In the final week of September, a pair of numbers surfaced in a sell-side note that should have stopped every compute trader cold. Short-term AI capacity was clearing at $30 to $50 per watt. The same analyst who recorded those trades built his long-term model on $17.60 per watt. No reconciliation. No explanation. Just a 1.7x to 2.8x spread parked in the middle of a $420 billion valuation.
I have audited enough valuation decks to recognize what a spread like that means. It is not a rounding error. It is an admission that nobody — seller, buyer, or analyst — actually agrees on what a watt of AI compute is worth. And when an asset class cannot settle its own unit price, the number stapled to it stops being a measurement and becomes a story. That story is what I want to take apart.
The frame being sold is clean. SpaceX, freshly public, is no longer merely a rocket company; its AI division now devours 86% of capital expenditure. The bull case — a $300 price target — rests on the claim that the market misprices this pivot, that the AI business trades "about 40% cheaper than large AI peers." It is a familiar sales pitch, and I have watched it before.
From the ashes of 2017 to the fluidity of DeFi, every cycle mints a new primitive that everyone agrees is "the future" long before anyone agrees how to price it. In 2017 it was the ERC-20 token. In 2020 the liquidity pool. In 2021 the NFT floor. Each was real. Each was also, for a window, priced entirely by narrative — until unit economics arrived and sorted the survivors from the corpses. What is proposed here is the same move, one layer down the stack: compute itself, denominated per watt, sold as a tradeable asset. DePIN projects have spent three years trying to tokenize exactly this — physical infrastructure, metered and priced in real time. SpaceX is simply proposing the centralized version, with a balance sheet instead of a blockchain.
Here is the arithmetic the note left unsaid. The sum-of-parts almost exactly equals today's price. Rockets plus Starlink: $127 per share. AI: $32 per share. Total: $159. The stock trades at $158.96. The market has already accepted the analyst's present valuation. There is no disagreement about the now.
The disagreement is entirely about the future, and the target price conceals how extreme that future must be. Strip the $127 of rocket-and-Starlink value from the $300 target and you are left with $173 per share of AI. Against today's $32, that is a 5.4x revaluation of a single division — roughly $2.28 trillion of implied value. The "doubling" headline is not a broad repricing. It is one bet, on one business line, at five times the current mark. That is not a re-rating. It is a leap of faith expressed in decimal places.
And the line itself is the problem. "AI business" bundles two assets with nothing in common. There is Grok, a model — a margin-compression business that lives or dies on benchmarks, developer adoption, and enterprise contracts. And there is compute rental — an asset business, priced per watt, closer to industrial real estate than to software. Lumping them under one multiple is not consolidation; it is concealment. The two carry different revenue curves, different depreciation schedules, and different failure modes. An analyst who values a software margin and a power plant at the same multiple has told you which of the two he understands.
The compute-rental side is where the innovation actually sits. Pricing capacity by the watt is the moment AI infrastructure stops being a cloud service and becomes a commodity — the same transition crypto made when tokens turned block space into a market. It is also why the $30–50 short-term prints matter so much. If those trades reflect genuine scarcity, they are a floor. If they reflect a supply squeeze that cools the moment demand softens, they are a narrative floor of the kind that collapsed the blue-chip NFT market — a label that held until liquidity vanished and revealed that, underneath, nothing had ever been bid. I have watched a six-figure floor become a rumor in a single weekend. Compute is not immune to the same gravity.
The model side is where the anxiety lives. Investors' top stated fear is blunt: whether Grok can keep pace with rival models. That is not a technical footnote. A laggard does not earn a leader's multiple, and a "40% discount" to peers may be the market pricing competitive reality rather than an error to be arbitraged. The note offers no benchmark, no API volume, no user count — a data vacuum that reads less like modesty and more like omission.
Then there is the question nobody asks: who buys the watts? Customer concentration is the quiet killer of every infrastructure thesis. If compute rental depends on three or four frontier labs, the revenue is as fragile as a single frozen counterparty — the lesson stablecoin holders learn when a compliance switch is flipped and an address goes dark. Diversified buyers mean a business. A handful of whales mean a hostage situation. A wholesale compute market with four buyers is not a market; it is a negotiation, renewed every quarter, where the price of a watt is whatever the largest tenant decides it is worth.
And the pricing itself has a clock. Infrastructure capacity rarely stays scarce or abundant for long. Post-Dencun blob space was supposed to be cheap for years; it is already being rationed, with rollup fees set to climb again. Compute may follow the same arc — today's $30–50 could be tomorrow's squeeze or tomorrow's glut, and the model's $17.60 assumption may be the only sober number in the room.
The negatives cluster tightly around the bull case. The AI division lost $1.26 billion last quarter. The analyst himself concedes that buyers currently pay for rockets and Starlink, not for AI. He asked forty clients whether they held the stock; not one raised a hand. On September 24, roughly 328 million shares unlocked. The president sold $52.5 million of stock — into a bullish note.
Here is the contrarian read, and it cuts against bulls and bears at once. Grok is xAI's product, not SpaceX's. To call compute rental plus Grok a "SpaceX AI business" presumes xAI has merged into, or been folded under, SpaceX — a possibility Musk has floated but never confirmed. If that assumption holds, it is the single largest undisclosed event in the story. If it does not, the valuation rests on a factual error. Either way, the note never says a word. And if the merger is real, the competitive set shifts entirely: the rival is no longer a frontier lab but AWS, Azure, and CoreWeave — wholesale compute landlords with their own power contracts and depreciation curves.
The "forty clients, zero holdings" signal is equally double-edged. Bulls read it as untapped demand. I read it as professional money voting with its feet while the retail narrative does the buying — the oldest structure in every cycle I have covered.
So watch three signals. Starship Flight 15, expected late October into early November, is the nearest catalyst — a successful ship catch is the strongest near-term event. The third-quarter report will show whether the AI division's losses are widening or flattening, and whether compute revenue appears at all. And every new $/W transaction is a referendum on whether $30–50 is an anchor or a spike. A $420 billion valuation that carries no regulatory risk premium is not conservative — it is optimistic by omission.
The deeper question is not whether SpaceX doubles. It is whether compute has truly become an asset class — or merely a narrative wearing an asset class's clothes. From the ashes of 2017 to the fluidity of DeFi, I have learned the difference is only ever revealed after the liquidity leaves. When the last watt is unsold and the last buyer is gone, we will finally learn what a watt was worth all along.