$150 Million a Month, Locked In for Three Years
On July 1, the deal between SpaceX and Reflection AI went live. The numbers are staggering: up to $6.3 billion, $150 million per month, locked in for three years.
Who is Reflection AI? An NVIDIA-backed startup building open-source LLMs. What they signed is access to SpaceX's Colossus 2 data center, giving them direct access to NVIDIA's latest GB300 GPUs.
This doesn't read like a typical data center lease. It reads like Reflection AI lashing itself to Musk's compute infrastructure — three years of guaranteed GPU supply, no matter what the market does.
What Colossus 2 Actually Is
Colossus 2 sits in Memphis, Tennessee. It started as an xAI facility. SpaceX took over and massively expanded it. Now it's one of the largest single-site AI data centers on the planet.
The setup is unusual: SpaceX operates the facility, xAI is a tenant, Tesla consumes AI chips, and Reflection AI plugs in as an external customer. One infrastructure stack, multiple "internal customers" splitting the cost.
Why Would a Startup Spend $6.3 Billion on Compute Leases
Here's the short answer: no compute means no model training.
Models are getting enormous. GPT-5.6 runs at trillion-parameter scale. To build an open-weight model that competes with closed ones, you don't need hundreds of GPUs — you need tens of thousands. And not for three months. You need sustained, large-scale compute over years.
Reflection AI's CEO Alex Graveley put it bluntly:
"If you don't have guaranteed compute supply, your model release roadmap is a joke."
That sentence captures the anxiety running through the entire industry. In 2026, NVIDIA GB300 chips are in such short supply that allocation queues stretch beyond a year. Big tech locks in procurement quotas. Startups that don't sign long-term contracts simply won't get chips.
NVIDIA's Role
Here's the interesting part: NVIDIA is an investor in Reflection AI.
So the deal essentially goes: NVIDIA invests in Reflection AI → Reflection AI rents SpaceX compute → SpaceX compute runs on NVIDIA chips.
The money takes a lap and ends up back in NVIDIA's pocket.
What This Actually Means
AI's competitive moat has shifted. It used to be "who has better algorithms." Now it's "who has stable compute supply."
Algorithms are open. Model weights can be downloaded. But tens of thousands of GB300s aren't something you just buy. That's why Reflection AI is willing to sign a three-year, $6.3 billion contract — they're buying a ticket to play, not a tool.
You might wonder: isn't cloud computing enough? AWS, Azure, GCP all offer GPU instances. The problem is that large-scale training doesn't need generic cloud GPUs. It needs customized, high-performance interconnected clusters. GCP can't give you ten thousand GB300s connected via NVLink running a single training job.
So no, Reflection AI isn't crazy. The industry just got to the point where compute itself became the hardest currency.