SPACEX IS NOW THE AI INDUSTRY’S LANDLORD AND THE RENT JUST KEEPS GOING UP
So Elon Musk built a rocket company that also built the biggest AI supercomputer in the world, and now every major AI lab in the business is writing him fat checks every month. SpaceX just signed a deal with Reflection AI worth up to $6.3 billion, with Reflection agreeing to pay $150 million per month starting July 1 for access to Nvidia GB300 chips at the Colossus 2 data center in Memphis. The deal runs through 2029 and either side can walk with 90 days notice after the first three months.
Now here is the part that should make you laugh. Colossus was originally built as xAI’s private supercomputer. Musk’s own AI company, the one that makes Grok. It was supposed to be the competitive weapon he would use to fight OpenAI and Anthropic. And now Anthropic pays SpaceX $1.25 billion per month to use it. Google pays $920 million per month. And Reflection, a rival open-source AI lab, is going to pay $150 million per month starting next week. Musk has somehow turned his AI infrastructure into a commercial real estate empire, renting floorspace in his own house to his competitors while charging them for the privilege.
Reflection is an interesting outfit to watch here. They are valued at $25 billion, trying to build American open-source AI models that can actually compete with the closed frontier systems from OpenAI, Anthropic and Google. They need serious compute to do that, and SpaceX has serious compute. The math works, even if the optics are a little weird. An open-source AI company giving money to Elon Musk so they can build an alternative to Elon Musk’s AI company is exactly the kind of recursive absurdity that defines this industry right now.
What is really happening here is that SpaceX stumbled into one of the best businesses in technology without really trying. The Colossus data center was built for internal use. Now it is a profit center. And after SpaceX went public last week and raised $85 billion, you can bet the company is going to lean into this hard. The pattern is obvious: Google pays nearly a billion a month, Anthropic pays a billion and a quarter, now Reflection adds another $150 million. If SpaceX keeps adding customers at this rate, Colossus becomes a business that rivals anything they do with rockets. Someone should put that in the IPO prospectus.
The 90-day exit clause is the interesting wild card. It means this could all unwind quickly if anyone builds a competitive alternative. But right now there is not one, which is why everyone keeps writing the checks.
GOOGLE DEEPMIND WRITES A $75 MILLION CHECK TO HOLLYWOOD AND CALLS IT AI RESEARCH
Demis Hassabis won a Nobel Prize for solving the protein folding problem that stumped biology for fifty years. His reward, apparently, is now spending his time making deals with indie film studios. Google DeepMind announced it is investing $75 million into A24, the independent film studio known for movies like Everything Everywhere All at Once, Hereditary, and Midsommar. The two companies say they will build AI tools for filmmaking together, with DeepMind getting feedback and guidance from A24’s stable of artists.
Let me translate that. “Feedback and guidance from leading artists” means A24’s directors get to complain about the AI tools in a way that actually influences the product. Which, when you think about it, is the most expensive focus group in film history. Seventy-five million dollars to have some of the best filmmakers in independent cinema tell you what your AI is doing wrong. That’s either brilliant or insane, and honestly it might be both.
Here is the strategic angle that makes this make sense. Hollywood has spent years in open warfare with AI companies. The guilds fought, the studios sued, and the whole creative industry watched in horror as their work got scraped to train models they had no say in. The relationship between AI and film has been adversarial from the start. Google just bought itself a seat at the table by literally buying equity in one of the most respected studios in the business. You cannot sue your investor, or at least it gets a lot more complicated when you try.
A24 is also the right choice if you want artists to actually trust you. If DeepMind had invested in a Marvel assembly line or a streaming content factory, people would have rolled their eyes. A24 makes films that win Oscars and get argued about in college film classes. It carries cultural credibility that Google’s brand does not have right now with the creative community. Paying $75 million for a slice of that credibility is, by Silicon Valley standards, a bargain.
What DeepMind actually gets in return beyond goodwill and feedback is not entirely clear yet. The most likely outcome is AI tools that integrate with production workflows in ways that real filmmakers helped design, which would make them actually usable, which is not something you can say about most AI creative tools right now. The worst outcome is $75 million buys them a press release and a handshake. We will know which one it is when the tools show up.
RESEARCHERS BUILT AI AGENTS THAT REWRITE THEIR OWN OPERATING RULES AND GOT 60 PERCENT SMARTER OVERNIGHT
Scientists at the Shanghai Artificial Intelligence Laboratory published a paper on something they are calling Self-Harness, and the basic concept is exactly as interesting as it sounds. They built a system where AI agents can systematically rewrite their own operating rules to fix the recurring mistakes they keep making. The performance jump was between 33 and 60 percent depending on the model, which is an enormous improvement for something that requires no new training data and no changes to the underlying model weights.
Here is what is actually happening under the hood, because the distinction matters. When you deploy an AI agent, you do not just point it at a task and let it run wild. You give it a “harness” – a set of operating instructions that tells it how to approach problems, what to do when it hits an error, how to format its outputs, and so on. Think of it like a standard operating procedure for the model. The harness is usually written by human engineers after a lot of trial and error. What Self-Harness does is let the agent analyze its own failures and rewrite those procedures itself, targeting the specific patterns where it keeps going wrong.
The practical result is an agent that gets better at its job over time without you doing anything. You deploy it, it fails a few times, it rewrites its own rules based on those failures, and it stops making the same mistakes. If you have spent any time fighting with AI agents in production you know that the harness engineering is often the hardest and most tedious part of the whole operation. This is a framework that does that work automatically.
The philosophical angle is harder to shake. If an agent keeps rewriting its own operating rules, at what point is it still the agent you originally deployed? If the rules change enough, the behavior changes enough, you end up with something that started as one thing and became another. This is not a reason to panic but it is a reason to keep good version control and pay attention to what your agents are actually doing. “Set it and forget it” has never been great advice with AI, and a system that modifies itself autonomously makes that advice worse, not better.
Still, sixty percent is sixty percent. When researchers release this publicly, developers are going to be all over it.
NOBODY WANTS DATA CENTERS IN THEIR TOWN SO NOW SOMEONE WANTS TO LAUNCH THEM INTO ORBIT
Here is a sentence that a reasonable person in 2022 would have dismissed as science fiction: the AI industry is now seriously discussing whether to put its data centers in space. CNBC ran a full analysis of the economics this week, and the short version is that the idea has become slightly less ridiculous precisely because the world’s most powerful rocket company just went public with $85 billion in new cash and has more incentive than ever to find things to launch.
The reason this is even a conversation is straightforward. More than 100 communities across the country have proposed moratoriums or restrictions on AI data center construction. The things are loud, hot, they drink enormous amounts of water, they strain local power grids, and they show up in places that did not ask for them. Local governments are fighting back. So the AI industry, which needs to keep building at a pace that would stagger a normal person, is looking at alternatives. Space, with its unlimited solar power and complete absence of NIMBY zoning boards, is starting to look appealing.
The economics are still brutal though. A 1 gigawatt orbital data center would cost roughly $42.4 billion to build, almost three times what the same capacity costs on the ground. A tiny 1 megawatt center would burn through $140 million a year just in rocket fuel. Jeff Bezos, who has his own space company and every reason to want this to work, said a 2 to 3 year timeline for space data centers is “a little ambitious.” When the guy who funded Blue Origin tells you the timeline is too aggressive, maybe pump the brakes.
But here is the irony that is hard to get past. Every community that fights a data center contributes, in a small way, to the pressure that might eventually lead to launching AI infrastructure beyond the atmosphere where no one can object. The people trying to protect their towns from the noise and heat of server farms may end up being the reason we get a rack of Nvidia chips orbiting at 400 kilometers. That is not a comfortable thought for anyone involved. The AI industry moves the problem into space, the communities win their local fight, and all of us end up with our artificial intelligence running on computers circling the planet. Nobody planned for that outcome. That does not mean it will not happen.