SpaceX Colossus signs a $6.3 billion deal with Reflection AI

SpaceX Colossus signs a $6.3 billion deal with Reflection AI

🔎 $150 million per month for a startup with no product

On June 22, 2026, SpaceX announced a compute contract with Reflection AI valued at $6.3 billion over three years. The amount is colossal. But the detail that catches the eye: Reflection AI has not yet released any open-source model.

The deal provides for $150 million per month starting July 1, 2026, for access to Nvidia GB300 chips in the Colossus 2 data center near Memphis. The contract runs until 2029, with a 90-day termination clause after the first three months. It is not a firm long-term commitment, but it is the largest infrastructure contract ever signed by an open-source lab.

The timing is not insignificant. The US government just banned Anthropic's closed models (Fable and Mythos), creating a strategic void that players like Reflection intend to fill.

Key points

• Reflection AI will pay $150M/month to SpaceX for Nvidia GB300 chips on Colossus 2, totaling $6.3B if the full term is honored (CNBC, June 2026).
• Colossus 2 is unified Blackwell-only, designed specifically for frontier training, unlike Colossus 1 which was dedicated to inference.
• SpaceX becomes the world's largest AI compute provider, surpassing traditional cloud providers thanks to contracts with Anthropic ($1.25B/month), Google ($920M/month), and now Reflection.
• Reflection AI, founded in 2024 by two former Google DeepMind employees, is valued at $25B and raised $2.5B in 2026, but has no public model.
• The deal is set against a backdrop of geopolitical tension around open-source, following the US government's ban on Anthropic's closed models.

Recommended tools

| Hostinger | Web hosting for deploying AI apps | Starting at 2.99€ (June 2026, check on hostinger.com) | AI startups and developers |

The deal numbers — $150M/month, but with an exit clause

The headline amount is $6.3 billion. The contractual reality is more nuanced.

Reflection commits to paying $150M/month starting July 1, 2026. The contract provides for a duration of 42 months (until 2029), which indeed amounts to $6.3G in full terms. But Elon Musk himself downplayed the significance of the three-year term, emphasizing that Colossus contracts are cancellable.

The 90-day termination clause after the first three months changes the game. In practice, Reflection can test the infrastructure during the summer of 2026 and bow out in the fall if the results don't follow. It's a calculated risk for SpaceX, but a limited one: the GB300 chips are already there, and excess capacity will find a buyer.

To put things in perspective, here is the pricing structure of known Colossus contracts:

The Reflection contract is the smallest of the three, but it is strategically significant because it opens Colossus to an open-source player. This positioning is detailed in our analysis of the $6.3 billion SpaceX × Reflection AI deal.

Colossus 2 vs Colossus 1 — two infrastructures, two missions

SpaceX does not manage a single homogeneous data center. Colossus 1 and Colossus 2 have radically different architectures and objectives.

Colossus 1, initially built by xAI in 122 days, is now entirely dedicated to Anthropic. With over 220,000 GPUs and a consumption of 300 MW, it is used for the inference of Claude models. Anthropic uses a mixed architecture that optimizes latency rather than training throughput. This Anthropic-SpaceX contract for Colossus 1 represents the largest single agreement on the platform.

Colossus 2, on the other hand, is unified Blackwell-only. It is designed specifically for frontier training — the training of next-generation models. It is on this infrastructure that Reflection AI will work with Nvidia GB300 chips.

The difference is not trivial. Frontier training requires maximum inter-GPU bandwidth, specific network topologies (NVLink, InfiniBand), and stable power supply at the gigawatt scale. Wikipedia notes that Colossus is currently expanding towards 2GW of total capacity.

For current models like GPT-5.5 (agentic score of 98.2) or Gemini 3 Pro Deep Think (95.4), the training has already taken place on previous infrastructures. But the next wave — the models that will break the 100 barrier on the agentic benchmark — will require precisely this type of platform.

Reflection AI — who is this $25 billion startup?

Reflection AI was founded in 2024 by two researchers who left Google DeepMind. In two years, the startup has raised $2.5 billion and reached a valuation of $25 billion. Nvidia is among its investors.

The problem: there is no public open-source model signed by Reflection. Zero. No weights published on Hugging Face, no API available, no independent benchmarks.

What is public, however, are its government contracts. Reflection is working with the Department of Energy on the "Genesis Mission" and collaborating with the Pentagon. The company describes its approach as "American open intelligence" — a deliberate oxymoron that blends technological openness with alignment with US national interests.

The strategy resembles that of OpenAI in its early years: raising massively, committing to infrastructure before having a product, and betting on talent to make up for lost time. Except that OpenAI had published GPT-2 and GPT-3 before signing multi-billion-dollar deals. Reflection is skipping the public proof-of-concept stage.

The bet is risky but not irrational. The US government's banning of Anthropic's closed models (reported by Axios in June 2026) creates a structural opportunity. If Reflection can deliver a high-performing open-source model within 6 to 12 months, it will capture a market that is closing off to proprietary models.

Musk's strategy — turning xAI into the AWS of AI

The absorption of xAI by SpaceX in February 2026 was not a simple internal restructuring. It was a strategic pivot.

xAI had built Colossus to train Grok. But internal training efforts had "weakened," according to TechCrunch. Rather than leaving the infrastructure underutilized, SpaceX repositioned Colossus as a commercial compute platform. The logic is simple: why risk billions on a single proprietary model when you can rent the capacity to multiple clients and guarantee recurring revenue?

SpaceX's $60 billion acquisition of Cursor fits into the same vertical logic. Cursor provides the application layer (the AI code editor), Colossus provides the infrastructure layer, and clients like Anthropic, Google, and Reflection pay for the compute.

In parallel, OpenAI signed a $300 billion deal with Oracle Cloud to make GPT and Codex accessible via existing credits. The message is clear: the AI battle is now playing out on infrastructure, not just on models.

With $1.25B/month (Anthropic) + $920M/month (Google) + $150M/month (Reflection), SpaceX potentially generates $2.32 billion in monthly recurring revenue on Colossus. That is more than most traditional cloud providers in their AI segment.

Forbes analyzes this positioning as a direct attempt to compete with AWS, Azure, and Google Cloud on AI compute. Except SpaceX starts with an advantage: its construction costs are lower (Musk has always optimized engineering costs), and its rapid deployment capability is proven (122 days for Colossus 1).

What the deal means for open-source AI

The Reflection contract is the largest open AI infrastructure commitment announced to date, according to TechCrunch. It's a strong signal for the ecosystem.

Until now, the training frontier was the exclusive domain of closed labs. Open-source models like DeepSeek V4 Pro (88 on average, according to June 2025 benchmarks) or Kimi K2.6 (84) were trained on proprietary or ad-hoc rented infrastructures. None had guaranteed access to a Colossus-level platform for three years.

The situation changes the competitive dynamic. If Reflection manages to produce an open-source model trained on GB300s at $150M/month, it could theoretically rival top-tier proprietary models. The score of GPT-5.5 (98.2 agentic) or Claude Opus 4.7 Adaptive (94.3) would serve as a benchmark.

But there is a major caveat. Compute isn't everything. Anthropic cut access to Fable and Mythos not out of a lack of power, but for security and regulatory compliance reasons. A powerful but poorly aligned open-source model does not solve the fundamental problem the government is seeking to address.

Reflection implicitly understands this, hence its "American open intelligence" label. The idea is to combine the transparency of open-source with safeguards acceptable to federal agencies. It's a delicate balance, and the market does not yet have a clear precedent for success with this model.

The geopolitics of compute — why the Pentagon is interested

The involvement of the Department of Energy and the Pentagon in Reflection's contracts is not anecdotal. It reflects a reality: compute has become a matter of national sovereignty.

The United States recently banned Anthropic's closed models (Fable/Mythos) for reasons that have not been fully published, but which seem linked to security and control concerns. This decision immediately strengthened the argument in favor of open-source: an open model is auditable, modifiable, and does not depend on a single company for its updates.

The Reflection-Colossus deal fits into this logic. By indirectly funding (via government contracts) Reflection's access to cutting-edge compute, the US government supports the creation of an American open-source alternative. This is as much a defensive strategy as an offensive one: defensive against the risks of closed models, offensive against Chinese competition in the field of open-source AI.

The question is whether $150M/month is enough. Training a frontier model requires thousands of GPUs for months. The GB300s of Colossus 2 are among the most powerful chips in the world, but the quantity allocated to Reflection (not publicly specified) will determine the speed and quality of the result.

Risks for each party

No deal of this magnitude is without risk. Let's examine the potential pitfalls.

For Reflection AI

The main risk is execution. The company has never released a model. Going from zero to a frontier model in a few months on rented infrastructure is a major technical challenge. Even with ex-DeepMind talent, the probability of delays is high.

The financial risk follows: $150M/month is $1.8 billion a year. With $2.5B raised in 2026, Reflection has about 16 months of runway at the current rate. If no revenue-generating product emerges by the end of 2027, investor pressure will become intolerable.

For SpaceX

The risk is minor but real. If Reflection defaults or terminates after 3 months, SpaceX is left with unallocated GB300 capacity. Given current demand (Anthropic and Google are absorbing the bulk), this is a manageable risk. But at the scale of the planned 2GW expansion, every unmonetized megawatt weighs heavily.

For the market

The systemic risk is concentration. If SpaceX becomes the dominant compute provider, it controls a bottleneck. Anthropic, Google, and Reflection all depend on the same platform. A physical incident (power outage, natural disaster in Memphis) or a unilateral decision by Musk could paralyze a significant portion of the AI industry.

❌ Common mistakes

Mistake 1: Confusing Colossus 1 and Colossus 2

Colossus 1 is used for inference for Anthropic (220K+ GPUs, mixed architecture). Colossus 2 is Blackwell-only, dedicated to frontier training. Mixing the two in an analysis distorts the understanding of the issues. The solution: always check which client is on which infrastructure before comparing.

Mistake 2: Treating the $6.3B as a firm commitment

The contract includes a 90-day termination clause after 3 months. The amount of $6.3 billion is a theoretical maximum, not a guarantee. Musk himself pointed this out. The solution: present the figure with its conditions, not as a done deal.

Mistake 3: Equating "open intelligence" with "traditional open-source"

Reflection speaks of "American open intelligence", not open-source in the community sense. The contracts with the DOE and the Pentagon suggest access or usage restrictions. The solution: do not put Reflection in the same basket as Meta or Hugging Face.

❓ Frequently Asked Questions

How many GPUs will Reflection AI use on Colossus 2?

The exact number has not been published. What is confirmed: the chips are Nvidia GB300, the infrastructure is Colossus 2 (Blackwell-only), and the $150M/month amount suggests a significant allocation but less than the capabilities allocated to Anthropic or Google.

Why doesn't SpaceX keep Colossus for Grok?

xAI was absorbed by SpaceX in February 2026, and internal training efforts have decreased. Renting out the capacity generates more predictable recurring revenue than developing a proprietary model in the face of fierce competition (GPT-5.5, Claude Opus 4.7, Gemini 3 Pro).

What is the connection between Anthropic's ban and this deal?

The US government banned Anthropic's closed Fable and Mythos models, creating demand for open-source alternatives. The Reflection-Colossus deal arrives exactly in this window of opportunity, according to Axios analysis (June 2026).

Will Reflection AI release a model soon?

No date has been announced. The company has no public model to date. The compute deal suggests an intensification of training, but the delay between compute access and the release of a frontier model is typically 6 to 18 months.

✅ Conclusion

The $6.3 billion deal between SpaceX and Reflection AI is less of a bet on a product than on positioning: SpaceX is becoming the AWS of frontier AI, and Reflection is attempting to become the open-source champion that the US government is expecting. The figures are staggering, the risks are real, but the economic logic is sound. It remains to be seen whether Reflection can turn $150 million a month into a model worth checking out.