How much Anthropic and Cursor spend on Amazon Web Services

Original Article ↗ Hacker News Discussion ↗

Leak and AWS Spend Concerns

  • Thread centers on leaked AWS bills showing Anthropic spending slightly more than its estimated revenue on AWS and Cursor’s AWS bill doubling MoM.
  • Some see this as clear evidence of an unsustainable business model and an imminent AI bubble deflation.
  • Others argue the numbers are raw R&D/training spend, not structural long‑term COGS, so early “selling $20 for $5” is normal for high‑growth startups.

Startups, Unit Economics, and Bubble Debate

  • Pro‑growth side: early infra vs revenue comparisons are misleading; past giants (e.g. ride‑sharing) looked terrible pre‑IPO yet later built moats. This is what venture capital is for.
  • Skeptical side: scale of current losses and circular financing (clouds “invest” then recapture via compute spend) looks like a bubble with large eventual fallout.

Inference Costs, Hardware Limits, and Usage Growth

  • One camp expects inference costs per token or per capability level to keep falling via better architectures (e.g. Mixture‑of‑Experts) and optimization.
  • Others note state‑of‑the‑art models remain similarly priced, usage (tokens, context) explodes as costs fall (Jevons paradox), and physics/power limits may cap hardware improvements.
  • Long subthread disentangles:
    • Cost of inference as provider COGS vs
    • User spend vs
    • Price per token vs
    • Dollars per end user.
      Disagreement often comes from mixing these.

Revenue Metrics and Cursor

  • Debate over Cursor’s “ARR”: critics say annualizing the latest high month overstates real revenue; defenders say that’s standard for fast‑growing, non‑seasonal SaaS.
  • Confusion over whether AWS numbers capture all compute (article itself says no; most compute comes via Anthropic).

Role of AWS and Strategic Investing

  • Some emphasize AWS as the shovel‑seller: earning huge cloud revenue while also owning a significant equity stake in Anthropic.
  • Others note AWS is actually behind Azure/GCP in AI services despite leading in generic compute.

Critiques of the Article and AI Skepticism

  • Many like the leak but call the financial analysis shallow, biased, or numerically confused (especially around Cursor’s pricing change).
  • Others defend the writer as one of the few consistent skeptics, though even some skeptics say the work is polemical, fixation‑driven, and underestimates current AI usefulness.

Enterprise Pricing Power and Adoption

  • One view: enterprises will happily pay hundreds–thousands per employee per month if they see ~10% productivity gains.
  • Counterview: much LLM‑accelerated work is “bullshit jobs” with little bottom‑line impact; most firms are too irrational and politicized to translate small productivity gains into cash savings, and will switch providers if prices get too high.

Cheaper / Chinese Models

  • Some argue Chinese/open models optimized for training and inference cost may win long‑term.
  • Others note Western adoption is still rare due to tooling, integration friction, unclear reliability, and data‑sovereignty concerns.

Meta: Hype, Shorts, and Forum Dynamics

  • Tangents on whether critics should “short” AI stocks, conflicts of interest for bulls vs bears, and analogies to previous bubbles.
  • Discussion of HN’s “flamewar filter” explains why a heavily commented, contentious thread is downranked.