The Musk Paradox: From ‘Evil Detector’ to Compute Partner
Anthropic CEO Dario Amodei took the stage at Code with Claude 2026 to announce a deal that reeks of cognitive dissonance. The company struck an agreement with SpaceX to use the entire compute capacity of the Colossus 1 supercomputer in Memphis, Tennessee, a 300+ megawatt facility packed with over 220,000 NVIDIA GPUs. Just months ago, Elon Musk publicly declared that ‘Anthropic hates Western Civilization.’ Now, after a cozy week with senior Anthropic staff, Musk claims ‘no one set off my evil detector.’ The about face is shameless, but compute is compute, and Claude is hungry.
The deal lets Anthropic double Claude Code usage limits for Pro and Max subscribers, scrap peak hour throttling on those plans, and raise API limits for the Opus model. The company frames this as a victory for users frustrated with outages and rationed access. But the real story is that demand is exploding while infrastructure remains brittle. Anthropic is scrambling to ink partnerships with Microsoft, Google, Amazon, Nvidia, and now SpaceX, because building your own compute is hard and slow. The company would rather cosy up to a mercurial billionaire than fix the underlying supply chain.
Orbital Compute: Vaporware or Strategic Hedge?
More eyebrow raising is Anthropic’s stated interest in building ‘multiple gigawatts’ of orbital compute capacity with SpaceX. The pitch is that AI training will soon outpace terrestrial power, land, and cooling. This concept is unproven and speculative, but it signals that Anthropic is willing to chase any shiny object Elon dangles. Meanwhile, users fleeing OpenAI over military contracts and multi agent coding workflows are piling onto Claude Code, exacerbating the very shortages this deal claims to solve. Raising limits now is a Band Aid, not a cure. The company is betting that orbital data centers and sweetheart GPU deals will paper over a fundamental mismatch between hype and hardware. Investors should watch closely. Users should manage expectations.
Source: Arstechnica
