Anthropic is paying $15 billion a year for access to Elon Musk’s data centers

SpaceX disclosed in its IPO filing that Anthropic will pay $1.25 billion per month, or $15 billion annually, through May 2029 for access to its Colossus AI training centers. The deal includes a 90-day termination clause and reduced fees during the current capacity ramp-up, highlighting rising demand for large-scale compute. SpaceX also said its AI business spent $12.7 billion on capex in 2025 and lost $6.3 billion on $3.2 billion of revenue, underscoring the strategic but costly push into AI infrastructure.

Analysis

This is less a single contract than a market signal that compute scarcity remains severe enough to support take-or-pay economics at hyperscale. The biggest second-order effect is that AI training capacity is becoming increasingly financeable like infrastructure, which should widen the gap between owners of scarce power/land/permitting and model developers forced to prepay for access. That favors the physical stack—data center REITs, power equipment, grid interconnectors, and gas-fired generation—more than the pure model layer, which now looks increasingly capacity-constrained rather than demand-constrained.

The hidden risk is concentration and termination optionality. A 90-day out clause means the revenue stream is not as bond-like as the headline annualized figure suggests, so the market should discount it heavily until there is proof of utilization durability and payment discipline. Over the next 1-3 quarters, the key catalyst is whether other AI labs sign similar contracts; if they do not, this may expose a winner-take-most dynamic where only the most heavily capitalized frontier labs can secure enough compute, pressuring smaller peers to delay model releases or raise dilutive capital.

For public markets, the cleaner trade is not long the model winner but long the infrastructure bottleneck. If AI capex continues to migrate from software R&D into power, cooling, and buildout, earnings revisions should accrue to the picks-and-shovels ecosystem before they show up in the large-cap AI names. The contrarian view is that this may actually be a peak scarcity print: once enough private capital rushes into dedicated campuses, marginal returns on incremental GPU deployment could compress sharply, and the first sign of overbuild would be contract renegotiations or slack capacity in 2H26.

A subtle positive for the ecosystem is that a massive external customer validating third-party compute monetization lowers the hurdle for other operators to pre-sell capacity, which should improve project finance for the entire data-center supply chain. That said, if compute gets commoditized faster than expected, the bargaining power shifts back to the model vendors, and the current pricing power could prove temporary rather than structural.

Key Decisions for Investors

• Long VRT / long ETN into the next 1-2 quarters: buy the power, thermal management, and electrical bottleneck rather than model-layer exposure; target 15-20% upside if AI capex remains elevated, with a tighter stop if hyperscaler capex guides down.
• Long DLR or EQIX vs short a basket of overowned AI software names over 3-6 months: the thesis is that contracted compute demand monetizes real estate and interconnect more reliably than frontier-model equity stories; expected outperformance if scarcity persists.
• Buy SMR or GEV on pullbacks as a 6-12 month power-supply proxy: data-center load growth should keep nuclear/gas/cooling capex bid; use options where available to limit downside if the contract proves non-recurring.
• Avoid chasing pure-play AI model beneficiaries after the headline; if anything, use rallies to fade names where valuation assumes indefinite compute scarcity monetization without owning infrastructure.
• If listed proxies to this theme exist in your book, consider a pair: long infrastructure enablers / short a basket of high-multiple AI application names for a 2-4 month relative-value expression.