The AI industry’s strategic pivot happened this week. Elon Musk’s SpaceX and xAI are leasing their Colossus 1 supercomputer - a 220,000 GPU facility - to Anthropic. The move solves Anthropic’s acute compute shortage and signals Musk’s retreat from the frontier model race. As Frank Downing noted on FYI, Musk is becoming a compute kingmaker, a utility provider like Nvidia, while Anthropic gets the raw power it needs to stop throttling its users.
The deal underscores a brutal new reality: hardware, not software, is the ultimate bottleneck. Nathaniel Whittemore’s analysis highlighted that Anthropic had the best coding model but couldn’t run it reliably. Musk had the world’s largest AI cluster but a mediocre model. The partnership is a marriage of convenience dictated by scarcity.
“Anthropic is aggressively routing public users away from high-performance Nvidia H100 clusters and onto inferior chips like Google’s TPUs and Amazon’s Trainium.”
- Theo, Nerd Snipe
This hardware triage has a direct user cost. Theo and Ben cited an internal AMD audit showing the tokens required for identical Claude Code tasks increased 170-fold from January to March, with costs exploding from $26 to over $42,000. Anthropic is sacrificing public performance to preserve its research pipeline, creating a two-tier system where employees use superior internal versions.
The shift in focus is evident in Anthropic’s product launches. At its recent ‘Code with Claude’ event, the company didn’t announce a new model. Instead, it unveiled ‘Managed Agents’ with features like Dreaming, which allows AI to review its own work, and Outcomes, which uses a grading agent to iterate until a rubric is met. Boris Cherny, creator of Claude Code, stated that manual coding has effectively disappeared within Anthropic’s own teams.
The endgame is an AI that never stops learning. Anthropic CEO Dario Amodei revealed the company is targeting ‘infinite’ context windows. As Whittemore explained, if a model’s context window never clears, it gains a persistent, evolving memory, reducing the need for constant retraining. The distinction between reasoning and processing is collapsing.
“If rented out as infrastructure-as-a-service, a gigawatt generates roughly $15 billion per year. If used by the model builder itself, like OpenAI or Anthropic, that same capacity can drive $30 billion in annual revenue.”
- Frank Downing, FYI
The economics justify the hardware arms race. Building a one-gigawatt data center costs $60 billion but can generate $30 billion annually for a vertically integrated AI company. This payoff is why the next scaling frontier may be space. Downing’s co-host Sam Korus argued that orbital data centers, launched by SpaceX’s Starship, could deploy compute faster and cheaper than fighting for terrestrial power and land.
Ben Horowitz’s framework explains the strategic mindset. On The a16z Podcast, he argued a founder’s only job is ‘right product, right time.’ Everything else is support. For Musk, the right product is now compute infrastructure. For Anthropic, it’s agentic workflows. Both are focusing on what they can uniquely deliver amidst a shortage - be it power or product - and outsourcing the rest. The model war is over. The utility war has begun.
