Morgan Stanley's $300 target on SpaceX contains an arithmetic disclosure the note never makes explicit. The bank's own sum-of-the-parts values rockets and Starlink at $127 per share. It values the AI segment at $32. Sum: $159. SpaceX trades at $158.96. The target price is therefore not a statement about the launch business, the constellation, or any cash flow that exists today. It is a single-variable bet. The AI segment must revalue from $32 to $173 per share โ a 5.4x expansion โ for the target to clear. Liquidity is a mirage; solvency is the only truth. I have spent twenty-five years reading valuation notes, and the pattern is invariant: when an entire thesis collapses into one line item, that line item is not a business. It is a hope with a spreadsheet attached.
The note, circulated in early October 2026, frames SpaceX as two businesses wearing one ticker. The first is aerospace: launch cadence, Starship, Starlink. The second is "AI," a category the note never defines with precision. That omission conceals two structural anomalies that determine whether the valuation holds at all.
The first anomaly is ownership. Grok is an xAI product, not a SpaceX asset. The note merges "Grok chatbot" and "compute rental" into a single line called "SpaceX's AI business," which presupposes that xAI has been absorbed into or tightly integrated with SpaceX. Elon Musk has publicly floated such a merger. If it happened, it is the most material unreported event in the document. If it did not, the valuation rests on a factual error. The note addresses neither. That silence is the largest blind spot, and I flag it before any number is trusted.
The second anomaly is valuation basis. Compute rental is an asset business โ capital-intensive, depreciating, utilization-dependent. Grok is a model business โ talent-intensive, competitive, margin-sensitive. These two assets carry different earnings curves, different risk exposures, and different terminal values. Packing them into one $32-per-share figure and then demanding a 5.4x revaluation of the combined blob produces a number that cannot be audited, because it corresponds to no coherent cash flow. This is the same category error that ran through the 2021 tokenization wave, when projects bundled a treasury, a staking yield, and a governance token into a single "protocol value" and priced the sum as if each component carried the risk of the safest one. Emotion is a variable I exclude from the equation. What remains is a weighted average of incompatible assets.
The broader context is a bull market in which compute has become a tradeable asset class. The arrival of a per-watt quote โ dollars per watt, $/W โ is the signal that matters. It is the moment AI capacity stops being a bespoke cloud contract and becomes a commodity with a ticker-like price. Crypto has spent three years trying to tokenize exactly this: idle GPU capacity, inference markets, decentralized compute networks. Most of those tokens died because they priced a subsidy, not a service. The SpaceX note is interesting because the per-watt quote emerges from a physical operator, not a token issuance. That is what makes the number worth auditing โ and what makes its ambiguity dangerous.
Start with the pricing signal, because it is the only genuinely novel data point in the note. Compute is quoted at $30โ50 per watt on short-term transactions. The analyst's own model assumes $17.60 per watt. That gap โ roughly 2x to 3x โ is the entire ballgame. If $30โ50 is the real clearing price, the asset is worth multiples of the modeled value. If $17.60 is the long-run equilibrium and $30โ50 is a scarcity spike, the model is generous and the market is paying for a temporary premium.
My due diligence background matters here. In 2020, I audited liquidity-mining mechanisms that advertised 5,000% APY. The number was real. The yield was not sustainable, because it depended on a token subsidy that decelerated with every block. The $/W quote has the same structure. A short-term transaction price during a supply-constrained period tells you what someone will pay for immediate capacity. It does not tell you the annualized rent, the utilization rate, or the depreciation schedule. The note never specifies whether $30โ50/W is a one-time capacity sale or an annualized figure. Those two readings differ by an order of magnitude. A price without a period is not a price. It is a rumor with units.
Now stack the negative disclosures the note reports against itself. The AI segment lost $1.26 billion in a single quarter and consumed 86% of SpaceX's capital expenditure. Capital is being pulled from launch โ a business with real revenue โ into a compute buildout with no demonstrated revenue. That is not diversification. It is a cash-flow-negative division being funded by a cash-flow-positive one, and the transfer is large enough to be visible on the consolidated statement. In DeFi terms, this is a protocol using the emissions of a productive pool to subsidize a pool that has never cleared. The subsidy can run for a while. It cannot run forever.
Buyer demand is thin to the point of absence. The analyst asked 40 clients whether they held the stock. No hands. He asked who pays for AI. His own answer: buyers pay for rockets and Starlink; they pay almost nothing for AI. So the revenue side of a roughly $4 trillion implied question is currently near zero, while the cost side runs at $1.26 billion per quarter. A business with near-zero revenue and billion-dollar quarterly burn is not early-stage. It is pre-revenue with a large fixed cost base โ and fixed costs kill capital structures.
Insiders are selling into the narrative. President Gwynne Shotwell disposed of $52.5 million in stock. A tranche of roughly 328 million shares unlocked on September 24. When the people with the deepest information about a business are net sellers while a sell-side analyst publishes the highest target on the Street, the signal is not ambiguous. This is the same configuration I documented in 2017, when a fundraising vehicle's founders moved tokens to exchanges in the weeks before their distribution contract released โ and the audit trail, not the announcement, told the story. The code does not lie about who is selling. Neither does a Form 4.
Then there is dispersion. Thirty-three analysts cover the name. The average target is $235.10. The low is $142. Morgan Stanley's $300 is not the consensus; it is the outlier, and the furthest from the current price on the upside. When the bull case is a single bank's maximum, the "discount" the note cites โ "40% below large AI peers" โ needs scrutiny. That 40% figure has no stated denominator. Is it EV/Sales? EV/EBITDA? $/W times capacity? Without the base, the discount is an assertion, not a measurement. And a discount to peers is only a mispricing if the asset is comparable. Grok, by the note's own admission, is a follower: investors' top concern is whether it can keep pace with competing models. Followers trade at a discount to leaders. The market may be pricing Grok correctly, and the analyst may be mislabeling fair value as opportunity.
There is a deeper technical problem the note ignores entirely, and it sits inside my current work. An AI-driven business that rents compute and routes data through models introduces an oracle problem. The training data feeding those models has provenance, bias, and verifiability properties that no financial statement captures. I have spent the last three months auditing input pipelines for AI-driven financial modeling and found systematic bias in the data before it ever reaches the contract. A $4 trillion valuation that prices compute capacity but not data integrity is pricing the hardware and ignoring the input. I do not trust the pitch; I audit the structure. The structure has a missing column: verifiable computation.
And the physical base is undisclosed. Compute assets are priced in megawatts, GPU counts, interconnect topology, and power contracts. The note provides none of these. It gives a per-watt transaction price and an aggregate valuation, with no bridge between them. I can reverse-engineer the implied capacity if I assume a price โ but the assumption is doing all the work, which means the valuation is unfalsifiable in its current form. An unfalsifiable valuation is neither conservative nor aggressive; it is unaudited. When I reviewed NFT metadata in 2021 and found 40% of "rare" traits were algorithmically impossible, the project's defenders said the flaw was cosmetic. It was not cosmetic. It was the difference between a collection and a spreadsheet error. The same distinction applies here: an undisclosed physical base is not a detail. It is the entire claim.
The bulls are not wrong about everything. The compute-as-asset-class thesis is structurally sound. AI capacity is becoming a standardized, tradeable commodity, and the arrival of a per-watt quote is a maturity signal, not a gimmick. Every prior commodity โ oil, bandwidth, electricity โ moved from bespoke contracts to quoted units. Compute is following that path, and whoever holds the capacity holds a claim on a scarce input. SpaceX's genuine differentiator is not Grok. It is the integration of compute, energy, land, and launch โ a physical footprint that pure-model competitors cannot replicate by writing better weights. Energy and cooling, not chips, are the binding constraints on the next decade of AI, and a company that can site, power, and cool a datacenter at scale owns something real. That is a moat โ simply not the moat the target price describes.
The second concession: "40 clients, zero holdings" cuts both ways. It can mean professional money has voted with its feet. It can equally mean the institutional base is unbuilt, and any narrative improvement triggers a flow of new buyers into a thinly held name. Early is not the same as wrong. In 2017, I was early on a reentrancy flaw that the market ignored for two months before it became the reason the project failed โ and being early cost my clients their momentum. Timing is a variable; it is not truth. But the concessions do not rescue the target. They rescue the asset. The compute thesis can be right while the 5.4x revaluation is wrong, because the revaluation requires the asset to reach a scale the note never quantifies โ no megawatt figure, no GPU count, no utilization assumption, no depreciation schedule.
The finding is narrow and it is arithmetic. The $300 target is not a bet on SpaceX. It is a bet on one segment revaluing 5.4x while three of the note's own disclosures โ quarterly losses of $1.26 billion, near-zero AI revenue, and insider distribution โ point the other direction. Watch the physical base, not the target. The next earnings report either discloses compute megawatts and a per-watt revenue line, or it does not. If it does not, the 5.4x is a number without a denominator, and numbers without denominators are how the last cycle ended. The compute era is real. The question is whether this valuation is a claim on it โ or merely a claim about it.