SpaceX Has Two AI Compute Stories; Only One Generates Revenue

Dave Friedman

May 26, 2026

SpaceX’s S-1 dropped on EDGAR on May 20th. Most of the coverage focused on the $28.5 trillion total addressable market claim, the dual-class governance, or the gap between Musk’s Davos rhetoric and the risk-factor language about orbital data centers that “may not achieve commercial viability.” All fair game. But the most interesting thing about the filing isn’t any single claim. It’s that SpaceX is telling two stories about the future of AI compute infrastructure, and the document never explains how they fit together.

Story one: SpaceX is spending billions building terrestrial data centers and has signed one disclosed external customer, Anthropic, to a deal worth $1.25 billion per month for capacity at COLOSSUS and COLOSSUS II. That contract runs through May 2029. At run-rate, it represents roughly $15 billion a year in revenue from infrastructure that already exists, on Earth, plugged into a power grid, cooled by conventional systems.

Story two: SpaceX argues that the future of AI inference belongs in orbit, powered by solar arrays in sun-synchronous orbit where energy is near constant and cooling is handled by radiative dissipation into space. The filing calls this the path to the lowest achievable cost per token and claims SpaceX is “the only company that has already accomplished the key technical challenges associated with evolving connectivity satellites into AI compute satellites.”

Both stories are presented with conviction. Neither is presented as contingent on the other being wrong. The S-1 never provides a bridge between them. No migration curve, no cannibalization model, no analysis of which workloads move to orbit first, no disclosure of whether COLOSSUS-class ground assets remain premium infrastructure once orbital capacity comes online, and no stranded-asset sensitivity. Investors are implicitly asked to capitalize both: the terrestrial scarcity rents and the orbital optionality. The filing does not explain why those two values should be additive rather than one haircutting the other.

The cash flow story: terrestrial scarcity

SpaceX’s AI segment, the former xAI, merged into SpaceX in February 2026, operates what the filing calls the largest AI training data center clusters on Earth. COLOSSUS and COLOSSUS II collectively provide approximately 1.0 gigawatt of compute power, with additional power capacity available for data center operations. SpaceX brought the first cluster of COLOSSUS online in 122 days and COLOSSUS II even faster, in 91 days, against an industry benchmark of roughly two years for a 100-megawatt greenfield facility.

That speed advantage has been monetized. The Anthropic Cloud Services Agreements, disclosed in the S-1, provide for $1.25 billion per month through May 2029, terminable by either party on 90 days’ notice. SpaceX frames this as a “dual monetization strategy,” using excess capacity to generate returns while retaining the right to reallocate for internal use if needed. The filing notes that SpaceX has “sufficient capacity to provide compute for our own A models, including support of our training and inference demands, and to satisfy the obligations under these agreements,” and expects to sign additional similar contracts.

Using the full 1 GW facility base as the denominator gives you a rough revenue-density benchmark of $15 million per megawatt per year. If Anthropic is not receiving the full gigawatt--and the filing suggests SpaceX retains capacity for its own models and future customers–then the implied price per contracted megawatt is higher. Either way, the point is not precision. The point is that the terrestrial business is being valued off current scarcity rents, and those rents are very real. Data center vacancy rates are at historic lows. The demand-supply imbalance in AI compute is severe. Anthropic is paying $1.25 billion a month to a competitor’s infrastructure because it has no better option. That is what scarcity pricing looks like.

The AI segment did $818 million in revenue and lost $2.5 billion from operations in Q1 2026, reflecting massive ongoing capex of $7.7 billion in the quarter alone. The Anthropic contract, at full run-rate, would roughly double the segment’s quarterly top line. This is the near-term financial engine of the AI segment, and it is entirely terrestrial.

The optionality story: orbital inference

Now the other story. SpaceX frames AI economics as a cost-per-token problem, and cost-per-token as a function of three inputs: the underlying model, the compute hardware, and energy. The filing claims a competitive advantage in the latter two: (1) hardware through proposed vertical integration efforts such as Terafab, whose collaboration framework with Tesla and Intel remains early stage, with specific projects subject to separate negotiations and neither party contractually obligated to remain in the project; and (2) energy through orbit.

The energy argument is specific. Space-based solar arrays generate more than five times the energy per unit area of terrestrial solar, thanks to continuous illumination, no atmospheric interference, and optimal orientation. Meanwhile, U.S. electricity generation grew at a compound annual rate of 0.1% from 2008 to 2023, and even with the AI-driven demand surge, has managed less than 3% annual growth between 2023 and 2025. Data center buildout is dramatically outpacing grid capacity. SpaceX is arguing that the binding constraint on AI scaling isn’t chips, but rather it is watts. And orbit has lots of watts.

But the more aggressive claim here is about timing. SpaceX argues that orbit bypasses the terrestrial bottleneck of grid interconnection, which is the permitting, substation construction, and utility negotiation that gate when a new data center can actually turn on. The filing claims that near-constant solar access will allow SpaceX to “consistently activate the highest performing hardware before our competitors without such access, shrinking the timeline to useful tokens on bleeding-edge hardware.”

This is a time-to-revenue argument dressed as an energy argument. If you finance data centers, your models have a line item for the gap between capex deployment and first revenue. SpaceX is asserting that orbit compresses it.

The deployment math is ambitious enough to be falsifiable. SpaceX projects 100 gigawatts of annual compute via satellites carrying over 100 kilowatts per metric ton. That requires “thousands of launches per year and the transport of approximately one million metric tons to orbit annually.” As of March 31, 2026, SpaceX had launched a cumulative 7,400 metric tons to orbit. The orbital compute program requires roughly 135 times that amount, every year. Initial deployment: “as early as 2028.” The 100-gigawatt target has no stated timeline. The filing concedes it “may be difficult or impossible to determine.”

Orbital inference does not need to replace all terrestrial compute to be valuable. It only needs to capture workloads where energy availability, autonomous batch processing, or model-side routing outweigh latency and serviceability constraints. That is the strongest version of the bull case. But the S-1 does not tell investors which workloads those are, what their revenue density looks like, or when they migrate.

The missing bridge

.The S-1 provides no quantitative model for orbital compute economics. No cost-per-kWh in orbit. No cost-per-token comparison between terrestrial and orbital. No projected launch cost for Starship at scale. No satellite unit cost. No capex buildout curve. No IRR. No breakeven timeline. The filing doesn’t even give orbital compute its own TAM. It’s folded into a $2.4 trillion “AI infrastructure” bucket that includes the terrestrial business.

But the filing does provide enough physical parameters to build a constraint map. This isn’t a forecast, but rather a boundary analysis that shows what would have to be true for the orbital thesis to work economically. The gaps tells you as much as the numbers SpaceX provided. I had Claude 4.6 build the spreadsheet; it’s available to review here.

From the filing, three parameters are fixed: Starship V3 delivers 100 metric tons to LEO per launch; AI compute satellites target 100 kilowatts of compute per metric ton; and COLOSSUS plus COLOSSUS II provide approximately 1.0 gigawatt collectively against the Anthropic contract. That gives you 10 megawatts of compute deployed per Starship launch and a terrestrial revenue-density floor of roughly $15 million per megawatt per year.

What is missing from the S-1 is the investable cost model: all-in satellite cost, replacement cycle, repairability, capital cost, and commercial pricing. Three undisclosed variables dominate the economics.

Starship cost per launch. Musk has claimed $2-10 million at full reusability. Industry estimates range from $30 million to $100 million or more. This turns out not to be the swing variable: at 10 megawatts of compute per launch, even a $100 million launch adds only $10 million per megawatt to capex. Launch cost matters but does not dominate.

All-in satellite cost per kilowatt of compute. This is the black box. The number bundles the satellite bs, solar arrays, radiative cooling systems, compute hardware, and integration into a single figure. Nobody has published this number with the precision needed for a capital allocation decision. Reasonable guesses span a wide range: $25,000 to $200,000 per kilowatt. This is the variable that dominates total capex at nearly every combination of assumptions.

Satellite operational lifespan. GPU architectures turn over every two to three years. Starlink broadband satellites are depreciated over roughly five. The S-1’s own risk factors note that orbital infrastructure “will not be easily repaired or upgraded, such that any component failures could result in permanent capacity loss, accelerated depreciation, decommissioning or need for replacement.” A three-year effective life roughly doubles the annualized cost versus six years. You can’t just swap the GPU. The whole satellite gets replaced. Every replacement requires another Starship launch.

At mid-range assumptions, of $30 million per launch, $75,000 per kilowatt all-in, four year satellite life, and 5% annual operating costs as a share of redeployed capex, deployed capex runs roughly $78 million per megawatt ($3 million from launch, $75 million from satellite hardware). Annualized over a four-year life with operating costs, that’s approximately $23 million per megawatt per year. This is about 55% above the Anthropic revenue-density benchmark.

But comparing orbital cost to the Anthropic benchmark is unfairly generous to orbit, because the Anthropic number is a scarcity price, not a cost benchmark. Beating it does not prove orbital compute is cheaper than Earth. It proves only that orbital compute might fit under a temporary price umbrella created by extreme capacity scarcity. The harder question, which investors should actually care about, is whether orbital beats terrestrial at fully loaded cost.

Terrestrial facility-and-power cost, before accelerator depreciation, can be amortized into the low single-digit millions per megawatt per year. Add GPU depreciation and the number rises materially. But even then, the orbital case has to beat not just today’s scarcity price but a terrestrial stack whose largest cost components are serviceable, financeable, replaceable, and located somewhere lenders can actually reach.

The model also excludes factors that tilt further against orbit: latency, ground station routing, and workload replacement constraints, along with GPU radiation degradation in LEO, thermal cycle effects on hardware lifespan, Starship development cost amortization, the capital cost of the Starlink download network, insurance, and debris liability. Every one of these factors make orbital look worse.

How should investors capitaliz both?

Spa

[truncated for AI cost control]