The Compute Landlord Paradox: SpaceX Is Selling 2026, But the Lease Is Priced for 2030

SpaceX reported $7.8 billion in Q2 2026 revenue, up 92% year over year. That growth figure looks heroic until you break the P&L apart. Back out the AI segment using the $1.26 billion operating loss and a 60% gross margin assumption, and the compute business is contributing roughly $3.1 billion per quarter. The two anchor tenants are Google and Anthropic. One reportedly pays $920 million per month for about 110,000 GPUs; the other pays $1.25 billion per month at the Colossus 1 campus. Those numbers do not survive arithmetic. $920 million divided by 110,000 GPUs is $8,400 per GPU per month, while the same analysis cites an $840 per-GPU market rate. Both cannot be true. A rocket company has become a compute landlord whose revenue model is a leaky spreadsheet. The question is not whether SpaceX can build a data center. The question is whether the lease terms can carry a 10GW capex plan. This is not an optional pivot. SpaceX has locked down Nvidia's Vera Rubin architecture ahead of the product's full-scale ramp: HBM4 memory, advanced 2.5D/3D packaging, and a claimed 25x per-GPU jump over H100. It has signed $6.7 billion in forward cloud service contracts. It is planning 2GW online by the end of 2026 and 10GW by the end of 2027. No single organization in history has deployed 10GW of AI compute in two years. The satellite branch is even more speculative - Starmind AI1 will carry Nvidia Space-1 Vera Rubin modules into low Earth orbit, where radiation-induced bit flips and radiative cooling limit effective throughput. The 25x benchmark is a ground-context number, not an orbital inference number. The core order flow has four layers. First, tenant concentration is the real risk. Google and Anthropic together make up nearly the entire AI rental book. If the leaked monthly figures are directionally correct, combined revenue runs above $26 billion annually from two counterparties. This is a two-tenant CLO, not a diversified REIT. A traditional property with 40% tenant concentration gets a valuation discount. SpaceX gets a premium because one tenant is Google. That premium evaporates the moment Google signs a build-to-suit with CoreWeave or shifts training capacity in-house. Smart money doesn't price a compute landlord as if the tenant is the asset. Second, unit economics are too clean to be credible. The spot market for H100 GPU instances in North America is around $100 per GPU per month for a base node. The reported Google rate of $840 per GPU-month is an 8x premium. That can be justified as priority access, liquid cooling, full-stack networking, or simply unreliable source data. But premium pricing on two concentrated leases is not an anchor for a 10GW buildout. When Nvidia supply catches up and CoreWeave, Lambda, and TensorWave keep adding 100MW campuses, compute rental prices compress. GPUs are not Bitcoin. Supply is a manufacturing function, not a fixed cap. The moat in any landlord business is unit cost, and SpaceX has not shown its cost per GPU-hour. Third, depreciation is the silent liability. A 10GW program at current industry capital intensity - roughly $30 million to $40 million per MW - implies $300 billion to $400 billion of cumulative capex. On a five-year straight-line schedule, annual depreciation lands between $60 billion and $80 billion. Compare that to the AI segment's roughly $12.6 billion annualized revenue base. EBITDA will look strong because EBITDA ignores the single largest cost in the model. This is the same trap I saw in DeFi during the summer of 2020. Yield farmers celebrated gross APY while ignoring collateral utilization and liquidation costs. The landlord's liquidation cost is depreciation. Smart money doesn't ignore depreciation; it prices it as the cost of future revenue. Fourth, the Nvidia lock-up is a call option with a counterparty. SpaceX is betting its entire compute roadmap on Vera Rubin's execution timeline. The Hopper-to-Blackwell transition already produced scheduler, networking, and software compatibility problems for early adopters. If Vera Rubin slips, SpaceX loses its first-to-2GW advantage. Nvidia also has an incentive to seed multiple compute landlords - CoreWeave, Lambda, Oracle - to balance pricing power. Exclusive architecture access is a double-edged sword. The manufacturer controls the tap. The bullish retail narrative paints "compute landlord" as recurring rental income, a digital real estate trust with a moat. The uncomfortable reality is that SpaceX owns xAI, the developer of Grok, and rents the same infrastructure to Anthropic, one of Grok's direct competitors. That is not a moat; it is a conflict. No contractual firewall fully fixes scheduling priority, data isolation, or access fairness when the landlord also trains a frontier model in the same building. Anthropic has publicly warned about AI compute concentration. Now it is paying a concentrated landlord that competes with it. If either anchor tenant walks, the revenue base is gutted. The satellite piece is equally double-edged. One million satellites would be the largest orbital constellation in history and the first orbital AI inference network. The competitive barrier is real - no one else can launch a million spacecraft - but so is the geopolitical response. A US company deploying orbital AI compute will trigger export control debates, spectrum conflicts, orbital debris disputes, and national security reviews. The same reason investors love the moat is the reason regulators will attack it. At current valuations, the market is not paying for today's $3.1 billion quarterly AI revenue. It is paying for a 2030 scenario where Google and Anthropic renew at a premium and Vera Rubin ships on schedule. That scenario requires no tenant defection, no depreciation panic, no Nvidia delay, and no regulatory intervention. Each assumption is a call option with negative carry. Smart money doesn't celebrate lease signings; it reads the early-termination clause. Sentiment buys the dip; data fills the position. The data I need is a unit economics table: per-GPU power cost, utilization, renewal pipeline, and EOL salvage value. Until that table exists, SpaceX is a borrower with a rental claim on future Nvidia deliveries. I learned this in 2020 when a 45% APY yield farm turned into a 12% exit. The yield was real until the collateral stopped behaving. The lease is real until the tenant stops paying. Can a landlord own both the building and the competing restaurant? That is the only question that matters for the long side.