SpaceX Just Inked a New AI Deal Worth Up to $6.3 Billion

SpaceX agreed to a $6.3 billion AI infrastructure deal with Reflection AI, under which Reflection will pay $150 million per month starting July 1 for access to Nvidia GB300 chips at SpaceX’s Colossus data center. Reflection AI, founded in 2024 and backed by a $2 billion Nvidia investment, is growing rapidly and positioning itself as an open-source rival to OpenAI and Anthropic. The contract supports SpaceX’s push to monetize its AI ambitions, though the article notes its AI division still posted $6.35 billion in 2025 losses and $2.47 billion in Q1.

Analysis

This is less about one contract and more about the emergence of a spot market for frontier compute. The most important second-order effect is that NVDA demand is no longer just hyperscaler capex plus sovereign AI; it is now being monetized through intermediaries who can arbitrage scarce GB300 supply into high-margin rental economics. That should support near-term GPU utilization and pricing power, but it also means the ecosystem is becoming more levered to customer concentration and cancellation risk if end-demand does not scale quickly enough.

For the platform owner, the deal helps validate the asset layer, but the economics are still asymmetric: compute revenue can ramp immediately while the model/application layer remains structurally loss-making and highly volatile. A 90-day termination clause after the initial period makes this look more like an option on capacity than a durable annuity, so investors should not extrapolate the headline annualized value into stable EBITDA without discounting churn and underutilization risk. The real catalyst over the next 1-3 quarters is whether this becomes the first of a multi-deal cluster that meaningfully de-risks the company’s AI capex curve.

The market may be underpricing the signal for private-market AI and overpricing the durability of the open-source narrative. Open-source model builders tend to attract talent and research attention, but monetization often lags until enterprise distribution or managed inference is proven; that creates a gap where valuation can outrun revenue quality. Meanwhile, the winning trade is not necessarily the start-up itself but the infrastructure vendors that capture recurring spend before model economics are validated.

A contrarian risk is that this is a prestige deal dressed up as demand proof. If the startup’s growth stalls or broader AI budgets tighten, this could reverse quickly and expose spare capacity just as GPU supply normalizes into 2026. In that scenario, the market’s current willingness to pay up for AI infrastructure could compress fastest in the names most exposed to third-party rental economics rather than direct, contracted enterprise demand.

Key Decisions for Investors

• Long NVDA on 1-3 month horizon into any post-headline pullback; this deal reinforces GB300 scarcity and should support pricing/utilization, but trim if implied AI supply tightness starts to normalize faster than expected.
• Overweight GOOGL as a relative beneficiary versus open-source challengers on a 6-12 month view; if open models gain adoption, distribution and cloud inference remain the monetization choke point, which favors platform incumbents.
• Avoid chasing private AI application names at current marks; prefer a short-dated call spread or no position until there is evidence of repeatable enterprise revenue rather than one-off capacity commitments.
• Pair trade: long NVDA / short a basket of high-multiple AI software names most dependent on narrative rather than revenue durability; the market is still paying for story while the compute layer has the cleanest near-term cash conversion.
• Set a watchlist alert for any second and third capacity deal over the next 60-90 days; if the contract cluster does not materialize, fade the infrastructure enthusiasm and expect mean reversion in AI beta.