Have you seen “The Founder”? It’s the story of McDonald’s and how Ray Kroc (played by Michael Keaton) transformed the company from a local burger joint to a global landlord. According to the movie, Kroc’s accountant gave him the revelation: “You’re not in the burger business. You’re in the real estate business.”
When brothers Richard and Maurice McDonald transformed their San Bernardino, CA barbecue restaurant into a fast-food burger joint in 1948, their model was to make and sell their own burgers. They would succeed or fail based on making better food products than other restaurants.
By the time Kroc bought out the McDonald brothers in 1961, the new model was leasing real estate to other people, and those other people would make the food. Whether the original San Bernardino McDonald’s succeeded or failed became irrelevant to the success of the McDonald’s corporation.
SpaceX has done the same thing. Its San Bernardino location, i.e. xAI, can now succeed or fail without affecting the success of the parent company, which is SpaceX.
The company this week signed a compute lease with Reflection AI, a pre-revenue startup founded by former Google DeepMind researchers. Under the agreement, Reflection pays $150 million per month for use of the Nvidia GB300 chips housed at Colossus 2, SpaceX’s expansion facility in Memphis, TN. If the lease runs for the full term, SpaceX as a landlord stands to make around $6.3 billion. (Reflection AI has shipped open-weight models, but has no widely adopted frontier model yet, and was reportedly raising capital at a $25 billion valuation.)
Earlier this month, an S-1 filing revealed that Google agreed to pay SpaceX approximately $920 million per month for 32 months. SpaceX stands to make around $30 billion. And last month, SpaceX disclosed that xAI made a big deal with Anthropic that could bring in to SpaceX as much as $45 billion in revenue.
Regardless of whether xAI — or, for that matter, Reflection, Google, or Anthropic — succeeds or fails, SpaceX still wins.
Like McDonald’s (which developed its super efficient burger-building system in order to succeed with its San Bernardino location and ended up using that system to succeed as a landlord), SpaceX is using the Colossus infrastructure it built for xAI’s Grok to succeed as an AI landlord.
Grok might succeed, or it might fail. But SpaceX makes bank if Grok’s competitors pay their rent.
That makes Elon Musk, who is the CEO of SpaceX, the Ray Kroc of AI.
Colossus originally went online in July 2024, powered by 100,000 Nvidia H100 Hopper GPUs housed in a Supermicro liquid-cooled HGX H100 chassis. The company doubled that to 200,000 GPUs within a short time and it now comprises more than 220,000 Nvidia GPUs including H100, H200, and next-generation Blackwell-class accelerators. The entire fabric runs on Nvidia’s Spectrum-X Ethernet platform, specifically the Spectrum SN5600 switch built on the Spectrum-4 ASIC.
Also: Nvidia is also heavily invested in companies that are renting compute power on Colossus. The company invests in Anthropic and Reflection AI — and, for that matter, xAI itself and, therefore, SpaceX.
Nvidia has positioned itself as the Mayor McCheese of the AI industry, collecting taxes at every node of the AI economy. It supplies the GPUs and networking fabric that every frontier lab must train on and collects hardware revenue from the winner’s rivals, even as it profits from the winner’s success.
So whether Anthropic, Google, Reflection AI, xAI, or some yet-unformed lab produces the dominant model, the computing power was bought from Nvidia and the landlord’s machine was built from Nvidia silicon.
Apple’s AI strategy is even more brilliant than Nvidia’s.
It’s built on a three-tier routing system. When you ask Siri to do something, a built-in orchestrator in the operating system decides how complex the task is. According to third-party estimates, around 85% of requests are handled on your Apple device by Apple’s own small, efficient models. (It does things like summarizing text, prioritizing notifications, cleaning up photos, or suggesting replies.) Roughly 12% of all queries get sent to Private Cloud Compute, Apple’s own server infrastructure running Apple’s larger models on Apple silicon in Apple-owned data centers. Only the hardest 3% of queries get routed to an external partner model.
This design lets Apple avoid the ruinous cost of training a frontier model from scratch. Microsoft, Google, Meta, and Amazon each spend tens of billions of dollars per year on GPU clusters, energy, and research teams to build and run trillion-parameter models. Apple doesn’t. Its own models are deliberately small and run on chips Apple already sells you, so the inference cost is basically absorbed into the device. It only needs a frontier model for that tiny sliver of hard queries, which is where the partnership strategy kicks in.
While frontier labs are collectively spending trillions to build AI infrastructure, Apple is paying Google a mere $1 billion per year to license a custom Gemini model that powers the rebuilt Siri and Apple Intelligence’s complex-query path. The reason Apple can swap partners is the Foundation Models framework, a native Swift API with a published LanguageModel protocol that any provider can implement. Google’s Gemini conforms to it. Anthropic’s Claude probably conforms to it. Any future model can theoretically conform to it. Apple’s orchestrator routes to whatever model fits the interface, so switching providers means changing routing logic, not rebuilding the whole system.
Apple profits through several channels. Apple Intelligence requires recent hardware, driving upgrade cycles. Advanced features push users toward higher iCloud storage tiers.
And so while everyone else is investing trillions, creating what is essentially debt that has to be repaid somehow, Apple is mainly just collecting billions without the massive investments needed by the frontier model companies.
The AI industry has been telling itself a story: that the companies building the best models will win, that intelligence is the product, that the chatbot with the most capabilities and the cleverest training run will capture the market. That story is wrong. Some of the companies building the best models are tenants. The companies that rent out the compute are landlords.
Ray Kroc would recognize SpaceX’s strategy immediately. The burger doesn’t matter; the land does. Right now, the most valuable land in the world isn’t in Silicon Valley. It’s a data center complex in Memphis full of hundreds of thousands of GPUs.
The man who owns it just realized that he’s not in the AI business at all. He’s in the real estate business. (And he’s probably lovin’ it.)
AI disclosures*: I don’t use AI for writing. The words you see here are mine. I used a few AI tools via Kagi Assistant (disclosure: my son works at Kagi) as well as both Kagi Search and Google Search as one part of my fact-checking for this column. I used a word processing product called Lex, which has AI tools, and after writing the column, I used Lex’s grammar checking tools to hunt for typos and errors and suggest word changes. Why I disclose my AI use and encourage you to do the same. *