Amazon spends another $2.7B on Anthropic

Structure of the Amazon–Anthropic Deal

• Many commenters assume a significant portion of the $2.7B is AWS credits, not just cash.
• Debate over how to value these credits:
  • One view: Amazon “spends” retail value but only incurs underlying infrastructure cost, inflating the apparent investment and Anthropic’s valuation.
  • Counterview: If credits displace paying customers, they’re effectively worth retail; the economic impact is real.
• Some call such in‑kind, tied-to-cloud investments quasi–money laundering or misleading “gift cards”; others see them as standard, mutually beneficial strategic partnerships.

Competition, Exclusivity, and Antitrust Concerns

• Critics argue cloud‑credit investments are anti‑competitive: they lock startups into a single cloud and distort valuations and reported revenue.
• Defenders say:
  • These are voluntary deals between sophisticated parties.
  • Similar structures exist everywhere (e.g., sweat equity, strategic partnerships).
  • If investors or LPs can’t parse the structure, that’s their responsibility.

AWS Chips and Cloud Strategy

• Several point out Amazon does have its own AI chips (Inferentia, Trainium via Annapurna Labs).
• Unclear how these compare to TPUs or Nvidia GPUs, but commonly noted downside is weaker software tooling vs Nvidia.
• Some speculate Anthropic is playing Google and Amazon off each other for better terms.

Perception of Anthropic and Model Quality

• Multiple commenters are impressed with Anthropic’s Claude 3 (especially Opus), some preferring it over GPT‑4 for coding and writing.
• Others highlight strong alternatives (Mistral, Cohere, local models) and note Gemini’s relative underperformance in their experience.

Local Models and Apple/PC Hardware

• Active subthread on running open‑weight models (Mistral, Llama, Mixtral, Yi, Sterling) locally, especially on Apple Silicon.
• Disagreement over how “blazingly fast” Apple’s unified memory GPUs really are, and whether large local models are worth it versus smaller models or cloud APIs.
• Privacy is cited as a key reason to run local models instead of using GPT‑3.5-like cloud APIs.

Scaling, Costs, and AGI “Stall” Debate

• Discussion on rapidly escalating training costs (from hundreds of millions potentially toward billions per model).
• Concern: if performance gains plateau, investors may stop funding 10× annual scaling, potentially “stalling” AGI progress.
• Others argue even flat spend with improving hardware still yields growing compute; any slowdown would be reduced exponent, not a true halt.