Amazon to invest another $4B in Anthropic

Original Article ↗ Hacker News Discussion ↗

Deal structure & accounting questions

  • Many ask whether the “$4B investment” is truly cash or largely AWS credits.
  • One linked article claims it is all cash; others argue AWS’s 40% margins mean the economic cost is far lower, especially if the money boomerangs back as cloud spend.
  • Several comments describe this as “circular” or “self-dealing”: AWS books equity and also books revenue when Anthropic spends the funds on AWS, raising concerns about creative accounting and possibly even securities or tax issues.
  • Others note this mirrors Microsoft’s OpenAI deal: big cloud funds a “design partner,” gets a showcase customer and equity.

Amazon’s strategic motives

  • Anthropic will use AWS as primary cloud, including Trainium/Inferentia chips.
  • Commenters see this as a way to:
    • Bootstrap AWS’s AI infra and custom silicon using a large, sophisticated customer.
    • Secure top-tier models for Bedrock so AWS can compete with Azure/OpenAI and Google/Gemini.
    • Potentially reduce dependence on Nvidia long-term.

Anthropic’s business, valuation & alignment

  • Claimed revenue around $850M and heavy Bedrock-based usage; one breakdown estimates 60–75% of revenue via third‑party APIs, mostly AWS.
  • Some see Anthropic as overhyped, with open-source models catching up and unclear long-term moat.
  • Others argue Anthropic offers valuable IP, safety work, and strong models, especially for coding, making the valuation defensible.

Claude vs ChatGPT and other models

  • Many developers strongly prefer Claude 3.5 Sonnet for programming and general assistance, citing better comprehension, willingness to say “no,” and superior UX (Projects, Artifacts, “concise” mode).
  • Others find GPT‑4o or o1 superior in specific domains (e.g., Apple languages, some math, complex reasoning).
  • Guardrails: Claude’s web UI is described as stricter and sometimes inconsistent; API guardrails are seen as closer to OpenAI’s. Some note refusal patterns around copyrighted or sensitive text.
  • Capacity issues are a major complaint: rate limits, 529 “overloaded” errors, degraded quality under load, and Pro users being blocked for hours. OpenAI is seen as more reliable, especially for voice.

Monetization & AI hype debate

  • Unclear how Anthropic (and LLMs in general) reach strong profitability given training/inference costs, although some expect costs to keep falling and ad/freemium models to emerge.
  • Some argue big-tech AI investments partly “buy revenue” and prop up valuations; others point to real, growing cloud and AI revenues as evidence it’s not mere hype.