Hook: The number that doesn't reconcile
TD Cowen initiated SpaceX at Buy with a $200 price target, citing AI compute leasing as the driver. Run the arithmetic before reading the thesis. Public tender references put SpaceX at roughly $185/share in late 2024 (~$350B) and around $212/share by mid-2025 (~$400B), implying about 1.89 billion shares outstanding. At $200 per share, the implied valuation is approximately $378 billion. If the note postdates the mid-2025 tender, the "Buy" rating carries negative upside of roughly six percent against the last observable clearing price. That is not a growth call. It is a narrative explanation of an existing price.
The deeper problem is what "AI compute leasing" means physically. Four candidate interpretations exist: orbital GPU clusters, terrestrial GPU rental, government/defense AI contracts, and xAI-related-party demand. The valuation delta between these scenarios spans orders of magnitude. The report apparently discloses none of them. So the correct first move is not to model revenue. It is to test whether the most exotic version of the story survives first-principles physics.
It does not survive, for training workloads. The math doesn't lie, and it doesn't round up either.
Context: what an AI compute lease actually is
A compute lease is simple machinery. A provider operates racks of accelerators, sells capacity by GPU-hour or by reserved megawatt, and books revenue against utilization. The economics hinge on four variables: power secured, chips deployed, cooling deployed, and interconnect bandwidth available for the workload. Terrestrial neoclouds like CoreWeave, Nebius, IREN, and Oracle OCI compete on exactly these variables, plus contract backlog.
As of late 2024, CoreWeave operated on the order of 250 MW of active capacity with a stated path beyond 1 GW. A single NVIDIA GB200 NVL72 rack draws approximately 120 kW. One megawatt is therefore roughly eight racks, or about a thousand H100-class accelerators with supporting infrastructure. At contract rates near $1.5 to $2.0 per GPU-hour and 70 percent utilization, one MW generates roughly $9 to $12 million annually.
Now reverse-engineer the phrase "restructure SpaceX's revenue." SpaceX's 2024-2025 revenue sits in the $13 to $15.5 billion range, with Starlink contributing an estimated $8 to $10 billion and launch services $4 to $5 billion. To move the revenue mix by 20 percent, a new line needs $3 to $5 billion per year. At unit economics above, that requires roughly 250 to 550 MW of deployed capacity. CoreWeave spent years and billions in GPU-backed debt to approach that scale. SpaceX has disclosed zero.
The report contains no customers, no backlog, no pricing model, no timeline, no capital expenditure plan. For a private company, that absence is not neutral. It is either non-disclosure by necessity (contracts not signed or not quotable) or omission by the aggregator. Both readings weaken the note as evidence and strengthen it as signal.
Core: the physics of orbital compute
Thermal hard wall. Vacuum permits radiation only. The Stefan-Boltzmann relation gives radiated power as P = epsilon x sigma x A x T^4. With emissivity around 0.9 and a radiator temperature of 340 K, flux is roughly 682 W per square meter. At 300 K it drops to about 413 W per square meter. Therefore one kilowatt of orbital compute requires 1.5 to 2.4 square meters of radiator area. One megawatt requires 1,500 to 2,500 square meters.
Cross-check against flown hardware: the International Space Station's external thermal control system uses roughly 360 square meters of radiators to reject on the order of 70 to 100 kW. The orders of magnitude agree. Terrestrial data centers reject kilowatts to megawatts per square meter of floor area through chilled water and direct-to-chip liquid loops. Orbital and terrestrial thermal rejection differ by three to four orders of magnitude in areal density. That gap does not close with engineering iteration. It is thermodynamics.
Power is the easier half, but still an area problem. Solar constant is 1,361 W per square meter. At 30 percent cell efficiency, that yields about 400 W per square meter of array. Continuous one-megawatt supply needs roughly 2,500 square meters of wings. Dawn-dusk sun-synchronous orbits in low Earth orbit approach continuous illumination and eliminate eclipse battery mass, which is why any serious design converges there. That choice is verifiable and should be demanded before any revenue model is credited.
Bandwidth vetoes training outright. Public disclosures on Starlink inter-satellite laser links place single-link rates in the 100 to 200 Gbps range, with per-satellite aggregation in the single-digit Tbps range. Tensor parallelism and pipeline parallelism inside a training cluster require intra-rack NVLink at TB/s scale and dedicated 400 to 800 Gbps collective communication between racks for gradient synchronization. An orbital cluster cannot sustain this. Training is architecturally impossible, not merely expensive. Inference is different: batch outputs and result downlink fit within available bandwidth, especially if only conclusions rather than raw sensor data return to ground.
Launch cost is the only real moat, and it is a Starship function. Falcon 9 costs roughly $2,700 per kilogram. Starship targets $100 to 200 per kilogram. A one-megawatt orbital module (radiators at 7.5 t, arrays near 5 t, payload and structure near 10 t) masses approximately 25 to 40 tonnes. Falcon 9 launch runs about $68 to 108 million per module; Starship at target pricing runs $3 to 8 million. The entire orbital arbitrage therefore prices Starship's target, not its realized cost. Meanwhile, terrestrial one-megawatt AI capacity carries total capex near $30 to $50 million. Orbital solutions absorb a 3 to 5 times capex premium before launch economics normalize.
Maturity is proof-of-concept. Publicly tracked in-orbit AI compute consists of single-satellite demonstrations: Starcloud-1 carrying H100-class hardware launched as rideshare in late 2025, and Google's Project Suncatcher targeting two experimental satellites in early 2027. No commercial orbital cluster exists. Pricing a $200 target partly on orbital leasing means pricing proof-of-concept hardware as production assets.
The workload that actually fits. The only near-term commercially defensible orbital AI workload is Earth-observation inference: process imagery on orbit, downlink conclusions, cut downlink demand by two to three orders of magnitude. This aligns with Starshield and NRO/SDA contract lines. The realistic near-term reading of "AI compute leasing" may well be defense-adjacent orbital inference, priced as a government contract stream rather than a commercial cloud business. Cost-plus government pricing caps margin. That does not support a high-multiple growth story.
Unit economics of the terrestrial alternative. If the phrase means terrestrial GPU rental from Texas or California sites, SpaceX brings interconnection experience, capital, and government relationships. It brings no exclusive power contracts, no cloud sales organization, no Kubernetes and InfiniBand operations bench, no developer ecosystem, and no existing customers. It enters a market where CoreWeave and Nebius hold multi-year hyperscaler contracts. New entrants compete with capital subsidies, and margins degrade accordingly. Starlink's consumer and SMB subscriber base has zero overlap with frontier training buyers. There is no data flywheel in renting GPUs.
Contrarian: three blind spots the note does not price
First, the price target may be a floor, not a ceiling. Against a $212 tender print, $200 implies negative upside. In that framing, AI compute leasing functions as retroactive justification rather than forward catalyst. Sell-side coverage of private companies before a liquidity event typically follows narrative augmentation logic: a new high-multiple story is required to defend an already-high price. The report's real news value is not the target. It is that a bank initiated coverage on a private company at all, which is an indirect IPO-mandate positioning signal.
Second, related-party demand can manufacture circular validation. xAI requires massive training and inference capacity. If xAI becomes the anchor customer for SpaceX compute, the revenue is internal to the Musk entity complex, and both valuations lift on the same synthetic print. Private companies have no SEC disclosure obligation for related-party concentration. Investors taking SPV exposure inherit an structural information disadvantage the article never mentions. Watch for any disclosed xAI-SpaceX compute agreements. That is the red flag for circular valuation.
Third, retail participation costs silently consume the target. Crypto-media readers access SpaceX through SPV and offshore vehicles carrying 5 to 10 percent upfront fees, 1 to 3 percent annual management fees, and 10 to 20 percent carry. Even if the underlying appreciates 15 percent annually, net investor return lands near 5 to 8 percent. A "Buy" at $200 under that fee stack is effectively a market-weight position. Liquidity is an illusion until it is priced, and the fee-adjusted price is never shown on the tender sheet.
One more structural point from my own audit work: when a report presents a thesis without customers, backlog, or a single revenue line, the absence is data. In 2021 I traced Aave V2's liquidation path because documentation claimed mitigation the code did not enforce. Documentation without contract calls is a claim, not a fact. Here, the claim is that AI compute restructures SpaceX revenue. The evidence is one sentence and no numbers.
Takeaway: signals that resolve the ambiguity
The decisive variable is the publication date; the second is the business form. Before treating this as investable, resolve two questions: does TD Cowen later appear in a SpaceX underwriting syndicate, and does SpaceX post hiring for GPU cluster operations, enterprise sales, or power procurement? Hiring is the earliest observable proof of a real leasing business. Named customers or contract values would follow within quarters. Continued silence means the narrative decays toward Starlink and launch fundamentals as the only defensible valuation anchors.
If orbital compute advances, the bottlenecks are not chips. They are space-grade solar arrays, thermal radiators, inter-satellite laser links, and radiation-hardened electronics. Track orders in those segments, not press releases. And watch Starship's realized cost-per-kilogram curve: it is the single independent variable on which the entire orbital compute argument depends. Until that curve prints, the honest position is that AI compute leasing is an option, priced at zero in any rigorous model, and the $200 target is a description of yesterday's private-market price dressed as tomorrow's thesis.
