Twenty-six point five trillion dollars. That is the number SpaceX wants anchored in your memory when it describes orbital AI data centers — the business it now frames as potentially its most profitable ever. The trigger is a launch schedule. As early as next week, a Starship is set to lift the first batch of V3 Starlink satellites, hardware that only a super-heavy reusable vehicle can deploy at economic cost. On paper, this is the moment SpaceX stops pricing itself as a rocket company and starts pricing itself as an orbital compute utility.
Then there is the physical inventory behind the claim. No thermal demonstration. No radiation-hardened inference cluster operating in vacuum. No downlink architecture capable of moving training gradients at the bandwidth the story quietly assumes. A number. A launch window. That is the entire asset base of the twenty-six-point-five-trillion-dollar headline.
The market is not rational; it is resistant. And this week the resistance surfaces somewhere most crypto analysts never look — not in low Earth orbit, but in the terrestrial compute markets that blockchain-linked assets have spent two years quietly re-rating against.
Starlink's communications network produced close to fifty-five percent of SpaceX revenue in the second quarter. That figure is a valuation hinge, not a footnote. It marks the point where the company's center of gravity shifted from building rockets to operating a subscription connectivity business — from a manufacturing multiple to an ARPU-times-subscribers multiple. When a company changes what it is, the market changes how it pays for it.
Crypto has run this exact re-rating script before. Between 2024 and 2026, a cluster of listed Bitcoin miners — Core Scientific, Riot, Hut 8, IREN — stopped being valued primarily on hash rate and started being valued on megawatts of interconnectable power and signed hosting contracts with AI hyperscalers. The equity market did not reward them for mining bitcoin. It rewarded them for owning the physical inputs to compute: land, power, transformers, grid interconnection queues. The same repricing SpaceX is now attempting in orbit — not satellites as assets, but capacity as assets — already happened on the ground, and it happened to crypto-native balance sheets first.
The decentralized mirror of that trade is DePIN — decentralized physical infrastructure networks. Render, Akash, io.net, Filecoin. Each sells a version of the same thesis: that idle or distributed compute can be aggregated, verified, and sold cheaper than the hyperscaler alternative. The pitch is real in narrow slices. It is also structurally fragile, which I will get to, because the same fragility defines what orbital compute can and cannot become.
To read the Starship launch as a crypto story, you have to see the macro liquidity map underneath it. Capital in 2026 is not flowing evenly. It is converging on one commodity: verifiable compute with a power contract attached. Every dollar of growth capital is bidding for the same scarce inputs — energy, cooling, interconnect, and now orbital slots. SpaceX is not raising against a demand curve. It is raising against a supply constraint, and dressing the constraint as a market.
Strip the narrative away and the launch is a logistics event. Starship is a reusable super-heavy vehicle roughly four hundred feet tall, and it exists because V3 satellites are too heavy for any existing fairing to deploy in economic volume. The vehicle is not a symbol of ambition here; it is the physical gateway to the next generation of the constellation. If Starship flies on schedule and reuses cleanly, the marginal cost of orbiting capacity falls by an order of magnitude and the entire constellation expansion pays for itself. If it slips, the constellation pauses, and the orbital data center story has no delivery mechanism at all. The whole thesis is levered to one machine.
Zoom out and the macro layer reinforces the point. Capital in 2026 is converging on verifiable compute with a power contract attached, and it is doing so while the cost of that capital is higher than it has been for a decade. Artificial intelligence has become a capital-absorption machine, and anything that promises to expand the compute supply — terrestrial or orbital — is bid for on the promise of future abundance. That is precisely the condition in which TAM figures inflate fastest and scrutiny falls lowest. Liquidity does not fund the best idea; it funds the most quotable one.
Begin with the physics, because physics writes the only honest balance sheet in this trade. An orbital data center has three irreducible problems. The first is thermal. In vacuum there is no convection and no conduction to the atmosphere. Heat leaves a spacecraft almost entirely by radiation, proportional to the fourth power of temperature and the surface area available. A terrestrial ten-megawatt training rack dumps heat into cooling towers. The same rack in orbit must radiate it from fin area measured in thousands of square meters, which means either a spacecraft too large to launch economically or a compute load too small to matter. Compute density in orbit is capped by radiator area, and radiator area is capped by launch mass. The twenty-six-point-five-trillion-dollar story has to clear that equation first.
The second problem is radiation. Cosmic rays and solar particles flip bits in commercial silicon. Terrestrial data centers run on commodity accelerators that assume a quiet electromagnetic environment. In low Earth orbit, single-event upsets accumulate, and hardening costs performance. You can shield, but shielding is mass, and mass is the radiator problem again. The third problem is downlink. Training large models requires moving enormous volumes of gradient and checkpoint data between accelerators. Starlink's user downlink is building out toward high throughput, but it is optimized for asymmetric consumer traffic, not for the east-west fabric a training cluster needs. You cannot train a frontier model across an inter-satellite link that was designed to stream video to a fishing boat.
None of this makes orbital compute impossible. It makes it specific. It will be viable for inference at the edge — low-latency, small-model, sovereignty-constrained workloads that benefit from being physically outside any single jurisdiction. It will not be viable, in any near-term window, for the large-scale training that the twenty-six-point-five-trillion-dollar number implicitly assumes.
Run the arithmetic on a single rack. A one-megawatt compute load in orbit, running at modest efficiency, must reject roughly a megawatt continuously. Radiating that at a workable temperature demands on the order of several thousand square meters of radiator surface, depending on emissivity and view factor. That is a structure larger than the compute it cools. Fold in the mass of the solar arrays needed to generate the power in the first place, add the battery mass for eclipse periods, and the launch mass per unit of usable compute climbs by an order of magnitude over a terrestrial data center. Starship's cost-per-kilogram advantage is real, but it has to overcome a structural penalty that no cheaper launch fully erases.
Which brings the DePIN comparison into focus. When I built a framework for decentralized compute economics, the first question I forced every network to answer was not about token incentives. It was about the physical bill. Render's distributed GPU rendering works because rendering is embarrassingly parallel, latency-tolerant, and checkpoint-independent — the exact profile that tolerates fragmentation. io.net aggregates smaller clusters for inference and fine-tuning, where the parallelism is coarser but the latency budget is still generous. Akash sells containerized compute where the verification cost is low. Each of these networks succeeds precisely to the degree its workload tolerates its topology.
Orbital compute is the inverse case. Its topology — high-latency, radiation-exposed, thermally constrained — is the worst possible fit for the workloads with the highest willingness to pay. The DePIN networks that survive have matched workload to topology. SpaceX is proposing to match a topology to a market that does not yet exist in the shape the topology demands. Fractures in the ledger reveal the truth of value. Here the ledger is the power and thermal budget, and it is already showing the fracture.
The terrestrial repricing tells us what actually captures value. When Bitcoin miners pivoted to AI hosting, the market did not pay them for the model weights. It paid them for the interconnect queue, the substation, the water rights, the signed power purchase agreement. Value accrued to the bottleneck. Every crypto-native operator that owned a bottleneck got re-rated; every operator that owned only the compute got commoditized. This is the single most important lesson to carry into orbital compute: if orbital capacity is abundant relative to demand, the bottleneck is not the satellites. It is the downlink bandwidth and the launch cadence — and those are exactly the assets SpaceX controls and its competitors must rent.
The pattern is visible in the deal structures themselves. When AI hyperscalers signed hosting agreements with converted miners, the contracts were written against power, uptime, and interconnect — never against the model or the inference output. The miner is paid for being a reliable landlord to compute, not for owning intelligence. That distinction matters enormously for anyone pricing orbital capacity, because it tells you where the margin actually settles when compute is abundant and power is scarce: it settles with whoever controls the scarce input, and it never settles with whoever merely owns the chips.
This is also where the Bitcoin fee-market argument becomes unavoidable. Bitcoin's long-run security budget depends on transaction fees replacing the block subsidy, and the marginal fee revenue has been carried by Ordinals and inscription activity — an inflow that most of the industry dismissed as a fad and that quietly underwrote the security model during a period when it would otherwise have been strained. The lesson generalizes: the second-order revenue stream, not the headline use case, is what keeps an infrastructure alive. SpaceX's second-order revenue stream is not the AI market. It is the launch-and-connect franchise that the AI market narrative happens to justify.
Now dissect the number. Twenty-six point five trillion dollars is a total addressable market figure. TAM is a ceiling, not a forecast. The distance between TAM and serviceable addressable market, and between SAM and serviceable obtainable market, is where nearly every failed crypto valuation lives. In 2017 I audited more than fifty ICO whitepapers for a Stockholm fund, and the pattern was monotonous: a credible technical team, a real problem, and a TAM slide inflated by one or two orders of magnitude because nobody asked what fraction of the market could actually be captured within the token's life. We shorted specific altcoins on that basis and went long the infrastructure, and the portfolio returned forty percent in a year when most of the sector round-tripped toward zero. The twenty-six-point-five-trillion-dollar figure is the same TAM slide, printed on a Starship fairing.
The tell is that the number is quoted as a market, not as a margin. AI compute is a brutal-margin business at scale because it is capital-intensive, power-constrained, and rapidly obsolesced. Even the terrestrial hyperscalers live with depreciation cycles that would horrify a software investor. Put that business in orbit, add a five-to-seven-year satellite replacement cycle, add launch risk, and you have compounded the worst features of telecom and semiconductors into a single asset. Entropy is the only constant in liquid markets — and orbital hardware experiences entropy faster than almost anything investors routinely price.
The binding constraint is not silicon. It is capital and energy. In 2020 I spent three months modeling the liquidity depth of Uniswap v2 and Compound, specifically tracing how stablecoin pegs correlated with Ethereum gas spikes. The paper that came out of it, The Illusion of Infinite Liquidity, predicted the volatility cascades that peak congestion would produce, and it was dismissed by bullish peers until the cascade arrived. The structural lesson was that apparent depth evaporates the moment the underlying resource gets scarce. Compute has the same property. The apparent abundance of capital for AI infrastructure is a function of cheap energy and cheap credit. Both are tightening. Orbital compute adds a third scarce input — launch cadence — on top of two that are already rationed.
SpaceX's moat is real and it is unusual in its dimensionality. Vertical integration across launch, satellite manufacturing, spectrum, and terminals means it can price internal transfer of launch cost in a way no competitor can. Its first-mover advantage is not just reputational; it is the physical occupation of orbital planes and the spectrum filings that go with them, awarded on a first-come basis by the FCC and ITU. Spectrum and orbital slots are one-time, first-occupancy resources. That is a moat that compounds and cannot be rebuilt by a later entrant, and it is the strongest part of the whole story.
But the same integration creates a single point of failure. V3 satellite economics, the orbital data center, the lunar and Mars ambitions — all of them route through one vehicle, Starship, that has not yet demonstrated a stable reliability curve. The honest reading is that the next twelve to eighteen months of this story are not a market question. They are a launch-cadence question. If Starship slips, every downstream narrative slips with it, and the twenty-six-point-five-trillion-dollar number gets pushed into a future that the current cost of capital may not finance.
The competitive set is forming faster than the launch schedule suggests. Google's Project Suncatcher is exploring exactly the orbital compute thesis, paired with its terrestrial cloud franchise — and Google, unlike SpaceX, already owns the developer ecosystem, the model training stack, and the enterprise relationships that make compute sellable. Kuiper and the Chinese constellations are compressing the communications window from below. SpaceX's advantage is launch cost and orbital real estate. Its disadvantage is that it would be entering compute as a commodity supplier against incumbents who own the margin layer. Being cheapest to launch a data center is not the same as being best at running one.
The valuation mechanics deserve their own paragraph. SpaceX is being asked to accept a telecom multiple on its connectivity business and a speculative technology multiple on its compute ambition. Those two multiples pull in opposite directions. Telecom is priced on EBITDA and cash-flow conversion, because satellites depreciate and require constant replacement capex. Compute, when it is real, is priced on growth and margin expansion, because software scales without depreciation. A company that is fifty-five percent telecom and zero percent delivered compute should be valued on the first curve, not the second — unless the market is deliberately paying for the option. And option pricing, as any derivatives desk knows, does not survive contact with a missed delivery date.
Seen through an infrastructure lens, orbital compute is a deferred version of infrastructure-as-a-service. The terrestrial IaaS model sells metered compute, storage, and network with a developer API on top. If SpaceX ever opens orbital capacity to third-party tenants, it inherits that model almost unchanged — except that its hardware is unreachable, unrepairable, and depreciating against a launch calendar rather than a refresh cycle. That is not a fatal flaw, but it is a structural discount. You can rent a satellite's compute the way you rent a virtual machine, but you cannot reboot the satellite when the kernel panics. The service-level agreement that would let enterprises commit real workloads to orbit does not yet have a physical form.
There is one more layer worth tracing, and it is regulatory. Orbital compute is fundamentally a sovereignty arbitrage. Data processed in orbit crosses jurisdictions without crossing a border checkpoint, which is either the business's greatest feature or its fatal legal flaw depending on which capital you are asking. The jurisdictions that will matter are the ones that decide whether orbital bits count as foreign processing. This is the same dynamic that governed Hong Kong's virtual asset licensing regime — a framework that was never primarily about innovation, but about recapturing the financial-hub position Singapore had taken during the pandemic years. Nations compete for the right to define where value is recognized. Orbital compute turns that competition into a physics problem, and the country that wins the definition will not be the one with the best technology.
Here is the counter-intuitive claim, and it is the one worth arguing. Orbital data centers will not compete with terrestrial AI training at scale within any investment horizon that matters to a portfolio. They will serve three narrow, high-value niches: sovereignty-constrained inference, edge caching for latency-sensitive applications, and defense workloads that value jurisdictional independence above unit economics. The bull case that orbital compute becomes the substrate of the AI economy is not a forecast; it is a category error that conflates the scarcity of compute with the scarcity of a specific location for compute.
The deeper contrarian point is about what gets repriced. The market is treating orbital compute, DePIN compute, and Bitcoin-miner-to-AI compute as three expressions of one trade. They are not. They share a narrative and almost nothing else. DePIN owns verification and price discovery. Miners own power and interconnect. SpaceX owns launch and spectrum. When the AI capex cycle finally compresses — and cycles always compress — these three will decouple violently, because only one of them holds a bottleneck that cannot be substituted. My bet, from the audit seat rather than the trading desk, is that the bottleneck that survives will be physically located on the ground, not in orbit.
The decoupling will not be gradual. It will look like a single quarter in which the market realizes that three assets it priced together have three different fundamental drivers. DePIN tokens will follow verification demand and fee revenue. Bitcoin miners will follow power pricing and hyperscaler contract renewals. SpaceX, being private, will follow launch cadence and sovereign data policy. The moment a portfolio treats orbital compute as the tide that lifts all three, it is buying correlation it has not actually underwritten. Entropy is the only constant in liquid markets, and entropy is what dissolves false correlations first.
The signal to watch is not the launch. It is the downlink. Track Starlink's per-satellite throughput curve, not its satellite count, because throughput is the actual ceiling on any orbital compute thesis. Track Starship's reuse cadence, because cost per kilogram is the gravity of the entire story. And track whether any sovereign customer signs a binding inference contract that requires the data to physically leave the planet — a contract with a penalty clause, not a memorandum of understanding. If that contract appears, the niche is real and the repricing is justified. If it does not, the twenty-six-point-five-trillion dollars is a number looking for a physics problem to solve — and physics, unlike the market, does not negotiate.