On July 1, Reflection AI's compute lease with SpaceX went live: $6.3 billion through 2029, at $150 million per month, for access to Nvidia's flagship GB300 chips at the Colossus 2 data center near Memphis. Reflection joins Anthropic, Google, and Cursor as tenants of the same facility — which now reportedly carries more than $80 billion in committed outside revenue through 2029. One building in Tennessee has quietly become load-bearing infrastructure for a meaningful slice of the AI industry.
Who's writing these checks
Reflection AI was founded in 2024 by two Google DeepMind veterans: Ioannis Antonoglou, a founding DeepMind engineer who worked on AlphaGo, AlphaZero, and MuZero, and Misha Laskin, who led reward modeling on Gemini. The company positions itself as a leading open-source AI lab — and Nvidia has invested $800 million directly into it.
Sit with that structure for a second: Nvidia is an investor in Reflection, and Reflection is leasing Nvidia GB300 chips through SpaceX. The same handful of players — chip makers, hyperscale compute providers, well-funded labs — keep appearing on multiple sides of these deals, creating financial interconnections that antitrust regulators are already circling. When your supplier is also your investor and your landlord's supplier, "arm's length" gets blurry.
Why the deals keep getting bigger
Model efficiency keeps improving, and the checks keep growing anyway — because the constraint has shifted. It's no longer just training runs; agentic systems demand serious inference compute at serving time, continuously, at scale. For labs without the balance sheet to build hyperscale data centers, long-term leasing is the only way to guarantee capacity on current-generation chips. For SpaceX, renting out Colossus capacity has become a business line arguably as strategically significant as anything else the company does.
The most interesting clause is the exit
Buried in the deal structure: after an initial three-month lock-in, either party can terminate with 90 days' notice. That's a sharp departure from the multi-year, hard-to-exit compute commitments of earlier AI infrastructure deals, and it reads like an industry that learned something — chip generations, pricing, and AI company fortunes all shift fast enough that a rigid 2029 commitment is a liability for both sides.
Even if you'll never lease a GPU cluster, that structure is a useful template for your own cloud commitments: take the discount for committed capacity, but price in the option to walk when the technology or your needs change. The most sophisticated compute buyers in the world just decided flexibility was worth paying for.
What this means if you just buy API access
If your exposure to all this is calling models from Reflection, Anthropic, or Google over an API, three things flow downstream to you:
- Pricing reflects these economics. Your per-token costs sit on top of nine-figure monthly leases. Providers with committed compute runway can sustain aggressive pricing; providers scrambling for capacity can't.
- Concentration is systemic risk. Four major AI providers renting from one Memphis facility means a serious disruption there — power, physical, or otherwise — could degrade multiple model vendors simultaneously. "Multi-vendor" resilience strategies quietly assume the vendors don't share a building.
- Committed compute predicts capability. A lab locking in years of GB300 access is telling you it plans to keep training and serving frontier models. That's a real signal when assessing which vendors will still be competitive at your next contract renewal.
The takeaway for enterprise IT: the AI vendors you're evaluating are themselves managing enormous infrastructure dependencies and financial commitments, and those dependencies increasingly converge on the same chips and the same facilities. When you do your next AI vendor risk assessment, ask where their compute actually lives — the answer is more concentrated than the market's brand diversity suggests, and that concentration is your risk too.