Higher usage limits for Claude and a compute deal with SpaceX

Original Article ↗ Hacker News Discussion ↗

Usage limit changes & user impact

  • Five-hour Claude Code caps are doubled and peak-hour penalties removed; API limits for top models raised.
  • Weekly token caps appear unchanged. Many see this as mainly reducing annoyance: easier to work in long bursts, but you can now hit the weekly wall faster.
  • Users split: some frequently hit 5‑hour caps but rarely weekly, so this is a big usability win; others exhaust weekly quotas in 3–5 days and see little net benefit.
  • Several report having to micromanage workflows around 5‑hour windows and find that extremely frustrating.

Subscription practices & trust

  • Strong criticism that Anthropic has previously lowered limits “silently,” especially painful for annual subscribers who feel bait‑and‑switched.
  • Some argue heavy users should just move to API billing; others counter that subscriptions are effectively pre‑paid token bundles and should not be degraded mid‑term.

SpaceX/xAI compute deal

  • Anthropic will use all compute at Colossus 1; Grok/xAI apparently moved to Colossus 2.
  • Interpretations vary:
    • Smart monetization of overbuilt GPU capacity and revenue boost before a SpaceX IPO.
    • Evidence of weak demand for Grok (low GPU utilization numbers cited).
    • Strategic move to strengthen an OpenAI competitor and gain insight into Claude usage.
    • Part of a broader pattern of circular, potentially bubbly “AI capacity” deals.

Orbital data centers debate

  • Some treat plans for space-based AI compute as serious, leveraging continuous high‑intensity solar, no terrestrial permitting, and SpaceX’s launch capability.
  • Others call it economically and physically dubious: cooling in vacuum, radiation, maintenance, and launch costs likely make it far worse than land‑based options for years.

Environmental and ethical concerns

  • Heavy criticism of the Colossus gas‑turbine power setup near Memphis: alleged permit violations, NOx and particulate emissions in already burdened communities, and water risks from cooling and wastewater practices.
  • Others push back: site is in a heavily industrial zone with existing plants and refineries; turbines now permitted; pollution impact vs broader fossil‑fuel economy is disputed.
  • Several say this partnership undercuts Anthropic’s “safety” and “democratic” rhetoric; some cancel subscriptions over it.

Alternatives & model competition

  • Many users report migrating or hedging with Deepseek, Qwen, GLM, Kimi, Codex/OpenCode, Ollama Cloud, and local open‑weights; some find them “good enough” if work is carefully scoped.
  • Others maintain that frontier models like Opus/Sonnet still substantially outperform local setups, especially for complex coding and multimodal tasks.