SpaceX Is Quietly Becoming the World's AI Infrastructure Landlord

Something interesting is happening with SpaceX that doesn't get talked about nearly enough. The Colossus 2 data center near Memphis — originally built for xAI's own AI ambitions — is now being rented out to some of the biggest names in the industry. First it was Anthropic, then Google, and now Reflection AI is the latest to sign on.

Reflection AI, an open-weight AI startup founded in 2024 by two former Google DeepMind researchers, will pay $150 million a month starting July 1, 2026, for access to Nvidia's latest GB300 chips across the Colossus 2 facility. The deal runs through 2029 and is worth up to $6.3 billion total — though either side can walk away with 90 days' notice after the first three months.

How the Numbers Stack Up Against Anthropic and Google

Here's where things get interesting: Reflection's deal is actually the smallest of the three SpaceX has announced so far.

Anthropic is paying $1.25 billion a month. Google is paying $920 million. Reflection comes in at $150 million. That's a significant gap, but it makes sense — Reflection is a younger, leaner operation compared to the AI juggernauts it's now sharing data center space with.

What's worth noting is that all three contracts share the same exit clause: 90 days' notice, anytime after the first three months. Elon Musk has already been vocal about this, playing down the "through 2029" framing and emphasizing these are flexible arrangements. Make of that what you will.

Why Reflection AI Is Making a Bet on Open Weights

Reflection isn't just buying compute here — it's making a statement.

The startup has built its identity around open-weight AI models, which publicly release their trained parameters rather than locking them inside closed, proprietary systems like those from Anthropic or OpenAI. And honestly, the timing of this announcement couldn't be more pointed.

The US government recently banned Anthropic's closed models — Fable and Mythos — which put a spotlight on exactly what happens when your AI stack depends entirely on one private company's products. Reflection used this compute deal announcement to lean directly into that concern.

"Recent events highlight how important open source is to the AI ecosystem," a company spokesperson said, "with more nations and enterprises recognizing the risks and costs associated with exclusively depending on closed models."

It's a smart positioning move. The argument basically comes down to this: when closed models can be banned overnight, open weights start looking a lot more reliable.

The Colossus 2 Data Center: Built for xAI, Now Rented to Everyone Else

The Colossus 2 facility has a bit of a complicated backstory. xAI — Elon Musk's AI company, now folded into SpaceX — built it for its own AI development work. But as xAI's internal AI efforts have run into headwinds, SpaceX found itself sitting on an enormous amount of GPU capacity with no internal demand to absorb it.

Renting that capacity out to the world's top AI labs turned out to be a much better business.

The result is a strange but effective dynamic: SpaceX is simultaneously a rocket company, an internet provider, and now a major AI compute landlord — hosting labs that are often building things connected to Musk's broader competitive interests. The economics, though, make sense on all sides.

What This Deal Actually Means for Reflection AI

For Reflection, this is their first major compute deal, and they're treating it as a milestone worth celebrating.

Access to Nvidia's GB300 chips at scale is genuinely hard to secure. Having a guaranteed supply through 2029 gives Reflection a stable foundation to build and train the kind of frontier-level open models they've promised since raising $2 billion in 2025. That raise was explicitly positioned to make Reflection a credible open-source alternative to closed frontier labs — and compute is the raw material that makes any of that possible.

Reflection is calling this one of the largest open AI infrastructure commitments ever publicly announced. Given the scale and the explicit open-source framing, that's a reasonable claim to make.