Reflection AI Locks In $6.3B SpaceX Compute Lease, Paying $150M a Month for Nvidia GB300 Access

On July 1, a two-year-old AI startup began wiring $150 million per month to Elon Musk’s SpaceX. The payment is not for satellites or rocket launches. It is for computing time.

Reflection AI, a San Francisco-based startup building open-weight frontier models, activated a long-term compute lease at SpaceX’s Colossus 2 data center near Memphis, Tennessee, under a deal first announced on June 22. The agreement runs through 2029 and totals approximately $6.3 billion if it goes the distance — though either party can terminate with 90 days’ notice after an initial three-month commitment period expires.

The numbers are significant by any measure. But the deal’s real importance lies in what it reveals about the shape of the AI industry in mid-2026: the ability to train frontier models has become less a software problem and more a logistics problem, and the companies that solve it are changing faster than most analysts expected.

Who Is Reflection AI?

Reflection AI was founded in 2024 by Misha Laskin and Ioannis Antonoglou, both veterans of Google DeepMind. Antonoglou’s credentials are formidable: he was one of DeepMind’s founding engineers in 2012 and is credited as a key contributor to AlphaGo, AlphaZero, and MuZero — the programs that sequentially conquered Go, Chess, and Shogi with superhuman performance.

The company emerged from stealth in March 2025 with a valuation of $545 million. By October of that year it had climbed to $8 billion, and by early 2026 it was pursuing a fundraising round at a $25 billion valuation. Nvidia, sensing a major future customer, invested $800 million directly in the company — a strategic bet that has made Nvidia simultaneously a supplier and a financial partner in Reflection’s growth.

Reflection’s stated mission is to build American open-source frontier models capable of competing with the best closed models from OpenAI and Anthropic, while positioning itself as a domestic alternative to Chinese open-source labs like DeepSeek. The compute deal with SpaceX is the infrastructure backbone of that mission.

What Colossus 2 Is, and Why It Matters

SpaceX’s Colossus 2 facility near Memphis is one of the largest dedicated AI compute clusters in the world. The original Colossus facility — built initially to train xAI’s Grok models — demonstrated that SpaceX could construct and operate massive GPU clusters at a speed that traditional hyperscalers could not match, leveraging SpaceX’s internal logistics, power procurement, and construction capabilities.

Colossus 2 is a commercial extension of that capability, designed from the outset to serve multiple tenants. Within roughly two months of opening it to third-party clients, SpaceX has secured committed revenues exceeding $80 billion through 2029 — from Anthropic, Google, Cursor, and now Reflection. The facility runs Nvidia’s GB300 chips, the successor to the H100 and H200 that dominated earlier training workloads, offering significantly higher memory bandwidth and compute density.

The commercial compute business has become a meaningful revenue line for SpaceX outside its core aerospace operations. For an aerospace company, the diversification is notable; for the AI industry, SpaceX’s entry into hyperscale compute is a structural shift that gives a non-traditional player significant leverage over AI development timelines at some of the fastest-growing labs.

The Economics of the Deal

At $150 million per month, Reflection AI is committing to an annualized compute budget of $1.8 billion. That figure is larger than the total venture capital raised by many successful AI startups in their entire histories, and it must be sustained every month for the deal to run to its 2029 end date.

How does a startup at even a $25 billion valuation fund a $150 million monthly commitment? The answer involves multiple revenue streams that Reflection is building in parallel with its model development. The company is reportedly licensing access to compute-efficient inference for business clients while training its flagship open-weight models, generating cash flow that offsets training costs. Nvidia’s $800 million investment provides a capital buffer. And the 90-day exit clause in the SpaceX contract provides a safety valve: if Reflection’s financial position deteriorates, it can exit the deal without being trapped in a multi-billion-dollar obligation.

The contract structure also reflects how SpaceX approached the deal. Unlike a traditional cloud contract with variable usage fees, Reflection pays a flat monthly rate — a financial arrangement that gives SpaceX reliable revenue while giving Reflection predictable capacity and protection against spot-market GPU price spikes, which have been brutal for AI companies without long-term supply agreements.

An American Bet Against DeepSeek

The strategic subtext of the Reflection deal is geopolitical. The company has been explicit that one of its goals is to build a capable open-source frontier model that is American in origin and governance — a counterpart to DeepSeek’s V3 and R1 models, which demonstrated that Chinese labs could achieve near-frontier performance at a fraction of the cost assumed by U.S. incumbents.

DeepSeek’s releases earlier in 2025 and 2026 rattled U.S. AI investment assumptions by showing that training efficiency could substitute for raw compute. Reflection’s response is to combine both: aggressive investment in training efficiency (Laskin’s research background at DeepMind centered on sample efficiency and reinforcement learning) combined with brute-force compute access that no efficiency shortcut can fully replace at the frontier.

The combination of $6.3 billion in committed compute time, $800 million in Nvidia backing, and a founding team with AlphaGo-era credentials positions Reflection as the most credible challenger in the open-source frontier model race that does not have Big Tech parentage. Whether a $25 billion valuation can be sustained on open-weight model releases — which by definition are free to use — remains the central business model question hanging over the company.

SpaceX as AI Infrastructure Kingmaker

The Reflection deal raises a structural question for the AI industry that has not yet been fully reckoned with: what does it mean for a single company — SpaceX — to hold committed compute contracts worth more than $80 billion from the organizations building the world’s most advanced AI systems?

SpaceX is not a neutral infrastructure provider in the way that AWS or Google Cloud are. It is controlled by Elon Musk, who also controls xAI, which competes directly with Anthropic, OpenAI, and Reflection for AI leadership. The governance questions this concentration creates — who has priority access in a capacity crunch, what data protections exist between tenants — have not been publicly addressed by any of the parties involved.

For now, the AI industry appears willing to accept those risks in exchange for the one thing Colossus 2 offers that no other facility can currently match: scale, speed of deployment, and access to the latest Nvidia silicon, all under one roof.