ARK Invest’s latest research note is a masterclass in narrative engineering. It claims SpaceX will derive 90% of its future growth from AI infrastructure. That is a lie—wrapped in a cost-per-kilogram assumption and sealed with a buzzword. I’ve read this script before. It’s the same one I saw in 2018 when a project called “Aether” promised a decentralized cloud with zero technical proof. The code did not lie then. The balance sheet does not lie now. Only the founders do.
SpaceX, according to ARK, is no longer just a rocket company. It is now an “AI infrastructure company.” The narrative runs like this: vertical integration of rockets, satellites, and data centers will slash launch costs to under $100 per kilogram. Orbital data centers will operate at 25% lower build cost than ground-based equivalents and near-zero energy cost. The company is already leasing compute capacity to clients like Anthropic and Google. The valuation story shifts from a mature launch provider to a high-growth AI disruptor, justifying a massive post-IPO market cap.
I don’t trust the investor deck; I trust the launch manifest. ARK cites a $100/kg target for Starship—a vehicle that has not yet completed a single successful orbital re-entry with payload. The current Falcon 9 commercial launch price hovers around $2,500/kg. The gap between $2,500 and $100 is not a linear improvement; it is a factor of 25. In crypto terms, that is like promising a Layer 2 that scales 25× without any fraud proof mechanism. Possible in theory. Non-existent in practice. The entire orbital compute thesis rests on this single variable. If Starship fails to hit sub-$500/kg, the economics collapse.
Let’s dig into the orbital data center claim. ARK says “build costs 25% lower” and “energy costs near zero.” This ignores the physics of space. Solar panels in low Earth orbit generate roughly 1.3 kW per square meter, but thermal management for high-performance GPUs consumes massive power for active cooling. A single H100 GPU dissipates 700W of heat. In vacuum, you cannot use air cooling. You need radiative panels—heavy, expensive, and failure-prone. The added mass eats into the launch cost advantage. I have audited hardware-backed token projects where the team assumed off-the-shelf servers would work in a shipping container on the moon. The real cost of radiation hardening, vibration isolation, and redundant electronics often triples procurement budgets. ARK’s 25% lower build cost is an assumption that has never been validated by a prototype. It is a term in a spreadsheet, not a weld on a satellite.
What about the customers? The note mentions Anthropic and Google as lessees of computing resources. It does not state the scale, the contract duration, or the price per ExaFLOP. In my experience auditing crypto protocols, a “partnership” with a major exchange often means a $10,000 marketing spend, not a liquidity injection. The same applies here. Anthropic may rent a few racks of GPUs out of strategic curiosity. That does not signal a mass migration of training workloads to space. The missing data point is the unit cost. How much does SpaceX’s compute cost compared to AWS’s spot instances or CoreWeave’s dedicated clusters? Without that, the claim is empty. Reentrancy is not a bug; it is a feature of trust—and trust without data is just a marketing budget.
Ark’s thesis suffers from three systematic failures that I have catalogued across dozens of crypto project autopsies. First, the single-point-of-failure dependency. Everything hinges on launch cost. If Starship development slips two years, the AI narrative loses its foundation. Second, the neglected edge case. Orbital data centers face radiation-induced bit flips, thermal cycling fatigue, and a total inability to service hardware. SpaceX has never demonstrated a space-grade data center that runs continuously for more than 90 days. In DeFi, a rounding error in a borrow rate can cause insolvency under high volatility. Here, a single solar panel failure could strand an entire compute cluster. Third, the valuation mismatched to reality. ARK wants to price SpaceX as an AI infrastructure company, yet the revenue today comes from launch and Starlink internet. The 90% growth claim implies that AI will generate billions in incremental revenue within five years. That would require SpaceX to capture an improbable share of the $500B data center market. I saw the same math in 2021 when NFT projects projected million-dollar floor prices based on infinite demand. The difference is that those contracts were audited—poorly, but audited. SpaceX’s AI business has no audit trail. The code does not lie; only the founders do.
But let’s not dismiss the bull case entirely. Bulls argue that SpaceX’s vertical integration gives it a structural advantage no cloud provider can replicate. Amazon relies on third-party data centers, power grids, and fiber. Spacex owns the rockets, the satellites, the ground stations, and the energy source. If launched costs truly collapse, SpaceX could become the world’s lowest-cost compute wholesaler. The “wholesale” angle is key: they don’t need to build a full cloud platform like AWS. They can sell raw GPU clusters to hyperscalers, letting Amazon and Google handle the software stack. This is analogous to crypto mining farms selling hash power to pools rather than building their own exchanges. It is a narrower but more defensible moat. And the orbital data center, if it works, offers unique value for latency-insensitive workloads like geological modeling, pharmaceutical simulations, or AI training on privacy-sensitive data (no territorial data residency issues when the data never touches any sovereign soil). The contrarians might have a point: the future of compute could have a space component, and SpaceX is the only company with the engine to get there.
The takeaway is simple: ARK’s note is a brilliant piece of capital markets storytelling. It converts the uncertainty of rocket science into the certainty of a growth curve. But as an auditor, I see a missing line of code—the line that proves the $100/kg assumption is real, the orbital data center is built, and the customer commitments are binding. Without that, the entire thesis is a speculative token with no lock-up period. The rug was pulled before the mint even finished. the real question is not whether SpaceX can build orbital AI infrastructure—it can, eventually. The question is whether the market is pricing that eventuality as if it has already happened. History shows that when narratives run ahead of reality, the correction is swift. In 2018, I watched a project with zero code raise $40 million. In 2021, an NFT collection with a single line of metadata raised $2 million before the rug. The pattern is always the same. The only difference is the suit.
Gas fees don’t lie. But when the gas is rocket fuel, the meter stops working. Investors should demand a full audit of the economic assumptions, not just the engineering. Until then, I’ll trust the launchpad, not the pitch deck.
