The temperature in Shandong dropped suddenly these past two days; it took two cups of pour-over coffee to bring me back to life. By the way, today’s beans seem over-roasted. I just opened Twitter and stumbled upon news that almost made me spit out my coffee: Anthropic and SpaceX are shacking up.
No 220,000 GPUs, No Seat at the Table
Anthropic’s chief, Dario Amodei, spilled the beans in a recent interview. In the first quarter of this year, driven by the explosive growth of the Claude Code developer ecosystem—yes, that magical tool that lets programmers sip tea while watching AI automatically refactor code—the company’s revenue and API calls surged an astonishing 80 times year-over-year. Absolutely fierce.
But this massive windfall almost made their server rooms burst into flames.
Here lies the problem. Squeezed by immense traffic, Anthropic had to scramble for reinforcements. Ultimately, they made a sweeping move and exclusively leased SpaceX’s Colossus 1 data center in Memphis. That’s a whopping 300 megawatts of new compute capacity. To put that in perspective, it’s roughly equivalent to a massive cluster of 220,000 Nvidia GPUs.
Think about it: an AI software company that constantly preaches “human safety and value alignment” is now forced to mingle with a heavy-industry rocket manufacturer. Frankly, it’s like going to the neighboring steel mill to borrow their blast furnace just to bake a cheese crostini. Isn’t this scene a bit surreal?
Musk’s “Rent-Seeking” Game: Who’s the Real Winner?
Everyone treats this as a standard Silicon Valley giant alliance. But the waters run much deeper.
Notice that the contracting entity in the announcement is called “SpaceXAI”—the new label after xAI merged into SpaceX this February. This is a crucial detail.
While everyone else is obsessing over model parameters, Musk took a completely different path. He pivoted into becoming a “compute landlord.” It only took him 122 days to build Colossus 1, and now he turns around and leases it entirely to his former rival, Anthropic. As for his own xAI training tasks, rumor has it they’ve all been migrated to the newly built, gigawatt-scale Colossus 2.
Honestly, this maneuver is brilliantly shrewd. Last month, Musk leveraged his compute resources to secure the acquisition rights for the AI coding tool Cursor, casually helping them establish the foundation for their next-generation models. Now, he has grabbed Anthropic by its infrastructure lifeline. I originally thought the moat in the AI era was top-tier mathematicians and genius hackers. It turns out Musk is telling us: whoever controls the transformers, cooling towers, and tons of compute clusters is the true dealer at this table.
A Multi-Billion Dollar Infrastructure Game
Comparing across the industry, the barrier to entry has become absurdly high.
Last month, Anthropic just signed a 5-gigawatt compute agreement with Google. What exactly does 5 gigawatts mean? That capacity could power hundreds of thousands of households running their air conditioners full blast in the summer. It’s said that for these 5 gigawatts, they have to pour in nearly $200 billion over 5 years. Add that to the $30 billion contract with Microsoft Azure, and their infrastructure bill alone is an astronomical figure.
To put it bluntly, the price of staying at the table now can’t be covered by a few stunning top-tier conference papers. The threshold is too high. If you can’t cough up tens of billions in cash, you don’t even have the right to watch others play. It’s like the 30% yield rate in the semiconductor industry—meaning for every 10 chips made, 7 must be scrapped—and every complete training run of a large model might burn enough electricity to power our entire neighborhood for several summers.
In the old internet startup days, three or five people renting a garage could try to change the world. That playbook doesn’t work anymore. The early AI venture capital circle used to think of large models as some lightweight digital magic. Frankly, this thing is fundamentally a heavy steam locomotive devouring minerals, water, and electricity.
Orbital Compute, or Sci-Fi Novel?
Interestingly, the news also mentioned that the two companies are interested in co-developing “multi-gigawatt orbital AI compute.”
That’s a blind spot in my knowledge. My superficial understanding is that these guys genuinely intend to launch data centers into the sky.
Sometimes I wonder: if this actually happens, next time we call a Claude API in our local code editor, will the data packets have to first go to space, say hi to the Starlink satellites, endure a baptism of minus 200 degrees of cosmic radiation, and then be sent back? (o_O)
But then again, Earth’s electricity is indeed being drained dry by large models. The aging power grids in the US are already fragile. Instead of fighting local residents for natural gas power generation on the ground, stuffing energy-intensive servers into space sounds crazy, but logically, it’s surprisingly consistent.
Or maybe I’m overthinking it. The radiation-hardened chips and vacuum cooling mechanisms in space will be enough of a headache for them. For today’s average developer, rather than some mysterious orbital compute, quickly maxing out these 220,000 GPUs and honestly lifting the API rate limits for Pro users would be the most practical comfort.
Closes notebook
The sky outside seems to be getting gloomier; time to move the succulents on the balcony to a better spot.
References:
- SpaceX Becomes Compute Leasing Giant: Anthropic Sweeps Up Entire Colossus 1 Capacity
- Anthropic Taps Elon Musk’s SpaceX for More AI Compute Power
—— Lyra Celest @ Turbulence τ.
