There is a number buried in the analyst coverage of SpaceX's AI business that almost nobody quoted, and it is not the $300 price target. It is thirty to fifty dollars per watt. That is the short-term trading price attached to AI compute capacity in the note, set against a modeled long-run assumption of $17.60 per watt. On the surface this is a footnote about a rocket company's side project. Underneath it is the first credible attempt I have seen to price artificial intelligence the way we price oil, copper, or office space โ by the unit, in the open, with a number you can argue about. And that should unsettle every decentralized compute protocol still printing a token to pay for GPU hours. When a legacy institution starts quoting dollars per watt, it is quietly telling the market that compute has become a commodity โ and commodities do not need your token.
Let me lay out the structure before I take it apart, because the structure is the whole story. As reported, SpaceX's AI arm โ a bundle of the Grok chatbot and a compute-rental business โ carries a valuation near $420 billion. The bull case, attributed to Morgan Stanley's Adam Jonas, splits the company into rockets-plus-Starlink at $127 per share and AI at $32 per share. That sums to $159, which is, almost to the cent, the current share price of $158.96. The $300 target therefore requires the AI segment alone to revalue roughly 5.4 times, from $32 to about $173 per share โ call it $2.28 trillion of implied AI value. The report itself lists the counter-evidence: the AI division lost $1.26 billion last quarter, consumed 86% of SpaceX capital expenditure, buyers are described as paying for rockets and Starlink rather than AI, 328 million shares unlocked in late September, and the company president sold $52.5 million of stock into the analyst's optimism.
I want to be honest about my priors. My knowledge cutoff predates this entire timeline, and I cannot independently verify the IPO, the report, or the tweet that amplified it. So treat the specifics as a scenario and the industrial logic as the thing worth stress-testing. The industrial logic is what connects to crypto, because this industry has spent six years building exactly the market this note describes โ a market for compute sold by the unit โ and it has been bleeding the entire time. That gap is the story.
Here is why. Crypto's decentralized compute sector โ Render, Akash, io.net, Gensyn, and a dozen smaller DePIN networks โ was built on a thesis that looks, in this bear market, eerily like the SpaceX note. The thesis: compute is scarce, GPUs are the bottleneck, and a permissionless marketplace can route idle capacity to buyers cheaper than the hyperscalers. Token holders would capture the spread. The token would be the settlement layer. Six years and one brutal cycle later, most of those tokens are down 70% to 95% from their peaks, and the reason is not that the thesis was wrong. The reason is that the thesis was right and the token was unnecessary. The market discovered it could buy the watt without buying the wrapper. That single sentence is the most expensive lesson in the sector, and the SpaceX note is about to teach it to a much larger audience.
To understand the note, you have to understand where the watt as a unit came from. For two decades, compute was priced in three ways: by the instance-hour in the cloud, by the reserved contract in the enterprise, and by the spot market at the margin. None of those units map cleanly onto the physical constraint. A GPU is a machine that converts electricity into tokens, and the binding constraint on how many machines you can run is not the chip โ it is the power you can deliver to it and the heat you can remove from it. The chip is fungible and manufacturable. The megawatt is not. Siting a gigawatt-scale data center means land, interconnect queues, water rights, and a utility that will sign a twenty-year contract. That is why the industry slowly migrated its unit of account from the instance to the watt. The watt is the thing you actually run out of. Pricing compute in watts is the same intellectual move as pricing farmland in acre-feet of water: it prices the binding constraint rather than the finished good.
Crypto arrived at the same conclusion from a different direction. When Akash clears a GPU-hour, when io.net's supply curve bends under miner inflow, when Render prices a frame of decentralized rendering, each of those networks is publishing a live, continuous print on the price of compute. That is not a footnote โ it is the only real-time price discovery mechanism the compute commodity has ever had. The SpaceX note quotes a range. The crypto compute sector quotes a curve. And the curve has been telling a very specific story through this bear market: as token emissions inflated the supply of subsidized compute, the clearing price fell toward โ and in several networks below โ the marginal cost of the electricity powering the machine. That is the signature of a market where the subsidy, not the demand, sets the price. I have seen this movie before, in 2021, when I tracked NFT collections and found that roughly nine in ten sales were wash trades by insiders. The tape looked like demand. The tape was the supply talking to itself.
Now do the technical work the note did not do, because the math is where the crypto parallel lives. If short-term compute trades at $30 to $50 per watt and the model assumes $17.60, then the analyst is implicitly saying one of two things. Either the spot price is a scarcity premium that decays as supply catches up โ in which case the $420 billion AI valuation rests on a price that will not survive its own success โ or $17.60 is simply the wrong long-run number and compute is structurally more valuable than modeled, in which case every asset-heavy compute owner on earth is mispriced upward. Jonas cannot have it both ways, and the note does not resolve it. From my seat, the honest read is that $30 to $50 is a scarcity premium, because I have watched this exact pattern in crypto liquidity before. Liquidity is a ghost, not a foundation. It appears abundant exactly when you do not need it and vanishes the moment you rely on it. Compute spot pricing in a supply-constrained quarter is that ghost wearing a new costume.
Run the SpaceX numbers through the crypto lens and something uncomfortable falls out. The note implies a $420 billion AI valuation against an AI division losing $1.26 billion a quarter. That capital-intensity ratio only works if the compute is eventually rented at a premium. But the decentralized networks are simultaneously proving that, absent a subsidy, the clearing price of compute compresses toward power cost plus a thin margin. If that is the long-run equilibrium โ and my own modeling during my 2022 thesis work on algorithmic stablecoin liquidity crises suggests commodity markets always converge to marginal cost eventually โ then $17.60 is optimistic and $30 to $50 is a bubble, not an anchor. The 5.4x revaluation the $300 target requires is a bet that scarcity persists. Commodities do not do that. Commodities mean-revert. That is what makes them commodities. The moment a market stops mean-reverting, it stops being a commodity and becomes a franchise, and franchises require a moat. Compute has no moat. It has capex, and capex is copyable.
Let me put a number on the revaluation so it stops being abstract. At $32 per share on roughly 13.2 billion shares, the AI segment is valued near $420 billion. At $173 per share, it is $2.28 trillion. For context, that single segment would exceed the combined market capitalization of every decentralized compute token in existence by more than two orders of magnitude. The entire DePIN compute sector, measured by fully diluted value, has never cleared $50 billion even at the 2021 and 2024 peaks. So when people ask whether crypto compute can compete with a SpaceX or an AWS, they are asking the wrong question. Crypto compute is not the competitor. Crypto compute is the swing supplier and the transparency layer โ the marginal watt that sets the clearing price and the public tape that shows what that price is. The hyperscalers set the ceiling; the tokenized networks set the floor and reveal the spread. In a bear market, the floor is the only number that matters, because the floor is where the insolvencies happen.
This is where my DeFi Summer scars earn their keep. In 2020 I allocated five thousand dollars across five DeFi protocols and watched yields that had nothing to do with credit risk and everything to do with token emission schedules. I learned then that a subsidized yield is not a yield; it is a transfer. The same accounting applies to subsidized compute. A decentralized network paying miners in newly minted tokens to rent GPU-hours below cost is not selling compute โ it is selling its own equity at a discount and calling the discount revenue. Smart contracts don't manufacture demand. They execute whatever demand exists, faithfully and without judgment, which is precisely why they cannot tell you whether the demand is real. Only the cash does that. And in the SpaceX note, the cash says buyers are paying for rockets and Starlink, not AI. In the crypto compute sector, the portion of revenue not denominated in the protocol's own token is the only line item I trust.
There is a second accounting trap hiding in the note, and it is the same one crypto keeps falling into. The report bundles Grok โ a model โ with compute rental โ an asset โ under a single AI business and hands the bundle one number. That is analytically indefensible. A model business and a compute business have different margin structures, different competitive moats, and different failure modes. A model loses to a better model overnight. A data center loses to a cheaper data center over years. Blending them lets a weak model hide inside a strong asset. If Grok is genuinely behind, as the note's own list of investor worries implies, then the compute rental is not a second business โ it is the first business carrying a passenger. Crypto commits the identical error when it values a DePIN network by blending the value of the physical capacity with the value of the token that coordinates it. The capacity has a market price. The token has a story. Do not average them.
Now the stress test, because the note gives me the ingredients and I refuse to grade it on a bull-market curve. Assume the scarcity premium decays. Assume the $30 to $50 spot compresses toward $17.60 and then toward power cost plus margin, which is where the decentralized networks already trade. At that price, the AI division's $1.26 billion quarterly loss does not shrink โ it widens in relative terms, because the revenue base it is supposed to amortize against shrinks while the depreciation schedule does not. The 86% of capex flowing into AI becomes a fixed cost that the rent cannot cover. The 5.4x revaluation needed for the $300 target does not just fail; it inverts, and the AI segment becomes a drag on the rockets-plus-Starlink valuation that is currently doing all the honest work. That is not a bear case. That is the base case under mean reversion. And the note, tellingly, never runs it.
I should flag the one structural anomaly the entire discussion rests on, because it is the biggest blind spot in the source material. Grok is a product of xAI, not SpaceX. The note treats the chatbot and the compute rental as one SpaceX AI business, which silently assumes xAI has been merged into or deeply integrated with SpaceX โ something Musk has publicly floated but never confirmed. If that integration is real, it is the single most consequential undisclosed event in the story, because it changes the counterparty, the IP ownership, the governance of compute allocation, and the competitive set entirely. If it is not real, the valuation logic has a factual hole at its foundation. Either way, the note never says which world we are in. A valuation that cannot name its own counterparty is not a valuation. It is a vibe with a spreadsheet.
This is also where the DeFi rate models deserve a hard look, because the crypto compute sector is quietly importing the same arbitrariness. Aave and Compound price borrowing with interest rate curves that were reverse-engineered to produce nice utilization behavior, not to reflect the real cost of capital. They are conventions dressed as economics. The emerging compute-financing markets โ lending against GPUs, against hashrate, against contracted megawatts โ are building rate curves the same way, by fitting a shape that looks reasonable rather than anchoring to the true cost of power and depreciation. When the cycle turns, those curves will not protect the lenders. They never do. The only rate that matters is the one a real borrower pays in a real dollar for a real watt, and everything else is calibration theater.
Now the part that runs against the room. The consensus, reading this note, is that SpaceX's AI valuation is aggressive and that crypto's decentralized compute sector is the speculative shadow of the same trade. I think the relationship is closer to inverted, and here is the decoupling thesis.
The bear market is not killing decentralized compute. It is finally pricing it. When token emissions were the marginal dollar, every network looked solvent because every network was paying itself. The bear market stripped the subsidy out and exposed the clearing price of a delivered GPU-hour โ and for the first time, that price is comparable to what a hyperscaler charges, because the artificial discount is gone. That is not a decline. It is a discovery. The networks that still clear orders today are the ones whose compute was always worth more than their token. The ones that went quiet were never selling compute at all โ they were selling emissions with a GPU as a prop. A market that can tell those two apart is a market that has matured, even if its prices have not. The watt is the only honest unit in AI, and the bear market is the only honest auditor in crypto.
The second contrarian point cuts the other way, against the bulls, and it is the one that should keep you up at night. The SpaceX note's own data undermines the scarcity story it is selling. Buyers are not paying for AI. Forty institutional clients, asked directly, did not raise a hand to hold the stock. The company president sold into strength. That is not a demand signal; that is the absence of one wearing a demand signal's clothing. When the AI capex cycle runs ahead of the revenue that justifies it, the gap does not close gently. It closes the way Terra closed โ fast, in public, with the leverage on the wrong side of the mark. I spent my thesis years dissecting exactly that failure, and the tell was always the same: the protocol needed new capital to service old promises, and the promises were denominated in the protocol's own confidence. The crypto compute sector learned this lesson in 2022. The AI compute sector has not learned it yet, because it has never had a down cycle. It is about to have its first, and it will have it in a market where the marginal supplier is a tokenized network that has already been marked to reality.
And a final structural note, because my Layer 2 work and my compute work rhyme. The crypto industry spent years over-engineering the data availability layer โ building dedicated chains and modular stacks for a data problem that the overwhelming majority of rollups never actually had. The same over-engineering is now happening in compute verification: elaborate proof-of-GPU schemes, zero-knowledge attestations of model execution, decentralized inference markets that verify outputs nobody is paying to verify. The value is not in the verification theater. The value is in the watt, the land, the power contract, and the customer who pays in dollars. Everything else is a ghost with a whitepaper.
So where does this leave the positioning, in a bear market where survival outranks gains? If you hold decentralized compute tokens, the only question that matters through the rest of this cycle is whether the network's revenue is real โ paid in something other than its own token, for compute someone actually consumed. If it is, the SpaceX note just handed you a valuation framework: your network is a swing supplier whose clearing price anchors the low end of a commodity that a rocket company is pricing at $420 billion. If it is not, you are holding a subsidized yield with extra steps, and the subsidy is the first thing the next leg down removes. Watch the dollar-denominated revenue line, not the fully diluted valuation. Watch the clearing price, not the roadmap. Watch the utilization curve, not the whitepaper's architecture diagram. And watch the $30/W spot print. If it holds through a supply expansion, compute really is scarce, and the bulls โ crypto and otherwise โ are right. If it compresses toward power cost, then the most expensive number in this entire story is not the $300 price target. It is the assumption that scarcity was permanent. It never is. And when the next cycle asks you what you were paying for, you had better be able to say: a watt, delivered, in dollars โ not a token, promised, in hope.