Reflection AI buys into SpaceX’s GPU cloud with a $6.3B compute deal

Reflection AI has signed a large compute agreement with SpaceX, giving the open-weight AI startup access to Nvidia GB300 chips at SpaceX’s Colossus 2 data center near Memphis, Tennessee.

The deal starts July 1, 2026. Reflection will pay $150 million a month through 2029, according to TechCrunch, putting the contract’s maximum value at $6.3 billion. Either company can terminate it with 90 days’ notice after the first three months.

That cancellation clause matters. AI infrastructure deals keep arriving with huge headline numbers, but the firm commitment can be much smaller than the total value implies. Even so, $150 million a month is serious money for any AI lab, especially one trying to position itself as an open alternative to closed frontier model providers.

Reflection joins a growing list of companies renting SpaceX’s AI infrastructure. Anthropic reportedly agreed to pay $1.25 billion a month for compute, while Google signed a $920 million-per-month deal. Reflection’s contract is smaller, but it points to the same problem: serious model training still depends on access to scarce high-end accelerators, and open-weight labs need industrial-scale infrastructure too.

Why GB300 access matters

The technical draw is Nvidia’s GB300 platform, the latest generation of its Grace Blackwell AI hardware. For frontier model training, that hardware affects training throughput, cluster efficiency, interconnect performance, power density, and the size of model runs a lab can realistically attempt.

Large language model training is constrained by several things at once:

• GPU memory capacity
• memory bandwidth
• inter-GPU networking
• storage throughput
• cluster scheduling
• power availability
• cooling
• software stack maturity

The chip matters, but the cluster matters more. A rack full of GB300s is useful only if the surrounding infrastructure can keep them fed and synchronized. Distributed training at this scale depends on fast networking, stable job orchestration, efficient checkpointing, and enough operational discipline to avoid expensive idle time.

That makes Colossus 2 interesting. The Memphis facility was originally associated with xAI’s internal model training efforts. Since xAI is now part of SpaceX, according to the source report, SpaceX has turned those GPU holdings into a rental business for other frontier labs.

It’s a practical pivot. If internal AI work stalls or doesn’t consume the full capacity, renting scarce accelerators is the obvious way to monetize the asset. The AI infrastructure market is supply-constrained enough that buyers will tolerate unusual arrangements if the hardware is available now.

Reflection’s open-weight bet gets more expensive

Reflection was founded in 2024 by two former Google DeepMind researchers and has pitched itself as an American open frontier AI lab. The company raised $2 billion in 2025, according to prior reporting, and has framed open-weight models as a strategic answer to closed systems from OpenAI, Anthropic, and others.

Open-weight models publish trained parameters so developers and enterprises can run, inspect, fine-tune, and deploy them outside the provider’s hosted API. A fully open source AI release may also include training code, datasets, evaluation tooling, and permissive licensing terms. The distinction matters. Open weights give builders far more control than a closed API, but they don’t automatically make a model reproducible or legally simple to use.

Reflection is using this compute deal to reinforce that position. The company said recent events show why open models matter, especially as governments and enterprises worry about depending only on closed systems. That argument gained more attention after the U.S. government’s ban on Anthropic’s closed Fable and Mythos models.

For technical buyers, the appeal is straightforward. Open-weight models reduce some platform risk:

• You can run inference in your own environment.
• You can fine-tune against private data without sending everything to a model vendor.
• You can inspect model behavior more directly.
• You can avoid some API pricing and availability surprises.
• You can build fallback infrastructure instead of depending on one hosted endpoint.

They also push work back onto the customer. Security reviews, deployment architecture, inference optimization, evaluation, monitoring, and abuse prevention become your problem. Strong AI engineering teams can handle that. Companies that want a managed API, support contracts, and predictable operations may find the trade-off less attractive.

Frontier compute is becoming a private club

The size of these contracts says something uncomfortable about frontier AI. Even open model development now requires capital and infrastructure that only a small group of companies can reach.

Training a competitive model has become a supply chain problem as much as a research problem. It requires chips, power, real estate, cooling systems, fiber, and large teams that can keep distributed workloads running across thousands or tens of thousands of accelerators.

That changes the meaning of open AI. A model can be released with open weights while the ability to create it remains concentrated. Reflection may publish models that developers can download and adapt, but those models still depend on billion-dollar infrastructure deals.

Open weights help downstream users. They don’t necessarily democratize frontier training.

The cancellation clause adds another risk. Reflection gets immediate access to top-end hardware, but the contract can be terminated after the initial period with 90 days’ notice. That gives both sides flexibility. It also creates planning risk. Training roadmaps for large models often stretch across quarters, and sudden capacity loss can break schedules, delay releases, or force expensive migrations.

For a lab trying to compete with closed incumbents, that’s a real operational caveat.

What developers should watch

Most developers won’t touch GB300 clusters directly. The effects will show up later in the models, pricing, licenses, and deployment options that come out of this compute race.

If Reflection’s strategy works, engineers could see stronger open-weight models that compete more directly with closed APIs in coding, reasoning, agentic workflows, and enterprise knowledge tasks. That would matter for teams building internal copilots, retrieval-augmented generation systems, evaluation pipelines, and domain-specific fine-tunes.

Open-weight releases can be especially useful when a team needs:

• low-latency inference close to users or data
• strict data residency controls
• custom fine-tuning
• offline or air-gapped deployment
• predictable per-token economics at scale
• deeper observability into model behavior

The trade-off is engineering load. Running open models well requires infrastructure competence. A model that looks cheap on paper can become expensive once you account for GPUs, serving frameworks, quantization, batching, monitoring, prompt safety, rollback systems, and evaluation.

For many teams, the practical answer will be a mix: closed models for the hardest general reasoning tasks, open-weight models for controlled workloads, smaller specialized models for high-volume internal automation, and retrieval systems to keep outputs grounded.

Reflection’s deal could improve that mix. It won’t make it simpler.

SpaceX is turning unused AI capacity into a business

SpaceX’s role is worth watching. Colossus began as infrastructure for Musk’s AI ambitions, but the company is now selling access to some of the most valuable compute on the market. Anthropic, Google, and Reflection are very different customers, which suggests SpaceX is acting less like an internal infrastructure provider and more like a high-end AI cloud supplier.

That’s good business if utilization holds up. Idle GPUs are financial dead weight. GB300 clusters are too expensive to leave underused.

The arrangement raises questions buyers will care about:

• How mature is the platform layer around the hardware?
• What service-level guarantees do customers get?
• How isolated are workloads between tenants?
• What are the data security and compliance terms?
• How does SpaceX prioritize jobs across major customers?
• What happens if internal SpaceX or xAI demand returns?

The public reporting doesn’t answer those questions. For AI labs spending hundreds of millions a month, the answers sit in contract language, operational dashboards, and hard lessons during failed training runs.

Hyperscalers have spent years building trust around cloud isolation, compliance, support, and procurement. SpaceX has chips and facilities. That may be enough in a constrained market, but enterprise-grade cloud behavior is harder than renting out racks.

Open models now share the same bottleneck

Reflection’s $150 million-per-month agreement makes the infrastructure problem plain: open-weight AI is competing in the same compute economy as closed frontier AI.

That’s good news if it produces stronger models developers can run and adapt. It’s also a reminder that the frontier is shaped by access to compute as much as by algorithms. Better training recipes matter. Better data matters. Without enough high-end accelerators, none of it scales.

Reflection has bought itself a serious seat at the table. Now it has to turn rented GB300 capacity into models that developers trust in production. The spending is impressive. The harder part starts when the training jobs do.