A $300 price target on a $2.1 trillion company reads like conviction. Run the arithmetic and it reads like leverage.
SpaceX trades at $158.96. Morgan Stanley's Jonas wants it at $300. Between those two numbers sits the entire bull case, and it does not spread across the business. It pools in a single line item: AI.
The headline skipped the math. Jonas's sum-of-the-parts assigns $127 per share to rockets and Starlink โ a $1.68 trillion valuation โ and $32 per share to AI, roughly $422 billion. Add them and you get $159. That is the current share price to the cent. Which means the full $141 per share of upside in the $300 target has to come from AI. The AI unit would need to be worth $173 per share, or about $2.28 trillion โ a 5.4x revaluation of a division that lost $1.26 billion last quarter.
The report calls the stock undervalued. The arithmetic calls it a single-asset wager wearing a portfolio costume. That gap is the trade.
Context: what the report says, and what it never does
Set the baseline first. The report, circulated through a Walter Bloomberg post, argues SpaceX is worth double its current price. It cites a $300 target, a current price of $158.96, and a share count near 13.2 billion. It flags a valuation discount of roughly 40% versus "large AI peers." It lists three catalysts: a new Grok release, additional compute deals, and a Starship flight that could bring a returning capsule home.
The structure of the argument matters more than the number. Jonas does not claim the whole company is cheap. He claims the parts are cheap and that the market will eventually re-rate them. Rockets and Starlink, in his framing, already justify today's price. AI is the option on top. That is a defensible construction. It is also a construction that puts every dollar of expected return into the asset with the least operating history and the most competitive exposure.
Now the omission. Grok is an xAI product. It is not, and has never been, a SpaceX business. The report folds "Grok chatbot plus compute rental" into one "SpaceX AI" bucket without stating the assumption that xAI has merged into, or been absorbed by, SpaceX. Musk has floated that combination in public before. If it happened, it is the largest structural event in the document and it goes unmentioned. If it did not, the entire valuation rests on a category error โ pricing a product the company does not own.
Either way, the report treats two economically opposite assets as one. Compute rental is an asset business: capex-heavy, revenue-predictable, depreciation-bound. A chatbot is a model business: margin-scalable, retention-dependent, competition-exposed. Bundling them under one multiple is not analysis. It is an averaging that hides both. The asset business looks more stable than it is. The model business looks more valuable than it is. A blended multiple is a place where two bad assumptions go to hide behind each other.
I audited reserve disclosures across five exchanges during the MiCA compliance race. The first lesson is that a blended number is where accountability goes to die. The second is that whoever does the blending usually benefits from it. Speed is the only currency that never depreciates, and this report spent its speed on a headline instead of a breakdown.
Core: the $/W paradigm and the crypto mirror nobody connected
Here is the part of the report with real information gain. Compute is being priced by the watt.
Short-term compute trades are clearing at $30 to $50 per watt. The analyst's model assumes $17.60 per watt. That spread โ roughly two to three times โ is the single most important number in the document, and it is buried beneath the valuation noise.
Why it matters. A watt of data center capacity is the lowest common denominator of AI. You are not buying a model, a framework, or a team. You are buying delivered electricity to silicon, plus the cooling and interconnect that make it usable. Strip a data center to its physics and you get dollars per watt. Once you can price a thing per watt, you can trade it per watt. Once you can trade it, you have a commodity.
This is the moment a resource becomes an asset class. Oil did it. Bandwidth did it. Carbon did it. Compute is doing it now. The $30โ$50 clearing price is not a marketing figure. It is a market discovering a unit of account.
The report never specifies whether $30โ$50 per watt is a one-time capacity sale or an annualized rental. That distinction is worth an order of magnitude. A one-time $40/watt sale on a gigawatt-scale build is a few billion dollars of transaction. An annualized $40/watt rental on the same build is a revenue stream that would justify an entirely different enterprise value. The report uses the number without defining it. An undefined metric cannot be verified and cannot be falsified. That is not a data point. That is a mood.
Take the metric seriously anyway, and a second market appears. The retail-facing mirror of wholesale compute is DePIN โ decentralized physical infrastructure networks that tokenize GPU capacity. Protocols that let holders rent out compute and earn yield are, structurally, the same trade as SpaceX renting watts to AI labs. Both price delivered compute. Both depend on the persistence of demand. They differ in liquidity and leverage.
When wholesale $/W is high, tokenized compute yields look attractive and capital rotates into DePIN tokens. When wholesale $/W normalizes โ and commodity prices always normalize โ those tokens reprice, because their yield was never a yield. It was a pass-through of a scarcity premium. The edge lies in the data others ignore, and the data here is the spread between the wholesale clearing price and the tokenized yield. That spread is the arbitrage, and almost nobody is watching it. Every DePIN yield farmer is long the same hidden variable: the spot watt.
Run the reverse calculation on SpaceX and the scale problem sharpens. If $30โ$50 per watt is real, then a $422 billion AI valuation implies a compute footprint the report never discloses. No megawatts. No GPU count. No interconnect topology. No power purchase agreements. A valuation built on a per-watt price, with no wattage attached, is a formula missing its variable. You cannot audit what was never stated. In my IBIT arbitrage work, the entire thesis lived in a 0.4% gap between the fund and spot, and it only held because the underlying shares were real and countable. Here, the underlying is invisible.
Now the demand side, and this is where the report indicts itself. Jonas asked forty clients whether they held the stock. Nobody raised a hand. The report frames this as untapped potential โ a vast base of non-holders waiting to be converted. There is a second reading, and it is stronger. Forty professional investors, handed the chance to own the most famous AI story on the tape, chose not to. That is not a pipeline. That is a verdict.
Combine it with the internal signals and the picture tightens. On September 24, roughly 328 million shares unlocked. President Gwynne Shotwell sold $52.5 million of stock. Insiders sell for many reasons, but the pattern โ unlock, insider sale, analyst upgrade, thin institutional ownership โ is a known structure. The informed money exits into the narrative. Retail arrives to hold the position. I watched the same choreography in the 2021 NFT floor collapse, when blue-chip collections held their prices right up until liquidity vanished, and then held nothing. The floor was never a floor. It was a bid.
The AI division's financials complete the case. A $1.26 billion quarterly loss. Eighty-six percent of SpaceX capital expenditure flowing into AI. This is not a business scaling into profit. It is a cash-flow-negative infrastructure build financed by a launch-and-broadband cash cow. That can work, for a while. It cannot work indefinitely, and the report offers no breakeven timeline, no utilization rate, and no customer concentration data. A compute landlord with one or two tenants is not a business. It is a counterparty risk with a roof.
The competitive picture is equally thin. The report's own top-listed investor concern is whether Grok can keep pace with rival models. That is the report admitting, inside its risk section, that the model business is a follower. GPT, Gemini, Claude, and Llama hold advantages in capability, developer ecosystem, or enterprise distribution. A follower does not earn a leader's multiple. It earns a discount, and the market is applying one.
Which brings the discount claim into question. Jonas says SpaceX trades about 40% below large AI peers. He does not say whether that is EV/Sales, EV/EBITDA, or dollars per watt. Without the denominator, "40% cheap" is unfalsifiable. And if Grok is genuinely behind, the discount may not be a mispricing at all. It may be the market pricing a follower correctly.
Look at the dispersion. Thirty-three analysts cover the name. The average target is $235.10. The low is $142. Jonas sits at $300 โ the top of the range, an outlier by nearly 28% over the mean. When one analyst is the loudest voice and also the most extreme, the target is a positioning tool, not a consensus. Resilience is built in the quiet before the crash, and the quietest thing in this report is the absence of a valuation methodology.
There is a final layer, and it is the one I care about most as a surveillance analyst. The report's framing of "AI" ignores that AI agents are becoming market participants in their own right. In mid-2026, I modeled a surveillance tool to track AI-generated wallet clusters, because autonomous agents were already driving a meaningful share of on-chain volume. When compute is priced per watt and rented by autonomous systems, the buyer is increasingly not a person. It is a machine with a budget and no disclosure obligation. A valuation that treats compute demand as a human enterprise, without modeling agent-driven volume, is missing the fastest-growing and least-transparent demand segment in the market. That segment does not fill out a client questionnaire. It does not raise its hand. It simply consumes watts and moves on.
Contrarian: the discount is not the error โ the blend is
Everyone will argue about whether SpaceX is cheap. That is the wrong question. The error sits upstream, in how the AI unit is assembled.
Here is the contrarian read. The market is not mispricing SpaceX AI. The market is correctly refusing to pay a leader's multiple for a follower, and correctly discounting an asset business whose clearing price may be a bubble. The "40% discount" is the market's way of saying it has seen the $/W spot print and does not believe it is durable. And it is probably right. $30โ$50 per watt looks less like a floor and more like a scarcity premium โ the price you pay when compute is short and everyone is building. Commodity spot prices in shortage are not anchors. They are spikes. The analyst modeling $17.60 per watt is not being conservative. He is modeling the long-run equilibrium, and the report quietly labels that equilibrium "undervalued."
If $/W normalizes toward the model, the revaluation math inverts. The 5.4x AI revaluation the $300 target requires becomes a de-rating, not a re-rating. And because tokenized compute tokens are levered to the same clearing price, the contagion runs straight into crypto. DePIN yields compress. Token prices follow. Retail holders who bought compute yield at the top of the scarcity cycle discover they bought the spike.
There is one more blind spot, and it is structural. The report assigns no regulatory risk to the AI valuation. For an entity operating in the United States, tied to the X platform, and renting compute to enterprise clients, that is a striking omission. Cross-border data transfer, sector-specific access rules, and model-reporting thresholds are all live compliance questions. In my MiCA audit, the exchanges that failed were never the ones with the worst technology. They were the ones that never priced the compliance line until it priced them. A $2.28 trillion implied valuation with a zero-dollar compliance provision is not optimistic. It is unmodeled. If xAI sits inside SpaceX, the compliance surface widens further โ data residency for rented compute, content liability through the X integration, and disclosure obligations on large training runs all land on the same balance sheet.
The bear market sharpens every one of these points. In a market where survival matters more than gains, the question is not "how much can this re-rate." It is "what happens to this cash flow if it does not." A division burning $1.26 billion a quarter, funded by a cash cow, is fine until the cash cow stumbles. That is the whole risk. Not the target price. The runway.
Takeaway: what to watch, and what it will tell you
The trade is not SpaceX. The trade is the watt.
Watch four signals. First, Starship Flight 15, expected late October into early November, with a returning-capsule catch as the strongest near-term catalyst. Second, the Q3 report, specifically the first disclosure of compute revenue and the AI division's loss trend. Third, any new compute deal with a stated $/W figure โ that single number will reveal whether $30โ$50 is a market or a moment. Fourth, Grok's next release and its third-party benchmark position, which will confirm or kill the follower thesis.
If the $/W spot price holds above the model, compute is genuinely becoming an asset class, and the entire tokenized-compute sector gets repriced upward with it. If it slides toward $17.60, the SpaceX AI story is a scarcity spike in disguise, and the same math will unwind every DePIN yield that borrowed its price.
Chaos is just data waiting for a pattern. The pattern here is simple. One number โ dollars per watt โ now prices a rocket company's future, a chatbot's relevance, and a thousand tokenized compute tokens at once. Watch that number, not the target. The target is a headline. The watt is the trade.