When a crypto-native outlet leads its coverage with a sell-side price target for an unlisted aerospace company, the mismatch itself is the signal. TD Cowen initiated coverage on SpaceX with a Buy rating and a $200-per-share target this week, attributing the thesis to "AI compute leasing." No named clients. No contracted backlog. No disclosed megawatts. No revenue contribution. Just a narrative container, priced as though it were a cash flow stream.
Code is law, but people are the protocol โ and the people writing this particular protocol have incentives the headline never surfaces. I have seen this exact shape before. โ Root: DeFi Summer, when every protocol needed a "real yield" story to justify multiples the underlying cash flows could never carry. The mechanic is identical: attach a high-multiple business to a scarce asset, let the retail channel absorb it, and defer verification indefinitely. In a bull market, that deferral is free. In a bear market, it comes due.
Here is what we can actually anchor. Public reporting places SpaceX's most recent tender near $212 per share, implying a valuation around $400 billion, up from roughly $185 per share and $350 billion in late 2024. Against a share count of about 1.89 billion, a $200 target implies roughly $378 billion. If the note postdates the mid-2025 round, the "Buy" sits below the last cleared price โ a negative implied upside dressed as enthusiasm. If it predates it, the upside is a modest single-digit percentage. Either way, the AI compute leasing story is not a growth engine. It is an explanation for why the current price is already fair.
That distinction matters more in a drawdown than any target number. โ Root: The 2022 Bear Market, when I watched tokens with immaculate narratives bleed 90 percent while their communities insisted the fundamentals were unchanged. In a downturn, narratives stop being leverage and start being liabilities. The reader's real question is not "how high" but "what breaks first." For a private, thinly disclosed asset wrapped in a retail vehicle, the answer is almost always the retail holder.
The venue choice is not incidental. Crypto media readers are the precise audience for SPV distribution of pre-IPO SpaceX shares, and "compute leasing" already has tribal recognition here from the miner-to-AI pivot โ IREN, Cipher, TeraWulf, Hut 8.
So let us test the physics, because orbital compute is a physics problem before it is a business model. Vacuum cooling happens only by radiation, governed by the Stefan-Boltzmann relation. At 340 Kelvin with an emissivity near 0.9, a square meter radiates roughly 682 watts; at 300 Kelvin, only 413. That means every kilowatt of orbital compute needs 1.5 to 2.4 square meters of radiator, and a single megawatt module needs 1,500 to 2,500 square meters of heat-rejection surface. For scale, the International Space Station's external thermal control system spans about 360 square meters to reject on the order of 70 to 100 kilowatts.
Now hold that against a ground-based AI rack. A single GB200 NVL72 cabinet draws roughly 120 kilowatts. The "one megawatt" unit that orbital advocates invoke is, in data center terms, about eight racks. You are spending a hectare of radiator to replicate a small server room.
Power generation is the easier half, and it is still area-bound. The solar constant is 1,361 watts per square meter; at 30 percent efficiency that is about 400 watts per square meter, so a megawatt of continuous draw needs roughly 2,500 square meters of solar array. A dawn-dusk sun-synchronous orbit would give near-continuous illumination and sidestep storage โ which is precisely why any serious proposal would choose it, and precisely the engineering detail the coverage omits.
Then there is the veto. Inter-satellite laser links run in the hundreds of gigabits per second per link, aggregating to terabits per satellite. Training workloads require hundreds of gigabits per node of dedicated collective communication for gradient synchronization. Orbital clusters cannot meet that bar. Training is architecturally impossible in orbit; only inference โ especially batched output and downlink of results โ is viable.
Launch cost is the only variable that could masquerade as a moat, and it is a function of Starship, not of today's fleet. Falcon 9 runs near $2,700 per kilogram; Starship targets $100 to $200. A one-megawatt orbital module โ radiators, arrays, compute payload, structure โ lands around 25 to 40 tonnes. That is roughly $68 to $108 million on Falcon 9, or $3 to $8 million on Starship at target. A ground-based megawatt AI data center costs $30 to $50 million in capex. The entire orbital arbitrage rests on Starship delivering its promised unit economics, and even then the orbital path carries a two-to-five-times capex premium. That only clears where grid power is unavailable, sovereignty is mandated, or defense demand is inelastic.
Consider what it would take to actually restructure SpaceX's revenue. The company runs roughly $13 to $15 billion annually today. To move the needle at 20 percent, a new line needs $3 to $5 billion a year. At industry benchmarks โ one megawatt approximating a thousand H100s, contracted at $1.50 to $2 per GPU-hour and 70 percent utilization โ a megawatt yields about $9 to $12 million a year. That implies 250 to 330 megawatts for the low case and 420 to 550 for the high. CoreWeave, the reference neocloud, ended 2024 near 250 megawatts of active power and targets over a gigawatt. SpaceX would have to out-deploy an entire leading neocloud's existing capacity in one move. Orbitally, 330 megawatts needs about 50 hectares of radiator and a hundred-plus Starship flights. That is not a five-year plan.
Which brings us to what the narrative probably means. The only orbital AI load with near-term commercial viability is on-board inference for Earth observation โ processing remote sensing data in orbit and downlinking conclusions rather than raw pixels, cutting bandwidth needs by two to three orders of magnitude. That aligns exactly with Starshield and government sensing contracts. If TD Cowen's "AI compute leasing" refers to a terrestrial GPU rental business, the differentiation is thin: no power advantage, no land advantage, no cloud-native operating stack, no developer ecosystem. If it refers to xAI coordination, we are looking at a related-party transaction โ and the accounting gets interesting.
Here is the blind spot the coverage never names. A sell-side initiation on a private company the market broadly expects to IPO is not neutral analysis. It is pre-mandate positioning โ an investment bank building goodwill ahead of a potential underwriting role. The rating's independence should be discounted, especially when its bull case rests on a business line that discloses no financials at all. The real news may not be the $200 target. It may be that coverage exists at all โ a signal that a liquidity event is approaching.
And then there is the retail channel. The crypto audience reaches SpaceX exposure through special purpose vehicles, and the fee stack is brutal: five to ten percent up front, one to three percent annually, ten to twenty percent carried interest. If the underlying appreciates 15 percent a year, the SPV investor nets five to eight. The "Buy" rating, after that friction, is closer to a hold โ or a loss.
Governance isn't a line item on a term sheet, but this is a governance problem. When the sell-side prices a private asset and the retail wrapper taxes the entry, the constituency absorbing the risk has the least information. That is the same asymmetry I audited during DeFi Summer, and the same one that emptied retail wallets in 2022. We didn't need a crash to see it. We needed a fee schedule.
The signal worth tracking is not the target price. It is whether SpaceX hires compute and cloud infrastructure teams, whether TD Cowen appears on a future underwriting syndicate, and whether any named customer ever signs. Watch the Starship cost curve โ it is the single independent variable underwriting the entire orbital thesis. Until one of those resolves, treat AI compute leasing as option value, not cash flow. The bear market does not reward the loudest story. It rewards the one that survives verification.