SpaceX signed a $6.3 billion agreement with Reflection AI to provide Nvidia GB300 computing power through 2029, days before the company finalizes its IPO pricing structure. The contract commits SpaceX to deliver dedicated AI infrastructure on a multi-year schedule, marking the first disclosed customer agreement for SpaceX's ground-based compute operations separate from its satellite launch and Starlink businesses.

Reflection AI, a late-stage model training firm with backing from Andreessen Horowitz and Sequoia, requires sustained access to Nvidia's GB300 architecture for multi-modal large language model training. The deal includes phased capacity delivery starting Q3 2025, with initial deployment of 12,000 GB300 units by year-end and staged expansion through 2027. Reflection's contract terms include uptime guarantees above 99.7%, suggesting SpaceX built redundant cooling and power infrastructure at its Texas facilities. The agreement references co-location at SpaceX's Boca Chica site, where recent satellite imagery shows new HVAC installations and transformer capacity consistent with hyperscale data center requirements.

This positions SpaceX as an infrastructure lessor in AI compute, not merely a satellite operator with edge compute optionality. The $6.3 billion figure exceeds SpaceX's total 2023 launch revenue of $4.7 billion, indicating Musk expects ground-based AI infrastructure to rival orbital services within five years. The timing—announcement two weeks before IPO allocation—signals to institutional buyers that SpaceX diversified revenue beyond NASA contracts and Starlink subscriptions. Goldman Sachs, lead underwriter on the IPO, will now model a third revenue stream in valuation decks circulating to family offices and sovereign funds this week. The contract also locks in Nvidia supply ahead of GB400 architecture availability in 2026, insulating SpaceX from spot-market volatility as Nvidia shifts production lines.

Allocators should monitor three follow-on events. First, whether SpaceX discloses additional AI infrastructure customers before IPO pricing closes, expected by April 18. A second contract above $2 billion would establish SpaceX as a Tier-1 compute lessor alongside CoreWeave and Lambda Labs. Second, Nvidia's Q2 earnings call on May 28, where management may reference GB300 shipment volumes to non-cloud customers—indirect confirmation of SpaceX's capacity scale. Third, Reflection AI's Series D raise, rumored for June, where valuation will reflect secured compute access. If Reflection prices above $8 billion post-money, it validates the strategic premium of multi-year Nvidia capacity in current market conditions.

The deal converts SpaceX's real estate and power infrastructure into a balance-sheet asset before public-market scrutiny. Boca Chica's 150 MW grid connection, originally permitted for rocket testing, now generates recurring revenue with margins exceeding launch services. The IPO roadshow begins April 7.