▶ YouTube Summaries

← All summaries
Theo#▶ Watch

Anthropic's accidental profitability: AWS reach, tier inflation, fatter tokenizers

2026-05-28 Thu ⏱ 27 min @t3dotgg

Theo breaks down rumors that Anthropic will hit its first profitable quarter (~$10.9B Q2 revenue, first operating profit). He argues profitability is partly accidental — driven by AWS distribution, a sneaky model-tier shift, a fatter tokenizer, and enterprises now paying API prices — rather than a clean efficiency win.

Why Anthropic is suddenly profitable

  • AWS advantage: Anthropic models are on AWS, GCP, and (via reroute) Azure; OpenAI is Azure-only. Real enterprises building in AWS effectively had one good option — Anthropic. Anthropic takes ~50% revenue share from cloud-hosted tokens while providing zero compute.
  • Compute conservation, not income, was the goal: Researchers want Nvidia GPUs/CUDA, not Trainium/TPUs. Cloud Code subscription users were burning Anthropic's own GPUs, competing with researchers. Limits + price hikes were meant to free compute.
  • Tier inflation: Opus 4.5 didn't replace Opus 4.1 — it replaced Sonnet. So the model most people use went from $15 to $25/M out. Haiku is now near-useless vs cheaper mini/open-weight models. Mythos is the new locked top tier.
  • Tokenizer change (4.7): 30–50% more tokens per text = immediate price bump. Wrong answers loop and cost millions of tokens; correct answers are cheap.
  • Benchmarks: GPT-5.5 scores higher using ~half the tokens of Opus 4.7. Opus 4.6→4.7 doubled token usage at same per-token price = doubled spend for same work.

Product-market fit (via Simon Willison)

  • Enterprises now pay $20/seat + API pricing (was bundled). Many discovering this at renewal as bills balloon. OpenAI did the same with credits.
  • Personal $100–200 tiers still give near-unlimited usage ($500–3000+ of inference); enterprise gets ~5% discounts despite rumored ~90% margins.
  • Coding agents are sticky enough that companies pay almost anything — unlike ChatGPT, where only ~5% of 900M weekly users paid.
  • Both labs heavily hiring enterprise sales (~30% of openings).
  • Debunks overblown stories (Microsoft "canceling" Claude Code = just rerouting via Copilot CLI; Uber blew its budget by not re-budgeting for Opus 4.5).
  • Bottom line: Opus 4.5 was the inflection — a sentiment + capability win that took ~3 months to reach enterprise spend. Profitability also helped by Anthropic under-committing on compute vs OpenAI.