Bridge Capacity Priced as Forever

On June 12, SpaceX raised $85.7 billion in the largest IPO in history. Shares opened at $135 on Nasdaq under the ticker SPCS, briefly pushing the company's market value toward $3 trillion — past Amazon and Microsoft. The filing that preceded it did not describe SpaceX as a rocket company. It called the company "an AI company with space operations" and claimed a total addressable market of $28.5 trillion, with $26.5 trillion — 93 percent — attributed to AI ^[1][2]. The filing envisioned up to a million solar-powered orbital data centers and argued that the sun, containing roughly 99.8 percent of the solar system's energy, offers the only truly scalable answer to terrestrial compute constraints ^[2]. That is the permanent story. The temporary story is in the lease agreements themselves. SpaceX signed three compute-lease deals in rapid succession before the offering. Anthropic pays roughly $1.25 billion a month for Colossus 1, a Memphis installation with more than 220,000 Nvidia GPUs drawing over 300 megawatts. Google pays $920 million a month for about 110,000 GPUs on a contract running October 2026 through June 2029. Reflection AI pays $150 million a month for Colossus 2, with a 90-day termination clause ^[3][4][5]. Combined, Anthropic and Google alone generate about $2.17 billion a month, or roughly $26 billion a year — the revenue stream that turned an aerospace company into a compute infrastructure story. But SpaceX's own filing describes these deals as something less than permanent. The structure, SpaceX wrote, allows the company to monetize unused compute capacity in its infrastructure while still permitting reallocation of the capacity for its own internal initiatives if needed in the future ^[4]. Google called its $920-million-a-month agreement a short-term, timely arrangement to ensure it had bridge capacity for Gemini Enterprise while it built its own infrastructure ^[4]. Bridge capacity. Short-term. With reallocation rights retained.

The contradiction is not between bullish and bearish analysts. It is between the two framings the parties themselves put on paper: the IPO prospectus prices AI revenue as a permanent moat, while the lease contracts call it a stopgap that expires when hyperscalers finish self-building. ^[4][1]

The gap matters because hyperscalers are self-building, fast. Jefferies reports that only 8.9 gigawatts of data center capacity came online in 2025 against 21.1 gigawatts of demand — a 12-gigawatt global deficit driven by supply-chain bottlenecks, not by a permanent shift in where compute lives. The four hyperscalers plan a combined $650 billion in capital expenditure in 2026, with Nvidia projecting more than $1 trillion by 2027 and McKinsey estimating $4 trillion in total data center spending by 2030 ^[6][7]. Each gigawatt that comes online on the ground is a gigawatt SpaceX no longer bridges. The compute-lease revenue is a function of a capacity shortage that the lessees are spending hundreds of billions to eliminate. The orbital data center future that would make the revenue durable — the million solar-powered nodes in SpaceX's FCC application — is not yet operational. Launch costs run about $3,400 per kilogram and need to fall to roughly $200 per kilogram for orbital compute to be cost-competitive, a threshold Starship may eventually reach but has not ^[8]. The engineering problems are real: no convective airflow for GPU cooling, radiation-induced hardware degradation, latency penalties for real-time inference. Lloyd's of London insurers are only now beginning to model orbital AI risk, and the industry lacks the historical data to price it ^[9]. The orbital story is pre-financing, pre-insurance, pre-debt. Nor is the orbital frontier exclusively SpaceX's. Starcloud placed the first H100 in orbit in November 2025. Axiom launched compute nodes in January 2026. Nvidia has developed an orbital-specific product line. Google, while paying SpaceX $920 million a month, is also exploring launch deals with other providers for its own orbital data center ambitions under Project Suncatcher ^[8][10]. Anthropic, which leases Colossus capacity from SpaceX, has comparable compute agreements with Amazon, Google, and Microsoft — SpaceX is one of several providers, not the sole option ^[3]. What SpaceX has that none of those competitors have is a captive AI lab merged in-house. xAI was folded into SpaceX in February 2026 as SpaceXAI, creating the only entity that simultaneously owns launch capability, terrestrial compute infrastructure, and an AI model developer ^[1][3]. Rocket Lab, SpaceX's nearest publicly traded launch competitor, is running the opposite play: defense specialization, not compute. Its Q1 2026 revenue was $200 million with 63 percent year-over-year growth, and its record $2.2 billion backlog is built on Pentagon hypersonic test contracts and a Space Development Agency missile-defense constellation. CEO Peter Beck cut his own salary to $1 to fund Neutron engineering. Rocket Lab has zero compute-lease revenue and no signal of entering that market ^[11]. No hyperscaler has its own launch capability. The asymmetry is real, and for now it is unreplicable. But SpaceX is already spending the valuation that asymmetry produced. Four days after the IPO, on June 16, SpaceX announced a $60 billion all-stock acquisition of Cursor, the AI coding tool built by Anysphere. The deal used Class A shares priced near the post-IPO peak — a market capitalization of roughly $2.97 trillion — to minimize dilution. Cursor's developer ecosystem, reaching 64 percent of the Fortune 500, is to be integrated with xAI's Colossus supercomputer ^[12]. SpaceX is simultaneously preparing $20 billion in investment-grade bonds to refinance the xAI acquisition bridge loan ^[1]. The compute-inflated equity valuation is being deployed as acquisition currency and as credit support before the market has settled on whether the pricing holds. And the market has not settled. SpaceX replaced the standard 180-day post-IPO lockup with a staged share-resale system: up to 20 percent of restricted shares unlock after Q2 earnings, with additional blocks tied to stock-price triggers. The first unlocks come in August 2026. Musk and major investors agreed to 366-day restrictions, but the structure, as one analysis put it, may spread volatility across six months rather than containing it to a single event ^[13]. The dual-category valuation — aerospace plus compute — has not yet been tested by selling pressure. Some institutions are already voting against it. Danish pension fund AkademikerPension, with $25 billion in assets, blacklisted the IPO, calling any valuation above $1 trillion difficult to justify and SpaceX's governance catastrophic. A coalition including the NYC Comptroller, the NYS Comptroller, and CalPERS sent a joint letter warning that Musk's control of more than 80 percent of voting rights as CEO, CTO, and board chairman dangerously undermines investor rights ^[14]. Morningstar estimated SpaceX's value at roughly $780 billion — less than half the IPO's lower target range ^[15]. The company reported a $4.9 billion net loss in 2025 driven by AI capital expenditure, and the stock trades at over 100 times sales ^[1][7]. Analysts at the IPO noted that Amazon, Microsoft, and Alphabet provide more reasonable valuations and clearer upside than SpaceX — the same hyperscalers whose bridge-capacity payments are the revenue justifying SpaceX's premium ^[7].

Then. SpaceX's filing frames compute leases as a permanent infrastructure moat: a $28.5T TAM with 93% from AI, orbital data centers, and the language of "the only truly scalable solution" to compute demand ^[1][2]

Now. The lease contracts themselves call the same revenue short-term bridge capacity with reallocation rights, while hyperscalers spend $650B+ in 2026 to self-build the supply that would close the gap ^[4][6]

There is a version of this story where the bridge capacity becomes a permanent platform. If Starship brings launch costs down by an order of magnitude, if orbital data centers clear their engineering and insurance hurdles, if the terrestrial capacity deficit proves harder to close than hyperscaler capex suggests — then SpaceX's compute revenue migrates from bridge to backbone, and the dual-category valuation is earned, not arbitragged. Musk projected revenue exceeding $1 trillion by 2031 ^[1]. The filing's vision of a million solar-powered orbital data centers is ambitious enough to justify that trajectory if it materializes ^[2]. But that future is not what is generating the lease revenue today. Today's revenue comes from a 12-gigawatt terrestrial shortage that Google itself called a bridge, from capacity SpaceX explicitly retains the right to take back, and from infrastructure — the Colossus complex in Memphis — that is still under active construction, with a 13-acre greywater treatment plant for cooling only recently finalized ^[16]. SpaceX is pricing the bridge as the destination, and spending the difference before August.