Starting today, Reflection AI is paying SpaceX $150 million per month for access to Nvidia GB300 chips at Colossus 2 in Memphis, Tennessee. The contract runs through 2029 — $6.3 billion if it runs to term, with a 90-day exit option after the first quarter.
Reflection has not shipped a frontier AI model. Its only public product, Asimov, is a code comprehension agent that launched to a waitlist on July 16, 2025. Nvidia contributed $800 million to a $2 billion round in October 2025 that valued Reflection at $8 billion, and by March 2026 the company was in talks to raise fresh capital at a $25 billion valuation — and it just locked up GPU time at the same facility serving Anthropic and Google.
That’s the anomaly worth understanding. Not whether Reflection will win, but what the structure of this bet tells you about how AI infrastructure is being allocated right now.
Who Reflection AI Is
Reflection was founded in March 2024 by two former Google DeepMind researchers:
Ioannis Antonoglou is the company’s co-founder, president, and CTO. He joined DeepMind in 2012 as its 25th employee, wrote an early DQN implementation used in the lab’s Atari-playing deep reinforcement learning research, and worked on AlphaGo, AlphaZero, and MuZero, which mastered Go, chess, shogi, and Atari without being told the rules. He later led the RLHF effort and reward modeling for Gemini.
Misha Laskin is the CEO. After a Berkeley postdoc in Pieter Abbeel’s reinforcement-learning lab, he joined DeepMind’s Toronto general agents team and collaborated with Antonoglou on reward models for Gemini.
Their thesis: the Western open-source AI ecosystem lacks a frontier-scale model trained with serious reinforcement learning. DeepSeek-R1, released in January 2025 with open weights and a public technical report, showed what a well-resourced team could build without closed-lab restrictions. Reflection is betting it can do the same for the US market, with open weights released for enterprise self-hosting while keeping training data and pipelines proprietary.
The company has been deliberate about what it has not released. Asked in a late-February 2026 interview how its model would look, Antonoglou said: “You have to wait for that.” As of that interview, Reflection had published no research papers, no model benchmarks, and no training methodology disclosures for its in-progress frontier model.
The Deal Structure
The terms signed on June 22, 2026 and effective today:
- $150 million per month, starting July 1, 2026
- Through the end of 2029 (approximately 42 months)
- $6.3 billion total if the contract runs its full term
- Exit clause: either party can terminate with 90 days notice after the initial three months
- Hardware: Nvidia GB300 chips at SpaceX’s Colossus 2 data center in Memphis, Tennessee
For context, SpaceX now has commitments from three major AI tenants leasing Colossus compute:
| Tenant | Estimated deal size | Monthly rate | Source |
|---|---|---|---|
| Anthropic | ~$45B through May 2029 | ~$1.25B/month | Teslarati, SpaceX S-1 filing |
| ~$29.5B through June 2029 | $920M/month | TechCrunch | |
| Reflection AI | $6.3B through 2029 | $150M/month | TechCrunch |
A fourth name that sometimes gets lumped into this list, Cursor (Anysphere), is not actually a compute lessee: SpaceX acquired Cursor outright for $60 billion in an all-stock deal that closed its merger agreement on June 16, 2026, folding it in as a subsidiary rather than a paying tenant.
Summing the three leases above puts SpaceX’s disclosed AI compute contracts at roughly $80 billion. The Anthropic and Google deal terms were confirmed in SpaceX’s own S-1 IPO filing with the SEC, filed May 20, 2026, not just company statements. Colossus has become a commercial compute platform, not just an internal training cluster for xAI.
The Nvidia Angle
Nvidia contributed $800 million to Reflection AI’s October 2025 funding round. It also manufactures the GB300 chips that SpaceX is deploying at Colossus 2 and leasing to Reflection.
This is a structurally interesting position. Nvidia profits from the chip sales to SpaceX, profits from its equity stake in Reflection, and benefits from Reflection’s research validating the GB300 architecture at frontier scale. Whatever happens to Reflection as a company, Nvidia has already captured value on two of the three sides of the trade.
This is not unusual for Nvidia — it has taken similar equity-plus-compute-purchase positions with OpenAI, Anthropic, CoreWeave, and other AI infrastructure customers, a pattern industry press has dubbed “circular financing” — but it is worth noting when evaluating Reflection’s valuation. Some of the capital flow feeding that $25 billion figure is circulating back through the same ecosystem.
What SpaceX Is Becoming
The Colossus data center started in 2024 as xAI’s internal training facility for Grok, built out of a converted Memphis factory in 122 days. SpaceX and xAI merged in February 2026 into a combined entity valued at $1.25 trillion, which is why SpaceX — not xAI — is now the counterparty signing these compute leases. In roughly a year and a half, Colossus has gone from an internal Grok training cluster to the preferred large-scale compute destination for multiple frontier labs that are not xAI.
The pattern suggests something important about what SpaceX offers that hyperscalers don’t. SpaceX has leaned on its own mobile natural-gas turbines to power the site rather than waiting on grid interconnects, and has used a vertically integrated, fast-build construction approach to bring clusters online in months rather than years. Leasing a “slot” at Colossus is closer to leasing a supercomputer than renting cloud instances.
Some analysts and press have started using the term “neocloud” for this model: private compute infrastructure operating at cloud-equivalent scale, accessible via long-term lease rather than metered API calls.
The Anthropic deal was the proof of concept. The Google deal confirmed it wasn’t a one-off. The Reflection deal shows that even startups — very well-funded ones — are choosing dedicated clusters over traditional cloud for frontier training runs.
What Reflection Is Actually Building
The company’s stated product goals:
A frontier open-weight general model: Reflection has said it is training with reinforcement learning at scale toward a model competitive with closed labs, in the broad tradition DeepSeek demonstrated — though, per its own founders, it has not disclosed its specific architecture or training methodology. CEO Misha Laskin has said the plan is to release model weights openly while keeping training data and the training pipeline proprietary, reasoning that “the model weights, because the model weights anyone can use and start tinkering with them,” while the infrastructure stack stays limited to companies with the resources to run it.
Asimov: Their only shipped product. An agentic coding system that ingests a company’s codebase, architecture docs, GitHub threads, and Slack/chat history to understand architecture context and answer questions about how systems are built. As of early March 2026, it remained in invite-only waitlist status.
Sovereign AI deployments: Enterprise and government customers who need to self-host a frontier model without dependency on US or Chinese closed-lab infrastructure. Reflection’s deal with South Korea’s Shinsegae Group to build a 250MW sovereign AI data center — reported at roughly 10 trillion won, or about $6.8 billion, announced in March 2026 — sits in this category.
The company has not shared benchmarks comparing its model to GPT or Claude. It has not published papers on training methodology. The $6.3B compute deal is not evidence of a shipped model — it is evidence of intent to train one at frontier scale.
Builder Implications
Three things worth tracking if you are building on or around frontier AI:
1. The open-weight frontier is getting serious compute
Until recently, the gap between open-weight models (LLaMA, DeepSeek) and closed-source frontier models (GPT-5, Claude Opus) was partly a compute gap. DeepSeek narrowed it with efficiency. Reflection is trying to close it with raw scale — the same scale Anthropic and Google are using.
If they ship a competitive open-weight model trained on $6B+ of compute, the calculus for self-hosting changes. A model competitive with Opus or GPT-5 that you can run inside your own infrastructure alters security posture, latency, and cost models for enterprise deployments.
2. Compute access is now a strategic resource, not a commodity
The Reflection deal starting today illustrates a supply constraint that is easy to underestimate. There are a small number of facilities globally with the power, cooling, and interconnect needed to train frontier models. Anthropic, Google, xAI, and now Reflection are locking up long-term capacity at them.
This means that if you are a mid-tier AI company needing GPU time at frontier scale in 2027 or 2028, you may be competing against contracts signed in 2025 and 2026. The capacity is not infinitely elastic.
3. The “no product shipped yet” valuation pattern is not accidental
Reflection was in talks for a $25 billion valuation without a flagship model. This is not irrationality — it is the market pricing the cost and difficulty of training frontier models correctly. The infrastructure required to enter the frontier tier is so large (compute contracts, energy, talent) that investors are buying into the option before the output exists, because waiting until the model ships means waiting until the moat is already built.
For builders, this has a practical implication: if you are waiting for a Reflection-style open-weight frontier model before building on top of open weights, you may be waiting a year or more. The better bet is to design your system to be model-agnostic at the API layer, so you can switch when the open-weight frontier catches up.
Reflection AI ships its first major product today — not a model, but a compute contract. Whether the model follows in 2026 or 2027 will determine whether this was the right call.
Correction policy: If Reflection releases model benchmarks or a public product between publication and your read date, treat the specifics here as a snapshot from July 1, 2026.