱7,600億、AI廣告6,000億、AI企業應用22.7兆。Anthropic已簽署450億美元36個月的算力租用合同,2028年計畫開始部署軌道AI計算衛星(利用太空真空散熱和無限太陽能解決地面數據中心的兩大瓶頸)。如果這個願景實現,SpaceX不只是「太空公司」,而是「全球成本最低的AI算力平台」。
這三層邏輯疊加後,1.75-2兆美元的估值就有了它的敘事結構。但問題在於:三層故事的確定性差距極大——第一層是已驗證的,第二層是有技術基礎但時間線不確定的,第三層幾乎完全是基於未來十年的假設。
"I think the future of humanity is going to bifurcate in two directions: Either it's going to become multiplanetary, or it's going to remain confined to one planet and eventually there's going to be an extinction event. I'd like to die on Mars — just not on impact."
「我認為人類的未來將走向兩條截然不同的路:要麼成為多行星物種,要麼永遠被困在一顆星球上、最終面臨某種滅絕事件。我想死在火星上——但不是撞擊著陸的那種死法。」
Elon Musk · SpaceX Founder & CEO
Company Overview
1. SpaceX (Space Exploration Technologies Corp.) was founded in 2002 and is the world's only private company operating commercial launches at scale with reusable rockets. Its primary vehicles — Falcon 9, Falcon Heavy, and the Starship currently in test flights — completed 170 orbital launches in 2025, capturing over 80% of global payload mass to orbit at a cost of roughly $2,700/kg, about one-seventh of traditional rockets.
2. The company operates three major businesses: Starlink (satellite broadband subscriptions, 10.3M subscribers in 164 countries as of Q1 2026), Launch (Falcon series commercial and government), and AI (xAI + X/Twitter, merged February 2026). Combined 2025 revenue was $18.674B, at a ~34% compound annual growth rate from 2023–2025.
3. SpaceX filed its S-1 prospectus with the SEC on May 20, 2026, planning to list on Nasdaq under ticker SPCX with a target valuation of $1.75–2 trillion. If completed, it would be the largest IPO in history, expected to raise over $30B — surpassing Saudi Aramco's ~$29B record in 2019.
The Three-Layer IPO Valuation Logic
To understand SpaceX's IPO, one key premise must be clear: this is not an ordinary tech listing. It is a complex valuation exercise pricing three distinct narratives simultaneously.
Layer 1: Starlink's Real Cash Flow
This is SpaceX's revenue engine today — the most tangible part of the entire valuation. In 2025, Starlink generated $11.387B in revenue (up 50% YoY) with EBITDA of $7.168B at a 63% margin. It operates in 164 countries with 10.3M broadband subscribers and 7.4M monthly active DTC (direct-to-cell) devices. This is a proven subscription business model, supportable at 30–40× price-to-sales multiples, implying a $400–600B standalone valuation.
Layer 2: Starship's Cost Revolution Promise
This is SpaceX's business amplifier. Falcon 9 reduced launch cost from $18,500/kg to $2,700/kg; Starship targets sub-$100/kg (a ~99% reduction). If achieved, Starlink's deployment cost drops dramatically, making each V3 satellite cheaper to orbit and deepening the competitive moat. Starship is also NASA's Artemis lunar lander. As of May 2026, Starship has completed 12 test flights with V3 having flown once, and commercial orbital payload launches are planned for H2 2026.
Layer 3: Space AI Infrastructure Vision
The most contested part of the valuation. SpaceX's prospectus claims a total addressable market of $28.5 trillion, with the AI segment accounting for $26.5 trillion (93%) — including $2.4T AI infrastructure, $760B AI consumer subscriptions, $600B AI advertising, and $22.7T AI enterprise. Anthropic has signed a $45B, 36-month compute rental agreement. SpaceX plans to deploy orbital AI compute satellites beginning 2028, leveraging vacuum cooling and unlimited solar power to overcome the two major data center bottlenecks. If realized, SpaceX would be not just a space company, but the world's lowest-cost AI compute platform.
These three layers combined give the $1.75–2T valuation its narrative logic. But the certainty gap between them is enormous — Layer 1 is proven, Layer 2 has technical foundation but uncertain timing, Layer 3 depends almost entirely on assumptions a decade out.
- Starlink: Revenue $11.387B · EBITDA margin 63% · 10.3M users
- Launch: Revenue $4.086B · EBITDA $653M · 170 launches in 2025
- AI: Revenue $3.201B · EBITDA -$1.237B · Colossus 1GW compute
三個事業,三種商業邏輯
火箭靠工業化複製降成本,星鏈靠衛星飛輪賺訂閱,AI靠算力租用建商業模型——同一家公司的三套截然不同的賺錢方式。
同一個公司,三個完全不同的生意
如果你問「SpaceX是什麼公司」,三個不同的人會給你三個不同的答案:做火箭的說它是「航天製造商」,用星鏈的說它是「衛星通信運營商」,AI圈子裡的人說它是「算力基礎設施平台」。三個答案都對,但又都不完整。
SpaceX的真正核心能力是:用工程的第一性原理(從物理定律而非行業慣例出發)把「複雜的高價東西」做成「簡單的低價東西」,然後大規模複製。這個能力在火箭領域的體現是可復用火箭,在衛星領域的體現是批量化低成本衛星,在AI領域的體現是「太空超算」的構想。理解這個底層邏輯,才能理解為什麼三個看似毫不相關的事業會在同一家公司裡共存。
火箭業務的商業邏輯:工業化複製,邊際成本遞減
怎麼製造:SpaceX 垂直整合了80%以上的火箭硬件製造,包括發動機(猛禽系列全自研)、箭體結構、推進系統、電子設備、整流罩等。這與傳統航天承包商依賴層層分包的模式形成鮮明對比。垂直整合帶來兩個好處:一是設計迭代速度快,設計-製造-測試全在內部閉環;二是成本透明,沒有層層外包的利潤漏損。
猎鷹9號新製造的內部成本約5,000萬美元,但一級火箭的邊際復用成本可降至1,500-2,000萬美元。復用5次時,邊際毛利率可達約60%;復用20次後,一級火箭的成本已接近維護費用(約1,500萬美元),形成極強的邊際成本遞減曲線。
怎麼銷售:SpaceX的客戶分為三類。第一類是政府客戶(NASA載人/貨運合同、美國軍方、情報機構),提供最高單次報酬但採購量有限。第二類是商業衛星客戶(電信公司、遙感公司),按每公斤有效載荷報價,猎鷹9號約2,700美元/公斤,遠低於競爭對手。第三類是SpaceX自己——猎鷹9號是自家星鏈衛星的主要發射運載,這讓「火箭」成為「星鏈」的垂直整合成本中心,不需要對外出售也有戰略意義。
怎麼定價:發射定價基於有效載荷質量和軌道類型,猎鷹9號LEO(低地球軌道)約2,700美元/公斤,GTO(地球同步轉移軌道)更高。相較之下,傳統火箭約18,500美元/公斤,猎鷹重型約1,400美元/公斤。星舰的目標是100美元/公斤以下,一旦達到,將對整個發射行業構成徹底的成本降維。
火箭業務的護城河:
星鏈的商業邏輯:衛星版Netflix,飛輪越轉越快
星鏈的商業模式從結構上看很像串流媒體:你付月費,拿到服務,下個月繼續付月費。但有一個關鍵差異:星鏈是SpaceX用自家火箭部署自家衛星,向全球沒有寬帶的地方提供服務,這個垂直整合創造了幾乎沒有其他公司能複製的成本優勢。
怎麼製造:星鏈衛星在華盛頓州雷德蒙德(Redmond)工廠生產,採用汽車式流水線,目前產能已達每週約70顆(每年約3,640顆)。SpaceX的衛星製造策略與傳統航天公司完全相反:大量使用汽車級或工業級商用現成組件(COTS),通過系統層面的冗余設計來替代單體的昂貴抗輻射元件。這讓單星成本從傳統衛星的百萬美元級降至數萬美元級(V2 Mini約50-100萬美元,V3預計進一步降低)。
怎麼銷售:星鏈採用直接面向消費者(DTC,Direct to Consumer)模式,用戶在官網訂購終端設備(約599-2,500美元)並繳月費。套餐從住宅寬帶(每月99-120美元)到航空(每月ARPU約25,000美元/機)和海事(每月ARPU約2,800美元/船)。2026年,星鏈還通過衛星直連手機(Direct to Cell)與全球30家運營商合作,月活設備達740萬台,把覆蓋範圍延伸到傳統蜂窩網絡無法觸達的區域。
怎麼定價:家庭寬帶月費在不同市場差異很大(美國約120美元,部分新興市場更低)。ARPU(每位用戶月均收入)從2023年的99美元下降至2026年Q1的66美元,下降原因是新興市場擴張(低收入用戶)和低價套餐推出。但這是「以量換質」的蓄意策略:ARPU下降但用戶量爆增(2023年230萬→2026Q1 1,030萬),總收入仍快速成長。
星鏈的護城河:
這是一個正在形成「飛輪效應」的業務:
`
更多發射(猎鷹9可復用降低成本)
→ 更多衛星入軌(9,600顆,全球75%)
→ 更廣覆蓋(164國)
→ 更多用戶(1,030萬)
→ 更多收入(EBITDA利潤率63%)
→ 資金支持更多衛星發射
`
關鍵壁壘在於:現有的9,600顆衛星佔用了大量低軌頻段資源,後進者即使建好衛星,也面臨頻段協調的「隱形壁壘」。此外,SpaceX用自家猎鷹9號部署星鏈,每顆V2 Mini衛星的入軌成本不到150萬美元,而任何依賴外部火箭的競爭者(如Kuiper使用ULA、Ariane 6等)部署成本都遠高於此。
AI業務的商業邏輯:最大的市場,最早期的商業化
AI板塊是SpaceX三個業務中最難評估的。它由兩個完全不同的資產組成:xAI(包含Colossus算力集群和Grok大語言模型),以及X(原推特,擁有13億賬戶的社群媒體平台)。
怎麼「製造」算力:Colossus超算集群位於田納西州孟菲斯,目前總算力1GW(2025年Q1為0.3GW,一年內成長3.3倍),搭載22萬張高端GPU(15萬張H100、5萬張H200、2萬張B200)。這個規模使xAI成為全球算力密度最高的私人算力提供者之一。
怎麼銷售算力:2026年5月,xAI與Anthropic簽署算力租用合同,Anthropic每月向SpaceX支付12.5億美元購買Colossus的計算容量,合同期2026年6月至2029年5月,36個月總合同額約450億美元。這是AI板塊第一個有確定性的大規模商業收入來源。
X(推特)的商業邏輯:X平台擁有13億活躍賬戶(月活約5.5億),每日3.5億條內容。收入來自廣告和訂閱(X Premium)。馬斯克2022年收購後,X的廣告收入一度大幅下滑,但平台作為「全球最大的實時輿論場」的媒體影響力仍然極強,且為xAI的Grok模型提供了訓練數據和分發渠道。
未來:太空算力
SpaceX的AI最大敘事在於「把算力搬到太空」。地面數據中心的兩大成本瓶頸是電力和散熱。太空環境理論上能同時解決:太陽能提供近乎無限的電力,真空環境提供近乎完美的散熱條件。SpaceX計畫2028年開始部署軌道AI計算衛星,每年向太空部署高達100GW算力。但這個計畫的物理前提是星舰每年能發射數千次——一個「至少5年後」的故事。
財務傳導:2025年AI板塊收入32.01億美元,但資本開支127.27億美元(大量GPU採購),造成EBITDA虧損12.37億、運營虧損63.55億。Anthropic合同若全額執行,2026年起AI板塊每月可貢獻12.5億美元算力收入,全年150億美元——但這還需要扣除電力、折舊和運維成本才能看到真正的利潤。
產品線全貌
火箭系列:
星鏈衛星:
AI/算力產品:
One Company, Three Completely Different Businesses
Ask "what is SpaceX?" and three different people give three different answers: rocket engineers say "aerospace manufacturer"; Starlink users say "satellite internet operator"; AI circles say "compute infrastructure platform." All three are correct — and all incomplete.
SpaceX's true core capability is: applying first-principles engineering (starting from physics, not industry convention) to make complex, expensive things simple and cheap, then replicate at scale. In rockets, this became reusability. In satellites, mass-produced low-cost hardware. In AI, the vision of orbital supercomputing. Only by understanding this underlying logic does it make sense why three seemingly unrelated businesses coexist within one company.
Launch: Industrial Replication and Falling Marginal Costs
How it's made: SpaceX vertically integrates over 80% of rocket hardware — engines (Merlin, Raptor), structures, propulsion, avionics, fairings. A new Falcon 9 costs approximately $50M to produce, but a reused first stage's marginal cost drops to $15–20M. By the fifth reuse, EBITDA margins on the reflight approach 60%; by the 20th, first-stage cost nears maintenance fees (~$15M), creating a powerful declining marginal-cost curve.
How it's sold: Three customer types — (1) government (NASA crewed/cargo, DoD, intelligence), highest per-launch revenue; (2) commercial satellite operators (telecoms, remote sensing), priced by kg to orbit; (3) SpaceX itself (Starlink deployments), where rockets serve as a vertically integrated cost center. Falcon 9 LEO pricing: ~$2,700/kg vs. competitors at $10,000–18,500/kg. Starship targets sub-$100/kg.
Launch moat — three pillars:
- Cost: $2,700/kg vs. $10,000–18,500/kg for alternatives — a structural competitive gap
- Reliability: Falcon 9 success rate 99.3% (579 flights), Falcon Heavy 100% (11 flights) — the purchasing criterion for government and high-value payloads that can't be bought with money
- Cadence: 170 orbital launches in 2025, 40 in 2026 Q1 — no customer needs to wait, an invisible but massive competitive advantage
Starlink: Satellite Netflix, a Flywheel Spinning Faster
Starlink structurally resembles a streaming service: pay monthly, receive service, repeat. The critical difference: SpaceX deploys satellites on its own rockets, serving regions without broadband globally — a vertical integration no competitor can replicate.
Manufacturing: Starlink satellites are produced in Redmond, WA on automotive-style assembly lines, currently ~70 satellites/week (~3,640/year). SpaceX uses commercial off-the-shelf (COTS) components with system-level redundancy replacing expensive radiation-hardened parts — reducing per-satellite cost from the traditional multi-million range to tens of thousands (V2 Mini: ~$500K–1M; V3 expected lower).
Sales: Direct-to-consumer (DTC). Users order the terminal ($599–$2,500) and pay monthly fees. Tiers: residential (~$99–120/mo.), aviation (~$25,000 ARPU/aircraft), maritime (~$2,800 ARPU/vessel). DTC (direct-to-cell) is live with 30 carriers globally, reaching 7.4M monthly active devices.
The flywheel:
More launches (F9 reuse → lower cost) → More satellites (9,600 in orbit, 75% of all operational sats) → Wider coverage (164 countries) → More users (10.3M) → More revenue (63% EBITDA) → Funds more satellite launches
Key moat: 9,600 in-orbit satellites have claimed large portions of low-orbit spectrum; late entrants face hidden spectrum-coordination barriers even after building their constellation. SpaceX deploys V2 Mini for ~$1M/satellite landed; competitors relying on external rockets face $3–5M+/satellite.
AI: Largest Market, Earliest-Stage Commercialization
The AI segment is the hardest to value. It comprises xAI (Colossus compute cluster + Grok LLM) and X (formerly Twitter, 1.3B accounts, 550M monthly active).
Colossus compute cluster: Located in Memphis, Tennessee. Current capacity: 1GW total (from 0.3GW in Q1 2025 — 3.3× growth in one year), housing 220,000 high-end GPUs (150K H100, 50K H200, 20K B200). xAI is among the highest-density private compute providers in the world.
Selling compute: In May 2026, xAI signed a compute rental agreement with Anthropic — $1.25B/month for Colossus capacity, running June 2026 to May 2029, $45B total. This is the AI segment's first large-scale, visible commercial revenue source.
X (Twitter): 1.3B active accounts, 350M posts/day. Revenue from advertising and X Premium subscriptions. Strategically, X provides: (1) the world's largest real-time text corpus for training Grok; (2) organic distribution for every SpaceX milestone.
Future: Space Compute. SpaceX plans to deploy orbital AI compute satellites from 2028 — leveraging vacuum cooling and nearly unlimited solar power. Target: 100GW of space-based compute per year. This requires thousands of Starship launches annually — a story for "at least 5 years from now."
Full Product Portfolio
- Falcon 9: 22.8t LEO, 99%+ success rate, max 34 reuses, $2,700/kg. Main workhorse since 2010, 620+ flights.
- Falcon Heavy: 63.8t LEO, 100% success (11 flights), for heavy government/commercial missions.
- Starship V3 (in test): 124.4m tall, target 100–200t LEO, fully reusable, target sub-$100/kg, 12 test flights through May 2026.
- V2 Mini satellites: Current production model, 250–300kg, argon Hall thrusters (first mass-produced argon electric propulsion), inter-satellite laser links.
- V3 satellites (mass production coming): 1Tbps downlink bandwidth per satellite (several times V2 Mini), 60 per Starship launch, H2 2026 deployment start.
- Colossus I/II: 1GW ground-based compute, H100/H200/B200 GPUs.
- Grok: xAI's LLM integrated in X; ~117M monthly users accessing AI features.
從「你的火箭會爆炸」到「萬億估值」的二十四年
五個高風險賭注,每一個都在最不可能的時機押下去,每一個都賭贏了。
五個關鍵決定,每一個都是高風險賭注
SpaceX的歷史是一部「在別人認為不可能的地方找到物理法則允許的可能性」的連續押注。以下五個決定,每一個在當時都看起來像賭博,但每一個都在事後被驗證為正確:
第一個決定:2002年,馬斯克決定直接去到第一原理
馬斯克在決定創辦SpaceX之前,去查了火箭的原材料成本。他發現:製造火箭的鋁合金、鈦、碳纖維等原材料,成本大約只佔市場售價的2%。剩下的98%是什麼?是行業慣例、分包商利潤、過時的設計規範、以及「我們一直這麼做」的惰性。
這個發現讓他得出結論:「造火箭成本貴,是組織模式的問題,不是物理的問題。」他以約1億美元個人資金創立SpaceX,核心使命是「降低太空運輸成本、推動人類成為多行星物種」。這個使命不是公關話術,它直接決定了SpaceX的所有技術選擇:可復用火箭、批量製造、商用現成組件(COTS)替代昂貴的太空級元件。
第二個決定:2008年,三次失敗後的第四次發射,和那份NASA合同
2006到2008年,獵鷹1號連續三次發射失敗。公司資金快耗盡,員工士氣低落,外界普遍認為SpaceX活不過2008年。馬斯克用最後的資金進行了第四次發射——成功。這次成功的象徵意義遠大於技術意義:它向市場和政府證明,私人公司可以進入軌道。
幾個月後,NASA授予SpaceX價值16億美元的商業補給服務(CRS)合同——用私人飛船向國際空間站(ISS)運送貨物。這份合同不只是資金,更是商業合法性的背書:如果連NASA都信任SpaceX,那誰還敢說這家公司不靠譜?它為後續猎鷹9號和龍飛船的研發提供了生存資金,也確立了SpaceX在美國政府商業航天供應鏈中的核心地位。
第三個決定:2015-2017年,完成行業普遍認為「不可能」的一級回收
2015年前,航天行業的共識是:火箭在大氣層高速再入後回收,物理上幾乎不可行。極端的熱力學條件、燃料消耗計算的精確性要求、着陸腿展開的時序控制……任何一環出錯都會導致失敗。
但SpaceX把「不可能」翻譯成了「需要解決哪些工程問題」,然後逐一解決。2015年12月,猎鷹9號一級首次在陸地垂直回收成功;2016年4月,首次在海上無人船(Of Course I Still Love You)回收成功;2017年3月,首次復用火箭發射並再次回收。
復用帶來的成本降幅是結構性的:猎鷹9號一級的邊際復用成本從5,000萬降至1,500-2,000萬美元,復用5次時邊際毛利率可達60%。這不只是成本節約,而是整個商業模式的重構——火箭從「一次性消耗品」變成了「可以反覆使用的基礎設施」。
第四個決定:2019-2023年,把星鏈從「最貴的互聯網項目」熬到現金流轉正
2019年首批星鏈試驗衛星發射時,行業評論普遍懷疑:在軌10,000顆衛星是天文數字,建網成本將遠超任何商業回報。星鏈前幾年確實大量燒錢——但SpaceX有一個別人沒有的優勢:自家的猎鷹9號是部署成本最低的火箭,而且隨著復用次數提升,成本還在持續下降。
更重要的是,SpaceX在批量製造星鏈衛星上應用了和火箭相同的邏輯:大量使用商用現成組件(COTS),把傳統航天的昂貴太空級砷化鎵太陽能電池換成改造過的商用硅基電池;把推進劑從昂貴稀缺的氪氣換成廉價的氬氣(全球氪氣年產量甚至撐不起萬顆衛星的規模,而氬氣近乎無限)。每一個替換都不是偶然,而是從「第一性原理」出發的系統性降本。
2023年11月,馬斯克在社群媒體宣布「星鏈業務年末首次實現盈利,現金流轉正」。從2019年試驗到2023年盈利,SpaceX用四年時間把一個「天方夜譚的太空互聯網夢想」變成了「63%EBITDA利潤率的現金機器」。
第五個決定:2026年,收購xAI+X,重新定義自己是「AI基礎設施」
2026年2月,SpaceX正式完成對xAI(含X推特)的收購。這個決定在商業邏輯上存在爭議:SpaceX真的需要一個月活5.5億的社群媒體平台,才能把太空算力這個生意做起來嗎?
馬斯克的答案是:是的,原因有三。第一,X是全球最大的實時文本數據庫,是訓練Grok大語言模型的核心數據源。第二,xAI的Colossus算力集群可以直接服務外部客戶(如Anthropic),提供可見的商業收入。第三,將來太空算力衛星的數據回傳,需要一個全球覆蓋的通信網絡——星鏈恰好可以扮演這個角色。
但批評者指出:xAI+X的2025年運營虧損高達63.55億美元,而SpaceX在IPO前夕把這個「燒錢機器」納入上市主體,是在稀釋星鏈和火箭業務的高質量現金流,換取更大的AI敘事空間。這個策略正確與否,取決於AI板塊能否在2027-2028年之前實現EBITDA轉正。
近五年策略轉變:從「燒錢造衛星」到「萬億AI估值」
2021→2022:星鏈飛輪啟動
用戶從50萬增至150萬,星鏈開始產生可見的訂閱收入。火箭發射次數超過60次,復用常態化讓成本持續下降。這一年,SpaceX用「規模+復用」組合,讓星鏈和火箭形成了正向循環的飛輪。
2022→2023:現金流轉正,從燒錢到造血
星鏈ARPU在高ARPU的早期用戶基數下仍相對較高(約$99/月),收入爆發(年增+88%至103.87億$),現金流歷史首次轉正。SpaceX從「需要持續外部融資的項目公司」成為「能夠自我造血的基礎設施平台」。估值突破1,000億美元。
2023→2024:估值翻倍,市場開始為星舰的未來定價
2024年發射次數突破134次,星舰成功回收(IFT-11),解鎖超大運力時代的想象空間。星鏈用戶過500萬,DTC(衛星直連手機)業務啟動。市場估值從2024年初3,500億美元躍升至年末8,000億美元,翻倍背後是「星舰+太空算力」的宏大敘事開始定價。
2024→2025:AI收購落地,資本開支激增
星鏈EBITDA 71.68億$,但全公司因AI板塊的127.27億資本開支導致整體運營虧損。馬斯克系的AI生態(xAI→Grok→X)正式合併進SpaceX。資本開支首次遠超收入,說明公司仍處於「為未來大規模投入」的階段。
2025→2026:IPO,打包三層邏輯上市
S-1於2026年5月20日提交SEC。招股書把三個事業、三種商業邏輯、三個時間尺度的故事打包成一個「萬億估值科技公司」的敘事。核心命題:你願意為哪一層故事付多少錢?
公司文化:第一性原理、垂直整合、失敗也要快
SpaceX的文化核心是馬斯克的工程師哲學,它有三個在工程決策上有具體體現的原則:
第一性原理思維:不問「行業怎麼做」,而問「物理上允許做到多便宜」。火箭原材料只佔售價2%,所以把剩下98%的組織成本去掉,是有物理依據的目標。每次面對「這是不可能的」評估,SpaceX的工程師都會問:「物理定律在哪裡說這不可能?」
垂直整合哲學:80%以上的火箭硬件內部製造。不是因為分包一定不好,而是因為垂直整合讓設計→製造→測試→迭代的閉環最短,能夠在別人還在等供應商報價的時候就完成了三次設計迭代。
快速失敗、快速迭代:獵鷹1號失敗三次,星舰爆炸多次,每一次失敗後都被視為「免費的工程數據」。這種對失敗的容忍度,是SpaceX能在試飛節奏上遠超NASA或歐空局的組織文化原因。公司有一個內部說法:「如果你不曾爆炸過你的火箭,你就沒有測試得足夠快。」
Five Defining Bets — Each Looked Like a Gamble at the Time
SpaceX's history is a series of high-conviction bets made where others saw impossibility. Five decisions shaped the company into what it is today:
Bet 1 (2002): Go to First Principles
Before founding SpaceX, Musk looked up the raw material cost of a rocket. Aluminum alloys, titanium, carbon fiber — the materials cost roughly 2% of the market price. The other 98% was industry convention, subcontractor margin, outdated specs, and "we've always done it this way" inertia. His conclusion: "Rockets being expensive is an organizational problem, not a physics problem." He invested ~$100M of personal capital and defined the mission: reduce the cost of space transport and make humanity multiplanetary.
Bet 2 (2008): The Fourth Launch and the NASA Contract
Falcon 1 failed three consecutive times (2006–2008). Money nearly exhausted, morale low, the outside world wrote SpaceX off. Musk spent the last of the capital on a fourth launch — which succeeded. Months later, NASA awarded a $1.6B Commercial Resupply Services (CRS) contract to deliver cargo to the ISS. This wasn't just money: it was a government endorsement that private companies could reach orbit — and it funded Falcon 9 and Dragon development.
Bet 3 (2015–2017): Landing Rockets — "Physically Impossible"
Before 2015, the industry consensus was that recovering a first stage after atmospheric re-entry at high velocity was nearly impossible. SpaceX translated "impossible" into "what engineering problems need to be solved?" — and solved them one by one. December 2015: first successful land landing. April 2016: first ocean drone-ship recovery. March 2017: first reused Falcon 9 launch and recovery again. The economics: first-stage marginal cost fell from $50M to $15–20M; EBITDA margin on a reflight approaches 60%.
Bet 4 (2019–2023): Making Starlink's "Most Expensive Internet Project" Cash-Flow Positive
When the first test Starlink satellites launched in 2019, analysts doubted any commercial return could justify the cost. But SpaceX applied the same logic to satellite manufacturing: COTS components, argon (cheap, abundant) propellant instead of krypton (scarce), silicon-based solar cells instead of expensive GaAs. Each substitution was principled cost reduction, not corner-cutting. In November 2023, Musk announced Starlink had achieved its first profitable year with positive cash flow. Four years from prototype to a business running at 63% EBITDA margin.
Bet 5 (2026): Acquire xAI + X, Redefine as "AI Infrastructure"
In February 2026, SpaceX completed the acquisition of xAI (including X/Twitter). The commercial logic is contested: does SpaceX really need a 550M-MAU social platform to build the space compute business? Musk's three-part answer: (1) X is the world's largest real-time text corpus — the core training data for Grok, accessible at zero marginal cost vs. licensing fees for any competitor; (2) Colossus generates direct external revenue (Anthropic contract); (3) future space compute data return needs a global communication network — Starlink already is one. Critics note xAI+X had $6.355B operating losses in 2025, and including this "cash-burning machine" in the IPO dilutes Starlink and Launch's high-quality cash flows in exchange for a larger AI narrative.
Strategic Pivots 2021–2026
2021–2022: Starlink flywheel ignites. Users grew 50K→1.5M. Launch cadence exceeded 60/year, normalized reuse driving continuous cost reductions. Starlink and Launch began a self-reinforcing cycle.
2022–2023: Cash flow turns positive. ARPU still high from early-adopter mix (~$99/mo.); revenue exploded +88% to $10.387B. SpaceX transformed from a project requiring continuous external capital to a self-funding infrastructure platform. Valuation crossed $100B.
2023–2024: Valuation doubled; market began pricing Starship's future. Launch cadence exceeded 134/year; Starship achieved successful recovery (IFT-11), unlocking the super-heavy payload imagination. Valuation leaped from ~$350B (early 2024) to ~$800B (year-end).
2024–2025: AI acquisition landed; capex surged. Starlink EBITDA $7.168B, but company-wide operating loss due to $12.727B AI capex. The Musk AI ecosystem (xAI → Grok → X) formally merged into SpaceX's consolidated entity.
2025–2026: IPO. S-1 filed May 20, 2026. Three businesses, three logics, three time horizons — packaged into a single "trillion-dollar tech company" narrative. The core question: how much are you willing to pay for each layer?
Company Culture: First Principles, Vertical Integration, Fail Fast
First-principles thinking: Don't ask "how does the industry do it?" Ask "how cheap does physics allow?" Materials cost 2% of rocket price — so eliminating 98% of organizational overhead is physically grounded. Every "that's impossible" claim is met with: "Where exactly does physics say so?"
Vertical integration: 80%+ of hardware manufactured internally. Not because outsourcing is inherently bad, but because it creates the shortest design → manufacture → test → iterate loop — allowing three design iterations while competitors are still waiting for supplier quotes.
Fail fast, iterate faster: Falcon 1 failed three times. Starship exploded multiple times. Each failure is treated as "free engineering data." This tolerance for failure is why SpaceX's test cadence far outpaces NASA or ESA. Internal saying: "If you haven't blown up your rocket, you haven't tested fast enough."
- Key milestones: 2002 founded ($1B) → 2008 NASA contract ($4B) → 2015 stage recovery ($10B) → 2023 cash-flow positive ($180B) → 2024 Starship recovery ($800B) → 2026 IPO ($1.75–2T)
IPO財務品質:真實的現金流 vs. 宏大的敘事
星鏈在賺錢,AI在燒錢,火箭維持穩定。三個業務放在同一張資產負債表上,才出現了90倍本銷比的估值。
核心判斷:星鏈在賺錢,AI在燒錢,整體收入增速仍健康
SpaceX 於2026年5月20日提交的S-1招股書,首次完整披露了三大業務板塊的財務數字。在解讀這些數字之前,需要先確立一個基本框架:這家公司的財務結構呈現「星鏈健康、火箭穩定、AI大量燒錢」的三速結構。2025年全公司出現整體經營虧損,但原因不是核心業務惡化,而是AI板塊的大規模前期投入。
Core Assessment: Starlink Makes Money, AI Burns It, Overall Growth Remains Healthy
SpaceX's S-1 filed May 20, 2026, provided the first complete disclosure of all three business segments. The financial structure is clear: "Starlink healthy, Launch stable, AI burning heavily." The company posted an overall operating loss in 2025 — not because core businesses deteriorated, but due to the massive front-loaded AI investment.
銷售歷史新高的背後,是星鏈業務的一枝獨秀
SpaceX 2023-2025年合併收入分別為103.87億、140.15億、186.74億美元,2026年Q1為46.94億美元,三年複合增速約34%,在這個規模下仍是相當強勁的成長。但成長的結構需要仔細解析:
星鏈業務是真正的增長引擎。2025年星鏈收入113.87億美元,同比增長50%,收入佔比從2023年的37%提升至2025年的61%,2026年Q1進一步達69%。星鏈的EBITDA(稅前折舊攤銷前利潤,是衡量業務現金創造能力的核心指標)達71.68億美元,EBITDA利潤率63%——這個數字遠高於一般電信公司(約20-30%),更接近頂尖SaaS(軟件訂閱)公司的水準。
火箭業務是穩定但增速放緩的現金流來源。2023-2025年收入分別為35.6億、38.0億、40.86億美元,三年累計增長15%,增速遠低於星鏈。EBITDA 6.53億,利潤率16%,合理但不突出。
AI業務(xAI+X)是最大的「燒錢黑洞」。2025年AI板塊收入32.01億,但EBITDA虧損12.37億,運營虧損高達63.55億——這意味著AI板塊每賺1美元,就要在運營上虧接近2美元。主因是GPU採購(資本開支127.27億,佔AI板塊收入的398%)和電力、薪酬等經常性支出。
整體經營利潤:負值,但有可解讀的原因
SpaceX 2025年整體經營利潤約為-25.89億美元(2026年Q1為-19.43億美元)。乍看之下似乎不健康,但需要拆解:
換句話說:如果沒有AI板塊,SpaceX 2025年的整體EBITDA仍可能達到70-80億美元,是一家財務非常健康的公司。AI板塊的虧損是「戰略性的」——在為未來的萬億市場投入,而不是核心業務失敗。
AI板塊的一次性沖銷已過,但商業化尚未到位
2025年AI板塊的大量燒錢(資本開支127.27億、運營虧損63.55億)是否會在2026年繼續?
Anthropic合同是第一個可見的商業化信號。2026年6月起,Anthropic每月向SpaceX支付12.5億美元購買Colossus算力,合同期36個月,總額450億美元。這意味著從2026年下半年起,AI板塊每半年的確定性算力收入達75億美元以上。如果Colossus的使用率維持高位,2026年AI板塊有可能首次出現EBITDA轉正跡象。
2026年Q1的資本開支(AI板塊77億,YoY+201%但季度上環比下降)也出現了「可能峰值已過」的信號——Colossus II完成建設後,GPU採購量應當逐漸下降。
資本效率:投入最多的地方是AI,回報最確定的地方是星鏈
2025年資本開支207.37億美元,分配如下:
這個分配透露了管理層的判斷:AI的長期潛在回報高於星鏈和火箭,所以把最多錢投在這裡。但從當前已驗證的商業回報率來看,這個排序是顛倒的——星鏈的CAPEX/收入比(41.78/113.87=37%)遠好於AI(127.27/32.01=397%)。
未來的關鍵問題是:AI板塊的資本開支什麼時候開始見到回報?Anthropic合同是第一個答案。
資料核實:以上所有財務數字均來自SpaceX S-1招股說明書(2026年5月20日提交),透過東吳證券和中航證券研究報告整理,各業務EBITDA和經營利潤由研究機構整理計算,個別指標(如淨利率)為研究機構估算值。
Record Revenue Driven Almost Entirely by Starlink
SpaceX's combined revenue: 2023 $10.387B → 2024 $14.015B → 2025 $18.674B → 2026 Q1 $4.694B. Three-year CAGR ~34% — impressive at this revenue scale. But the structure matters:
Starlink is the true growth engine. 2025 revenue $11.387B (+50% YoY), rising from 37% of mix in 2023 to 61% in 2025 and 69% in Q1 2026. EBITDA $7.168B at 63% margin — far above typical telecom (20–30%), closer to top-tier SaaS.
Launch is stable but slowing. 2023–2025 revenue: $3.56B → $3.80B → $4.086B (+15% over three years). EBITDA $653M at 16% margin — reasonable, not exceptional.
AI (xAI + X) is the largest cash drain. 2025 AI revenue $3.201B; EBITDA loss $1.237B; operating loss $6.355B — for every $1 earned, the segment loses ~$2 in operating costs. The driver: $12.727B capex (GPU procurement) = 398% of AI segment revenue.
Operating Profit: Negative, But Explainable
SpaceX 2025 operating profit: -$2.589B (Q1 2026: -$1.943B). Decomposing:
- Starlink EBITDA $7.168B — likely positive operating income after D&A
- Launch EBITDA $653M — positive
- AI EBITDA -$1.237B plus operating costs — the drag
- AI 2025 capex $12.727B → annual depreciation (5yr) ~$3–4B
Without the AI segment, SpaceX 2025 EBITDA would likely have been $70–80B — a financially very healthy company. The AI losses are "strategic" — pre-investing for a trillion-dollar market, not core business failure.
AI Segment: Commercialization Not Yet in Place, but Anthropic Contract Changes the Equation
Will the 2025 AI burn rate ($12.727B capex, $6.355B operating loss) continue into 2026? The Anthropic contract is the first visible commercialization signal. Beginning June 2026, Anthropic pays SpaceX $1.25B/month for Colossus compute capacity — 36 months, $45B total. From H2 2026, AI segment's annualized compute revenue becomes ~$15B. If Colossus utilization stays high, 2026 AI EBITDA could turn positive for the first time. Q1 2026 AI capex ($7.7B, +201% YoY but declining sequentially from Q4 2025) suggests the construction peak may have passed.
Revenue Mix Shift
| Segment | 2023 | 2024 | 2025 | 2026 Q1 |
|---|---|---|---|---|
| Starlink | 37% | 54% | 61% | 69% |
| Launch | 34% | 27% | 22% | 13% |
| AI (xAI+X) | 29% | 19% | 17% | 17% |
2025 EBITDA by Segment
| Segment | 2025 EBITDA (B$) | Margin | Notes |
|---|---|---|---|
| Starlink | +$7.168B | 63% | SaaS-like subscription model |
| Launch | +$0.653B | 16% | Stable, low-growth |
| AI (xAI+X) | -$1.237B | — | Strategic capex phase |
Capital Efficiency: Most Investment in AI, Most Certain Returns in Starlink
2025 capex $20.737B: AI $12.727B (61%), Starlink $4.178B (20%), Launch $3.832B (18%). The allocation reveals management's conviction: AI has the highest long-term potential return — even though current verified return rates are inverted (Starlink Capex/Revenue = 37% vs. AI 397%). The key question: when does AI capex start generating returns? The Anthropic contract is the first answer.
Data source: All financial figures from SpaceX S-1 prospectus (filed May 20, 2026), compiled via research reports from Guosen Securities and CATIC Securities. EBITDA and operating profit figures are research institution calculations; some metrics (e.g., net margin) are analyst estimates.
護城河的真實深度——星鏈和火箭
火箭的護城河是三重壁壘(成本×可靠性×產能頻率),短期幾乎無法被追上;星鏈的護城河顯著但有條件,Kuiper是真實的長期威脅。
評估護城河的框架
要評估SpaceX的護城河,需要先確立一個標準:護城河的強弱不在於「現在有多難被追上」,而在於「在競爭者有充足資金和時間的情況下,五到十年後是否仍難以被追上」。依此標準,SpaceX的兩個核心業務的護城河性質截然不同。
火箭業務:護城河是結構性的,短期(3-5年)幾乎無法被追上,因為壁壘是多維度疊加的(成本×可靠性×產能頻率×政府資質)。
星鏈:護城河是顯著的但有條件的,最大的長期威脅來自亞馬遜Kuiper的生態資源,以及潛在的頻段/軌道競爭。
火箭業務:三重壁壘的結構性護城河
成本壁壘
SpaceX透過可復用技術把猎鷹9號的入軌成本壓到2,700美元/公斤,相較傳統火箭的18,500美元/公斤,降幅85%。這不只是技術領先,而是商業模式的根本性差異:
當猎鷹9號的一級助推器可以復用34次時,每次發射的邊際成本已接近維護費用(約1,500萬美元),而不是重新製造一枚火箭(5,000萬美元)。競爭者要複製這個成本結構,需要同時解決:回收系統的工程設計、復用後的可靠性驗證、高頻次發射的供應鏈管理……每一步都需要時間和大量失敗的學習成本。
可靠性壁壘
猎鷹9號的發射成功率99.3%(579次)、猎鷹重型100%(11次)。這個數字在商業發射市場的意義是:對衛星運營商來說,一次失敗可能損失數億美元的衛星;對政府客戶來說,可靠性是非價格因素的採購標準。任何新進者需要用大量成功發射案例才能建立等效的信任,而這需要時間,不能用錢直接買到。
產能頻率壁壘
2025年170次軌道發射,2026年Q1已完成40次(年化160次+),這個發射節奏全球唯一。對商業客戶而言,「不需要等待,隨時可以發射」本身就是一種護城河。傳統發射服務商往往需要客戶等待18-24個月的排期,而SpaceX的高頻發射能力讓衛星運營商的部署計畫更靈活。
競爭格局矩陣
競爭格局正文分析:
在中型商業LEO發射領域,Rocket Lab的電子號雖然是最接近的商業競爭者,但其主攻150公斤以下的小型衛星市場,與猎鷹9號22.8噸LEO運力的定位完全不同。中子號(13噸LEO,中型可復用火箭)目前仍在研發中,最快2026-2027年才有可能首飛商業任務,且成本競爭力遠不及猎鷹9號。
在超重型市場,SpaceX幾乎沒有競爭對手。藍色起源的新格倫火箭(45噸LEO)剛完成首飛,但發射頻率和可靠性記錄仍遠不及SpaceX。ULA的火神半人馬(27噸LEO)在2025年首飛,但每次發射成本遠高於猎鷹重型,且不可復用。
星鏈:正在形成的飛輪,但Kuiper是真實的長期威脅
護城河的真實來源
星鏈的護城河不是單一的技術優勢,而是幾個相互強化的因素:
衛星數量的先發優勢:9,600顆在軌衛星佔全球可操控衛星的75%,形成對低軌頻段的先占優勢。低軌衛星的頻段資源是有限的,FCC批准SpaceX部署的衛星數量已佔全球低軌規劃總量的近70%。後進者即使建好衛星,也面臨頻段協調的隱性壁壘。
自家火箭帶來的成本護城河:每顆V2 Mini衛星的入軌成本估計不到150萬美元(猎鷹9號每次發射約6,000萬美元,可載60顆V2 Mini,平均100萬美元/顆),而依賴外部火箭的競爭者(如Kuiper)每顆衛星的入軌成本估計在300-500萬美元以上。這個成本差距直接影響到星座的盈虧平衡點和利潤率。
網絡效應與覆蓋廣度:164個國家的覆蓋,使星鏈在海事、航空等高ARPU場景中具備「全球唯一解」的地位——極地航線上,只有星鏈和OneWeb能穩定提供服務。DTC(衛星直連手機)服務與30家全球運營商的合作,也讓星鏈形成了更廣泛的分發網絡。
主要競爭威脅分析
OneWeb(Eutelsat合并後):648顆在軌衛星,成本遠高於星鏈,市場份額微弱。OneWeb缺少自家火箭,製造成本也高,短期不構成重大威脅。
亞馬遜Kuiper:最值得認真對待的競爭者。Kuiper規劃3,236顆衛星,2025年完成首批27顆衛星發射,2026年計畫商業化。背後是亞馬遜的電商生態(採購折扣)、AWS雲服務(潛在企業客戶協同)、Prime會員(潛在消費者分發渠道)。
但Kuiper有一個結構性弱點:它依賴外部火箭(ULA火神、亞馬遜自己投資的火箭公司)部署衛星,發射成本遠高於SpaceX。如果Kuiper的衛星入軌成本是SpaceX的3-5倍,即使兩者技術相當,Kuiper也難以在利潤率上與星鏈競爭。Kuiper的真正威脅是「生態整合」(亞馬遜給自家AWS用戶捆綁銷售),而不是純粹的成本競爭。
中國競爭對手(中國星網GW、千帆星座):主要在中國國內市場運作,短期不會與星鏈在國際市場形成直接競爭。但從頻段資源的角度,中國的大規模星座計畫仍在搶佔未來的軌道位置。
| 競爭者 | 競爭領域 | SpaceX 的優勢 | 主要觀點 |
|---|---|---|---|
| Rocket Lab Electron / Neutron |
小型衛星 LEO 發射 | 成本($2,700/kg vs $10,000+)、可靠性99.3%、年發射170次——Rocket Lab無法服務大型商業有效載荷 | Electron主攻150kg以下,定位無重疊;Neutron(13噸)最快2026-27首飛,成本差距仍大 |
| ULA Vulcan / Northrop Grumman |
超重型 / 政府軍事發射 | 成本+可靠性組合;已承接大量NRO和DoD任務,政府資質完備 | ULA在軍事任務仍有歷史地位,但SpaceX政府份額持續擴大;Vulcan不可復用,成本高於獵鷹重型 |
| Blue Origin New Glenn |
超重型商業發射 | 星舰(目標100噸LEO、$100/kg以下)vs 新格倫(45噸LEO);星舰已完成12次試飛 | 新格倫剛完成首飛,發射頻率和可靠性記錄遠不及SpaceX;超重型商業市場短期SpaceX幾乎獨佔 |
| Amazon Kuiper | 衛星寬帶互聯網 | 9,600顆在軌(Kuiper 27顆);自家火箭入軌成本 <$1M/顆 vs Kuiper約$3-5M/顆;頻段先占、164國覆蓋 | Kuiper唯一差異化是AWS生態協同(捆綁銷售),而非成本競爭;入軌成本結構性差距短期難以縮小 |
| OneWeb Eutelsat |
衛星寬帶互聯網 | 規模(9,600 vs 648顆);ARPU優勢(海事/航空高端套餐);覆蓋廣度 | 無自家火箭,製造成本高,市場份額微弱;短期不構成重大威脅 |
| 中國商業航天 藍箭 / 天兵 / 星河動力 |
中型商業 LEO 發射 | 中國可復用火箭仍在試飛,成本和可靠性均不及獵鷹9號;SpaceX持有全球最大政府和商業客戶基礎 | 政策保護下主要在國內市場競爭;短期不構成國際發射市場威脅 |
| 中國星網 / 千帆 GW / Qianfan |
衛星互聯網(國內) | 國際市場無直接競爭;頻段/軌道位置競爭是唯一隱性壓力 | 主要搶佔未來低軌頻段資源,不直接與星鏈在商業訂閱市場競爭 |
Moat Assessment Framework
The right question is not "how hard is it to catch up today?" but "five to ten years from now, with adequate capital and time, can competitors still not catch up?" By this standard, SpaceX's two core businesses have fundamentally different moat profiles.
Launch: Structurally unassailable in the near term (3–5 years) — the barriers are multidimensional and compounding (cost × reliability × cadence × government certification).
Starlink: Significant but conditional — the greatest long-term threat comes from Amazon's Kuiper and potential spectrum/orbit competition.
Launch: Triple-Barrier Structural Moat
Cost barrier: Falcon 9 launch cost: $2,700/kg vs. $18,500/kg for traditional rockets — an 85% reduction. When a first-stage booster can be reused 34 times, the marginal launch cost approaches maintenance fees (~$15M) rather than new rocket manufacturing (~$50M). Replicating this requires simultaneously solving: recovery system engineering, post-reuse reliability verification, high-cadence supply chain management — each requiring time and expensive failure.
Reliability barrier: Falcon 9: 99.3% success (579 flights); Falcon Heavy: 100% (11 flights). For satellite operators, a single failure means hundreds of millions in payload loss. For government clients, reliability is a non-price procurement criterion. New entrants need extensive successful flight records to earn equivalent trust — time cannot be bought with money.
Cadence barrier: 170 orbital launches in 2025; 40 in Q1 2026 (annualized 160+). This pace is globally unique. Traditional launch providers require 18–24 month booking windows; SpaceX's high cadence lets satellite operators deploy on flexible timelines. The invisible advantage compounds.
Starlink: Forming Flywheel, But Kuiper Is a Real Long-Term Threat
First-mover scale advantage: 9,600 in-orbit satellites = 75% of all controllable satellites globally, claiming priority low-orbit spectrum. The FCC has approved SpaceX satellite counts representing ~70% of all planned global LEO allocations. Late entrants face hidden spectrum-coordination barriers even after building their constellation.
Own-rocket cost moat: SpaceX deploys V2 Mini at <$1M/satellite landed (Falcon 9 ~$60M per launch, 60 satellites, ~$1M average). Competitors relying on external rockets face $3–5M+/satellite — directly impacting breakeven and margin profiles.
Network effect and global coverage: 164-country coverage gives Starlink "the only global solution" status in maritime and aviation (high-ARPU applications). On polar routes, only Starlink and OneWeb provide stable service. DTC partnerships with 30 carriers extend the distribution network further.
Competitive threats:
OneWeb (post-Eutelsat merger): 648 satellites, much higher costs, minimal market share. No own rocket. Not a near-term significant threat.
Amazon Kuiper: The only competitor worth serious attention. 3,236-satellite constellation planned; first 27 launched in 2025; commercial launch planned 2026. Backed by Amazon's e-commerce ecosystem (procurement discounts), AWS (enterprise channel synergies), and Prime (consumer distribution). But structural weakness: Kuiper relies on external rockets (ULA Vulcan, Blue Origin) — deployment cost likely 3–5× SpaceX. Kuiper's real threat is "ecosystem bundling" (Amazon offering Kuiper to its own AWS customers), not pure cost competition.
Chinese competitors (Guowang GW, Qianfan): Primarily domestic market; short-term no direct international competition with Starlink. But their large-scale constellation plans are still claiming future orbital slots.
AI板塊:最大的想象空間,最深的虧損
Anthropic 450億美元合同是AI板塊第一個可計算的商業化節點,但2028年「太空算力」願景至少還需要5年才知道答案。
核心判斷:AI是SpaceX估值的天花板,也是最大的風險來源
AI板塊在SpaceX的估值結構中佔據了一個矛盾的位置:S-1中的TAM(潛在市場規模)計算顯示,AI相關機會高達26.5兆美元(佔28.5兆總TAM的93%);但從當前財務數字看,AI是三個業務中收入最少(32.01億美元,佔17%)、虧損最大(EBITDA -12.37億、運營虧損-63.55億)的業務。
這個矛盾不是欺騙,而是「早期大型平台投資」的典型特徵——就像亞馬遜在2000年代初大量燒錢建AWS,最終AWS成為世界上最賺錢的雲計算業務之一。問題是:SpaceX的AI投資,能走出類似的路徑嗎?
xAI:從算力提供者到大語言模型的雙重身份
xAI的業務有兩個層面,性質完全不同。
算力層:Colossus超算集群
這是xAI最「硬核」的資產。Colossus 1位於田納西州孟菲斯,搭載22萬張高端GPU(15萬張H100、5萬張H200、2萬張B200),目前總算力1GW,已超越大多數科技巨頭的自有AI訓練設施。Colossus 2正在建設中,預計完成後總算力進一步擴大。
算力集群的商業模式是「租給外部客戶使用」,Anthropic的合同是第一個大規模案例。按合同條款:2026年6月至2029年5月,Anthropic每月支付12.5億美元購買算力,36個月總額450億美元。這讓Colossus每年有150億美元的確定性算力收入,相較2025年AI板塊32億美元的全部收入,這是量級的飛躍。
模型層:Grok和X生態
Grok是xAI開發的大語言模型,整合在X(推特)平台。截至2026年3月,Grok的月活用戶中約1.17億使用AI功能,佔X平台550萬月活用戶的約21%——這是一個可觀的滲透率,說明Grok在X生態中已有初步的用戶認可。
X平台本身擁有13億活躍賬戶,每日3.5億條內容,是全球最大的實時文本數據庫之一。對xAI而言,X既是Grok的訓練數據來源,也是Grok的分發渠道——每個X用戶都是潛在的Grok訂閱者。
Anthropic合同:從「燒錢」到「可計算商業化」的轉折
在Anthropic合同簽署之前,AI板塊的商業化路徑對外部觀察者幾乎不可見:GPU建設需要幾百億,但收入完全無法對沖資本開支。Anthropic合同改變了這個局面。
合同的商業邏輯:Anthropic的Claude AI模型需要大量算力訓練和推理,但Anthropic沒有能力自建與Colossus等量的算力集群。租用Colossus的GPU算力,是Anthropic在短時間內獲取大規模算力的最快路徑。對SpaceX而言,Colossus的算力使用率提升直接轉化為收入,前期投入的GPU資本開支開始見到回報。
但需要注意兩個風險:第一,Anthropic合同的真實利潤率取決於Colossus的電力成本(數據中心電費在高算力場景下可佔收入的30-50%)和GPU折舊(按5年折舊計算,22萬張GPU的年折舊可能超過100億美元);第二,Anthropic合同是「使用Colossus的計算容量」,而非固定訂閱,若Anthropic的需求下降,實際收款可能低於合同名義金額。
太空算力:最宏大的願景,最遠的時間線
SpaceX計畫在2028年開始部署「軌道AI計算衛星」——把算力放到太空,利用太空環境解決地面數據中心的兩大瓶頸:
SpaceX計畫每年向太空部署100GW算力,這需要:
這個規劃在物理上沒有矛盾,但在工程實現上至少還需要5-10年:星舰尚未開始商業運載,100GW太空算力的控溫、控制和數據回傳技術也尚未充分驗證。
這個願景是估值的「天花板想象空間」,但不是2026-2028年可以計算的財務貢獻。
X(推特):策略性資產,還是燒錢包袱?
X是AI板塊中爭議最大的組成部分。馬斯克2022年以440億美元收購推特(現改名X)後,廣告收入曾大幅下滑(主要因為大量品牌廣告主撤出),但平台的用戶規模(13億賬戶,550萬月活)和內容影響力並未顯著下降。
對SpaceX而言,持有X有兩個戰略意義:第一,X的文本數據是訓練Grok的核心語料,任何外部競爭者都需要授權費或抓取數據,而SpaceX可以直接使用;第二,X作為信息傳播平台,讓SpaceX的每次里程碑(發射成功、技術突破、IPO公告)都能獲得超乎尋常的有機傳播效果。
但從財務角度,X 2025年的廣告和訂閱收入還需要持續恢復,且平台的競爭壓力(TikTok、Instagram Threads等)持續存在。X是「有戰略價值但財務貢獻不穩定」的資產。
財務傳導:AI板塊若能在2027年EBITDA轉正(Anthropic合同收入全面計入,同時Capex峰值已過),SpaceX整體的自由現金流將顯著改善,ROIC(投入資本回報率)才能從目前的負值轉向正值,支撐1.75兆估值的財務基礎才算「落地」。反之,若AI虧損持續至2028年以後,IPO後的二級市場估值支撐將面臨壓力。
Core Assessment: AI Is Both SpaceX's Valuation Ceiling and Greatest Risk
AI occupies a contradictory position in SpaceX's valuation: the S-1 claims $26.5T in AI TAM (93% of the total $28.5T TAM); yet AI is currently the segment with the lowest revenue ($3.201B, 17%), highest losses (EBITDA -$1.237B, operating loss -$6.355B). This contradiction isn't deceptive — it's the characteristic pattern of an early-stage large-platform investment, like Amazon burning money to build AWS in the early 2000s before it became the world's most profitable cloud business. The question: can SpaceX's AI investment take a similar path?
xAI: Dual Identity — Compute Provider and LLM Developer
Compute layer: Colossus supercomputing cluster
Colossus 1 in Memphis, Tennessee houses 220,000 high-end GPUs (150K H100, 50K H200, 20K B200), totaling 1GW of compute — exceeding most Big Tech companies' proprietary AI training capacity. Colossus 2 is under construction, expanding capacity further.
The business model: rent compute to external clients. Anthropic's contract is the first major case. Terms: June 2026–May 2029, $1.25B/month for Colossus compute capacity, $45B total over 36 months. This gives Colossus $15B/year in contracted compute revenue — a quantum leap from the AI segment's total 2025 revenue of $3.2B.
Model layer: Grok and the X ecosystem
Grok is xAI's LLM integrated into X (Twitter). As of March 2026, ~117M monthly active users access AI features on X — ~21% of X's 550M MAU base. Meaningful initial penetration. X itself, with 1.3B active accounts and 350M posts/day, is one of the world's largest real-time text databases — both the training corpus and distribution channel for Grok.
The Anthropic Contract: From "Burning Cash" to "Calculable Commercialization"
Before Anthropic, AI's commercialization path was opaque: hundreds of billions in GPU capex with no offsetting revenue. Anthropic changes this.
The commercial logic: Anthropic's Claude model requires massive compute for training and inference, but Anthropic cannot self-build a Colossus-scale cluster. Renting Colossus GPU compute is the fastest path to large-scale capacity. For SpaceX, higher Colossus utilization directly becomes revenue — the front-loaded GPU capex begins to generate returns.
Two key risks: (1) Actual profit margin depends on Colossus power costs (data center electricity can run 30–50% of revenue in high-compute scenarios) and GPU depreciation (220K high-end GPUs, 5-year depreciation → annual D&A possibly exceeding $10B); (2) The Anthropic contract is for "use of Colossus compute capacity," not a fixed subscription — if Anthropic's demand drops, actual receipts could fall short of the nominal contract value.
Space Compute: The Grandest Vision, the Furthest Timeline
SpaceX plans to deploy orbital AI compute satellites beginning 2028 — putting compute in space to solve two ground data center bottlenecks:
- Cooling: Ground data centers consume massive energy cooling hardware; deep space vacuum provides near-perfect thermal dissipation conditions
- Power: Ground electricity is costly and supply-constrained; solar energy in space is nearly unlimited and free
The target: deploy 100GW of compute to orbit annually. This requires each satellite to carry ~100kW of compute, weigh ~1 ton, and Starship to achieve its 100-ton payload target — meaning thousands of Starship launches per year. Physically consistent, but engineering realization needs 5–10 more years.
X (Twitter): Strategic Asset or Cash Drain?
X is the most contested component. After Musk's $44B acquisition in 2022, advertising revenue fell sharply (brand advertisers pulled out), but the platform's user scale (1.3B accounts, 550M MAU) and real-time influence have not meaningfully declined. For SpaceX: (1) X's text data is Grok's core training corpus — zero marginal cost vs. licensing fees for any external competitor; (2) X as a media amplifier gives every SpaceX milestone organic reach that money can't easily buy. Financially, X's advertising and subscription revenue needs continued recovery, facing competitive pressure from TikTok, Instagram Threads, and others. X is "strategically valuable but financially volatile."
Financial transmission: If AI achieves EBITDA breakeven in 2027 (Anthropic contract fully recognized, capex peak passed), SpaceX's overall FCF improves significantly and ROIC finally turns positive — giving the trillion-dollar valuation its financial grounding. If AI losses persist past 2028, secondary market valuation support comes under sustained pressure.
SpaceX的錢花在哪裡,為什麼
2025年207億美元資本開支,61%流向AI——這是管理層用行動表達的判斷,也是整個估值邏輯最大的單一風險來源。
資本配置是管理層真實判斷的最誠實表達
一家公司的戰略文件可以說任何話,但資本開支的分配方式才能反映管理層真正相信什麼。SpaceX 2025年的資本開支207.37億美元,是收入186.74億美元的111%——這本身就是一個信號:公司仍處於高強度投入階段,不是在「收割」,而是在「播種」。
2025年資本開支分配:
這個排序說明:管理層判斷AI板塊的長期回報率最高,所以把最多資源投在這裡,即使它當前的EBITDA是負值。
AI:最大的賭注
127.27億美元的AI資本開支,大部分流向GPU採購(Colossus集群)和數據中心基礎設施。這個數字約等於:22萬張高端GPU的採購費用(H100約2-3萬美元/張、H200/B200更貴)加上機房建設、電力設施、冷卻系統等。
2026年AI Capex預計77億美元(YoY+201%但較2025年的127.27億下降了39%),這個下降是一個重要信號:Colossus I和Colossus II的建設基本完成後,後續算力擴充的邊際成本低於初建期。若Anthropic合同的150億美元年收入能夠從2026年下半年開始體現,AI板塊的Capex/收入比將從2025年的398%快速壓縮至2026年的50%以下——這才是業務走向成熟的財務信號。
星鏈:持續投入,鞏固護城河
41.78億美元的星鏈資本開支,主要用於:
星鏈的Capex/收入比(41.78/113.87=37%)在互聯網基礎設施公司中屬合理水準,且隨著用戶規模擴大、固定地面網絡成本攤薄,這個比例有望繼續下降。V3衛星單星下行帶寬達1Tbps(是V2 Mini的數倍),意味著同等數量的衛星可以服務更多用戶,實際上是在降低每用戶的基礎設施成本。
火箭:為未來做準備,星舰是最大投資
38.32億美元的火箭資本開支,核心是星舰量產和相關發射設施:
星舰的目標是每年1萬艘,對應「平均每天27艘」的生產速度。目前尚未達到這個規模,但Gigabay的設計產能是「每年365艘」的初始目標。這個資本投入的回收,完全依賴星舰在商業運載上的量產化和常態化。
IPO募資優先順序:管理層的公開排序
S-1招股書中,SpaceX公開了募資用途的優先順序(雖然具體金額尚未確定):
1. AI算力基礎設施擴展(最高優先)
2. 發射基礎設施和運載火箭升級
3. 衛星星座規模擴大
4. 其他一般公司用途
這個順序和2025年資本開支的分配比例(AI 61%、火箭18%、星鏈20%)高度一致,說明管理層的判斷是連貫的:AI板塊需要最多增量投入,火箭次之,星鏈的現有投入已足夠維持成長。
財務紀律的評估
SpaceX的資本配置有一個令人注意的特點:在AI板塊大量燒錢的同時,星鏈業務仍保持了63%的EBITDA利潤率,說明管理層沒有用星鏈的現金流「補貼」AI的日常運營,而是把AI和星鏈的財務結構保持相對獨立。
但整體上,公司2025年的資本開支超過收入,是在用外部融資(和IPO募資)支撐AI的建設期投入。這個模式在技術上是合理的,但財務上有一定的流動性依賴——如果IPO時間延後或市場條件惡化,資本開支計畫可能需要調整。
Capital Allocation Is the Most Honest Expression of Management's Judgment
Any company can write anything in its strategy documents — it is the actual capex allocation that reveals what management truly believes. SpaceX's 2025 capex of $20.737B exceeded its $18.674B revenue (111% ratio). The signal: this company is still in deep investment mode — not harvesting, but planting.
2025 Capex Breakdown:
| Segment | 2025 Capex | Share | Primary Use |
|---|---|---|---|
| AI (xAI+X) | $12.727B | 61% | GPU procurement (Colossus cluster), data center infrastructure |
| Starlink | $4.178B | 20% | V3 satellite production prep, global ground station expansion |
| Launch | $3.832B | 18% | Starship mass production, launch facilities (Starbase/KSC) |
This ordering says: management believes AI has the highest long-term return — even though current EBITDA is negative. Verified current return rates run in the opposite direction: Starlink Capex/Revenue = 37% vs. AI = 397%.
AI: The Biggest Bet
$12.727B AI capex = approximately 220,000 high-end GPUs (H100 ~$20–30K each, H200/B200 higher) plus data center construction, power infrastructure, and cooling systems. The 2026 AI capex is projected at $7.7B (down 39% from 2025's $12.727B) — signaling that the Colossus I and II construction peaks have likely passed. If Anthropic's $15B/year contracted compute revenue is recognized from H2 2026, the AI segment's Capex/Revenue ratio compresses from 397% in 2025 to below 50% in 2026 — the financial signal of a business maturing.
Starlink: Continued Investment, Deepening the Moat
$4.178B Starlink capex primarily funds: (1) V3 satellite mass production prep (scaled deployment via Starship from H2 2026); (2) global ground station expansion (supporting DTC and additional coverage); (3) terminal hardware R&D (smaller, cheaper user receivers). Starlink Capex/Revenue ratio (41.78/113.87 = 37%) is reasonable for an internet infrastructure company, and should decline further as user scale grows and fixed ground network costs are amortized. V3 satellites offer 1Tbps downlink bandwidth each — serving more users per satellite, effectively reducing per-user infrastructure cost.
Launch: Building for the Future, Starship Is the Core Investment
$3.832B Launch capex centers on Starship mass production and associated launch facilities: (1) Gigabay construction at Starbase (Texas) and LC-39A (Florida); (2) Raptor 3 engine mass production; (3) second launch mount (Launch Mount 2 at Starbase, used in the V3 test flight in May 2026). Starship's target: 10,000 units/year — averaging 27 per day. Current production is far below that, but Gigabay's designed initial capacity targets 365/year. Recovery of this capex depends entirely on Starship achieving commercial launch mass production.
IPO Use of Proceeds: Management's Public Priority Order
SpaceX's S-1 disclosed the priority order for IPO proceeds (specific amounts undisclosed):
- AI compute infrastructure expansion (highest priority)
- Launch infrastructure and vehicle upgrades
- Satellite constellation scale-up
- General corporate purposes
This order is highly consistent with 2025 capex allocation (AI 61%, Launch 18%, Starlink 20%) — management's conviction is coherent: AI needs the most incremental investment, Launch next, Starlink's existing investment sufficient to sustain growth.
Financial Discipline Assessment
One notable characteristic of SpaceX's capital allocation: while the AI segment burns heavily, Starlink maintains 63% EBITDA margins — management is not cross-subsidizing AI daily operations with Starlink cash flows; the two financial structures are kept relatively separate. Overall, 2025 capex exceeding revenue means the company relies on external capital (including IPO proceeds) to fund AI's construction phase. This is technically rational, but carries liquidity dependence — if IPO timing slips or market conditions deteriorate, capex plans may need adjustment.
IPO估值框架與觀察清單
1.75兆估值的合理性,依賴五個相互關聯的假設同時成立。五個觀察指標,每半年評估一次是否朝正確方向移動。
1.75-2兆估值的隱含假設
要理解SpaceX的IPO定價是否合理,必須先明確「1.75-2兆美元估值」背後隱含了什麼假設。
以2025年收入186.74億美元為基礎,1.75兆的估值對應約94倍本銷比(PS);以2026年預期收入220-240億美元計算,對應PS約63-68倍。這個倍數遠高於:
60-90倍PS的估值,通常只出現在「市場相信一家公司正在創造全新市場,且具備數倍於當前規模的長期成長空間」的科技成長股。SpaceX的招股書試圖說明這個「全新市場」是AI算力(26.5兆TAM),但AI板塊目前收入只有32億美元,且仍在虧損。
估值拆分:三層故事各值多少
依照市場通用方法,把SpaceX的三大業務分別估值:
星鏈(已驗證的現金流引擎):
火箭業務(穩定但低成長):
AI/星舰/遠期想象空間(完全基於未來):
三層相加:約1.25-1.75兆美元,這與市場給出的1.75-2兆估值基本在同一區間,說明市場的定價相對理性——但幾乎沒有為「事情比預期好」留下太多空間。
五個關鍵觀察指標
最後:如何使用這份觀察清單
SpaceX的IPO是一個「三層故事、三個時間尺度」的複合投資命題。第一層(星鏈現金流)是已驗證的,短期可觀察;第二層(星舰量產)是有技術基礎的,中期可追蹤;第三層(太空AI算力)是最宏大的,但也是最遠的,需要5-10年才能知道答案。
以上五個觀察指標,各自覆蓋了不同時間尺度的關鍵驗證節點。每半年對照最新財務數字和業務進展,評估哪一層的故事在按計畫推進、哪一層出現了偏差——這比僅僅追蹤股價更能判斷SpaceX真實的經營品質和估值合理性。
The Implied Assumptions Behind a $1.75–2T Valuation
To assess SpaceX's IPO pricing, one must first be explicit about what a $1.75–2T valuation implies. Based on 2025 revenue of $18.674B, $1.75T corresponds to ~94× Price-to-Sales (PS); on projected 2026 revenue of $22–24B, ~63–68× PS. These multiples far exceed:
- Traditional telecom companies (Comcast, Verizon: PS ~1–2×)
- Satellite communications companies (Iridium, Viasat: PS ~3–6×)
- High-growth SaaS (Snowflake, Cloudflare: PS ~20–40×)
60–90× PS typically only appears when the market believes a company is creating an entirely new market with room to grow several times its current scale. SpaceX's prospectus argues that "new market" is AI compute ($26.5T TAM) — but AI currently generates $3.2B in revenue and remains in deficit.
Sum-of-Parts Valuation: What Each Layer Is Worth
Starlink (verified cash flow engine):
- 2025 revenue $11.387B; 2026 projected ~$18B
- At 35× PS (mature satellite comms / SaaS hybrid business): ~$630B
- Rationale: 63% EBITDA margin, subscription model, global expansion potential — supports premium above standard telecom but below top SaaS
Launch business (stable, low growth):
- 2025 revenue $4.086B; 2026 projected ~$5.8B (including Starship commercial debut)
- At 25× PS (aerospace manufacturing + launch services): ~$145B
- Rationale: strong moat but low growth trajectory — infrastructure service comparable
AI / Starship / long-duration imagination (entirely future-dependent):
- Includes: Starship mass-production cost revolution improving Starlink margins; space compute (post-2028); X platform future value
- Valuation range: $500B–$1T (entirely determined by investors' beliefs about the next 10 years)
Sum: ~$1.25–1.75T — broadly in line with the market's $1.75–2T range, suggesting pricing is relatively rational. But there is almost no upside margin for "things going better than expected."
Five Key Watch Indicators
How to Use This Watchlist
SpaceX's IPO is a compound investment proposition with three layers of stories and three time horizons. Layer 1 (Starlink cash flow) is validated — observable near-term. Layer 2 (Starship mass production) has technical foundation — trackable mid-term. Layer 3 (space AI compute) is the grandest vision — but also the most distant, requiring 5–10 years for an answer.
These five indicators each cover a critical validation checkpoint across different time horizons. Evaluate every six months against the latest financial data and business progress — assessing which story layer is advancing on plan and which is showing deviation. This provides far better insight into SpaceX's true business quality and valuation rationale than simply tracking the stock price.