The Compute Lives in Memphis, and So Does the Power
Reflection AI's $6.3B lease to train open-weight models on rented SpaceX compute in Memphis exposes who really controls frontier AI — and who does not.
Frontier AI now belongs to whoever owns the land and the chips, not whoever writes the model. That is the plain lesson of a startup called Reflection AI agreeing to pay $150 million a month — $6.3 billion over the life of the deal — to rent Nvidia GB300 capacity inside a SpaceX data center in Tennessee. "Open" AI, it turns out, can still sit on ground you do not own.
On July 1, that lease activates. According to the original report from CNBC, Reflection AI switched on access to a slice of Colossus 2, the sprawling Memphis facility tied to Elon Musk's orbit of companies, and began drawing down capacity it will pay for by the month for years. The company, led by former DeepMind researcher Misha Laskin, wants to be the American, open-weight answer to closed frontier labs — a place where the model weights are published for anyone to inspect, adapt, and run.
But consider the shape of the arrangement. Reflection AI publishes its weights. Reflection AI does not own the machines those weights were trained on. The training happened on Nvidia silicon, housed in a SpaceX building, powered by a Memphis grid that has already strained under the demand of hyperscale compute. Reflection AI is a tenant. And a tenant, however brilliant, pays rent on someone else's terms.
What does 'open' actually mean when the hardware is rented?
Open-weight has always been the more honest cousin of "open-source" in AI. It means you can download the trained model and use it without asking permission. That is a real freedom, and it matters — a researcher in Nairobi or a hospital in Ohio can run the thing locally without begging an API gateway for access.
What open-weight does not touch is the layer beneath the model: the training run itself. A frontier model is not a clever idea; it is a physical event that consumes tens of thousands of GPUs for months and burns through a small city's worth of electricity. The freedom to copy the finished weights says nothing about who controlled the furnace that forged them. Reflection AI's deal makes the gap visible. The output is free. The means of production costs $150 million a month.
Who really controls frontier AI now?
For most of the last decade, the story of software was that it ate the world cheaply — a teenager with a laptop could build something that reached millions. The inversion now unfolding in Memphis is that the most consequential software of this era cannot be built in a bedroom, or even in a well-funded startup's office. It can only be built by whoever can commit billions to compute they will never touch. Naming that shift plainly matters: the barrier to frontier AI is no longer talent or code, it is access to physical infrastructure at a scale only a handful of players command.
The entities that matter, then, are not just the labs whose logos we recognize. They are the ones who own the buildings, the chips, and the substations. Nvidia sells the silicon. SpaceX-adjacent capital builds the halls to house it. And a startup with a mission statement about openness finds itself renting a corner of that empire, hoping the terms hold.
There is a phrase people use — "sovereign AI" — to describe a nation's desire to build models on infrastructure it controls, so that access cannot be revoked by a foreign firm or a hostile board. Reflection AI is pitching itself as the American, open-weight version of that idea. But sovereignty rented by the month is a fragile kind of sovereignty. A lease can be renegotiated. A landlord can raise the rent, or decline to renew, or decide the tenant's mission no longer suits the building.
The human detail
Memphis is not an abstraction in this story. It is a city where residents near the Colossus site have already raised alarms about the gas turbines running to feed the data center's appetite, and about who bears the cost of the load on the local grid. The compute that makes an "open" model possible is not weightless. It hums in a real building, on a real street, drawing real power from a community that did not vote to become the backbone of the frontier.
So when we say a model is open, we should ask the next question: open to whom, and built on whose back? The weights may be free to download in Nairobi. The turbines run in Tennessee. Openness at the top of the stack coexists, quietly, with concentration at the bottom of it — and the concentration is where the power lives.
Reflection AI may well succeed. Its bet is a serious one, and an open-weight American frontier lab is a genuinely useful thing to have. But the deal it signed is also a confession about the era we have entered. The idea has never been cheaper to share. The ability to have the idea in the first place has never been more expensive to hold. We spent years arguing about who owns the code. The argument that matters now is about who owns the ground it runs on.
FAQ
What is an open-weight AI model?
An open-weight model is one whose trained parameters are published for anyone to download, inspect, and run without permission. It differs from a closed model accessed only through an API. Crucially, open weights say nothing about who owned or controlled the compute used to train the model in the first place.
Why does renting compute matter if the model is free?
Because a frontier model can only be created by whoever can afford enormous amounts of specialized hardware and power. When a lab like Reflection AI rents that capacity by the month, the party owning the data center and the chips holds far more sway over what ultimately gets built than the lab publishing the weights.