I stood beneath the floodlights of Colossus and felt a small jolt: this was not a rocket yard but a machine farm the size of a small city. You would expect launch pads and lunar models; instead there were racks, cables, guards with clipboards and a deal that smells of money. For SpaceX, the future arrived not on a launchpad but through a loading bay.
I’ll say this plainly: I follow where the compute goes, and right now compute is parking inside a Tennessee warehouse you’ve probably never heard of.
On a Memphis loading dock, workers swipe in — Colossus is now the backbone of a new business
The hulking Colossus data center stretches over a million square feet and, by SpaceX’s account, houses more than 220,000 Nvidia GPUs. You may have imagined SpaceX as a Mars-first company; instead it has quietly become a landlord for the very hardware that runs modern AI. That matters because those GPUs are the currency of today’s artificial-intelligence economy.
How much is Reflection paying SpaceX?
Reflection has agreed to pay SpaceX roughly $150 million per month (€138 million) for access to Colossus starting next month, in a contract that runs through 2029. If both sides hold to the term, SpaceX will collect more than $6 billion (€5.5 billion) over the life of the deal, a windfall that redefines where the company’s revenue comes from.
You should see the arithmetic: rent out a giant compute facility, sign a few big AI labs, and what looked like a rocket company now earns like a data-center operator. I’m not romanticizing; I’m labeling a business model.
At a server rack, engineers swap CPUs — SpaceX is trading compute with the industry’s giants
On the same campus where crews once polished fairings, contracts are negotiated with Anthropic, Google, and now Reflection. SpaceX’s pattern is clear: it keeps the hardware and sells time on it. That has led to a circular flow of capital. Nvidia supplies the chips; startups and Big Tech buy the cycles; investors pour cash into software firms that then buy more Nvidia gear. SpaceX sits in the middle as a toll bridge for compute.
Reflection itself has drawn money from Nvidia — reportedly about $800 million (€736 million) — while SpaceX says Colossus houses hundreds of thousands of GPUs. You can trace how money and compute funnel through a few nodes: Nvidia, cloud providers, and now a formerly rocket-focused company.
Why would SpaceX rent compute instead of using it only for Grok?
Because renting is lucrative and strategic. After merging with xAI and acquiring Cursor, SpaceX could have hoarded compute for its own model, Grok. Instead it sells access. That brings immediate revenue, spreads operational risk, and makes SpaceX a hub for talent and traffic. For developers, paying for cycles at Colossus can be cheaper than building a private fleet of GPUs; for SpaceX, the same hardware becomes a recurring-revenue engine.
In a cramped coffee shop at a developer meetup, a grad student asks about open source — what Reflection is selling is influence
Reflection launched in 2024 and pitched open-source AI as a democratic alternative to proprietary labs. I spoke to developers who like the idea of models whose internals anyone can inspect, fork, or fix. That community argument has teeth: open development can surface bugs quickly, and it places control in the hands of many rather than a single corporate gatekeeper.
But open-source models hungry for frontier performance often need raw compute rather than clever shortcuts. Some projects use distillation to mimic bigger models; others need full-scale training runs. Reflection’s deal with SpaceX buys it the latter. The arrangement sets up a new axis of power — not just who builds the models, but who sells the time to build them.
There’s a political angle you can’t ignore: the federal government’s clash with Anthropic has made companies and users sensitive to centralized control of advanced models. If regulators or politics can press a single company to limit or shutter access, open-source alternatives that run on rented compute may look safer to some.
What does this mean for open-source AI?
It means open-source projects can scale if they secure predictable, high-volume compute. Reflection’s access to Colossus shows a path: combine communal model development with industrial-grade infrastructure. That could accelerate innovation — and make open models competitive with closed, heavily funded labs.
I’ve watched similar shifts before: infrastructure reshapes who wins. SpaceX’s data-center pivot feels like a power plant for algorithms, and it raises questions about where influence will sit as compute becomes the commodity that defines strategy.
SpaceX’s stock dipped nearly 17% before the market close on Monday, a reminder markets still prize narratives more than the quiet grind of server ops. But markets can pivot fast when revenues scale the way these deals promise to.
You and I are watching an industry that turns hardware into leverage, and a company once synonymous with rockets is now selling time on its machines — who benefits most when compute is king, and which side will you place your bet on?