Reflection AI just committed $150 million per month to SpaceX for access to Nvidia GB300 chips at the Colossus 2 data center — $6.3 billion through 2029. That is the largest infrastructure bet ever placed on open-weight frontier AI, and it was placed by a company that has never shipped a public model. If you currently pay Anthropic or OpenAI per token, this deal exists specifically to eventually end that arrangement.
SpaceX Is Now AI’s Landlord #
After the SpaceX-xAI merger in February 2026, Colossus became something nobody planned: the backbone of the entire AI industry. Anthropic rented all of Colossus 1 — 222,000 Nvidia GPUs, 300 megawatts — for $1.25 billion per month, a $45 billion commitment through 2029. Google followed in early June, signing a $30 billion deal for Colossus 2 capacity at $920 million per month. Now Reflection AI is joining them at $150 million per month for GB300 access through 2029.
Total committed revenue from Colossus through 2029: over $80 billion. A data center that started as Elon Musk’s Grok training cluster now generates more annual recurring revenue than most Fortune 500 companies. The world’s most important AI compute is concentrated in Memphis, Tennessee, and three companies — Anthropic, Google, and a startup most developers haven’t heard of yet — are betting their futures on it.
Who Is Reflection AI? #
Reflection was founded in March 2024 by two former Google DeepMind researchers. Misha Laskin (CEO) led reward modeling for Gemini. Ioannis Antonoglou (CTO) was DeepMind employee number 25 — the researcher who built the neural networks inside AlphaGo. Between them, they have more direct experience building frontier models than most AI labs in existence.
The funding trajectory is hard to ignore: $545 million valuation in March 2025, $8 billion by October 2025 after a $2 billion raise, and roughly $25 billion by March 2026 as Nvidia pumped in approximately $800 million across multiple rounds. The valuation grew 46x in under twelve months.
The uncomfortable fact: as of June 2026, Reflection has not shipped a public model. The company has been building in private, working with government customers, training on data at a scale it describes as “tens of trillions of tokens.” The $6.3 billion compute deal with SpaceX is the first concrete, auditable signal that the model is actually coming. You don’t lock in GB300 access for three years unless you plan to use it.
The Developer Case for Open-Weight Frontier Models #
The cost math is the starting point. Running a typical production workload on GPT-5.2 costs around $2,275 per month. The same workload on a self-hosted open-weight model costs approximately $168 per month. That is a 13x difference, and that is before accounting for the infrastructure cost of keeping your own API dependency alive through model deprecations, outages, and rate limit changes.
A frontier-class open-weight model collapses that cost structure further. You download the weights. You run it on your own hardware or any cloud instance. You pay $0 per token. You fine-tune it on your own data without sending that data anywhere. You can air-gap it completely. For regulated industries, government contractors, and any team that has been burned by an API access cut, that is a fundamentally different risk profile from what Anthropic or OpenAI currently offer.
The problem until now: the open-weight models available in 2026 — Meta’s Llama 4, Qwen 3, Mistral — are powerful and fast, but none of them match frontier-class performance against GPT-5.5 or Claude Fable 5 in demanding workloads. DeepSeek V4 comes closest, but its Chinese origin creates sovereign AI risk that disqualifies it for US government work and many enterprise deployments. Reflection is targeting the specific gap: frontier performance, open weights, American-made, Pentagon-cleared.
The Pentagon Signal #
In May 2026, the Defense Department cleared eight companies to deploy AI on its classified networks at Impact Level 6 and Impact Level 7 — the highest classification tiers in US government infrastructure. The list: AWS, Google, Microsoft, OpenAI, SpaceX, Nvidia, Oracle — and Reflection AI. Reflection is the only startup on that list. Every other name is a hyperscaler or established lab with decades of government contracting history.
That clearance matters beyond government contracts. IL6/IL7 certification signals that Reflection’s security practices, infrastructure controls, and model governance meet standards that no other open-weight AI developer has cleared. For enterprise teams in finance, healthcare, and defense contracting, that certification matters as much as the model’s benchmark scores.
Timeline and What to Watch #
Reflection’s GB300 access starts July 1, 2026. The company’s stated bottleneck is no longer compute — it is data-pipeline engineering. Late 2026 or early 2027 is the realistic window for a public model release, based on company statements and the compute timeline now locked in.
The risk is real: a $25 billion valuation for a company with no shipped model is fragile. If the model underperforms GPT-5.5 or Fable 5 on release, the compute bet becomes a very expensive lesson. But the evidence is accumulating in the other direction. Pentagon clearance, Nvidia’s $800 million investment, and $6.3 billion in GPU time are not things you assemble around vaporware.
Set up a benchmark pipeline now so you can evaluate the model on release against your actual workload. Reflection’s public blog is the best place to track progress. For deal specifics, TechCrunch’s coverage and CNBC’s breakdown of the Colossus terms have the details worth understanding now. For a deeper look at Reflection’s strategy, the Turing Post inside look is the most complete independent analysis available. And Axios frames why open-source AI needs this scale of compute to compete with closed frontier labs.
The open-source alternative to GPT-5.6 and Claude Fable 5 is being assembled in Memphis, on SpaceX’s Nvidia GB300 cluster, by two DeepMind veterans with Pentagon clearance and $25 billion in backing. The model isn’t here yet. But the bet is now large enough to take seriously.