Anthropic's growth just flattened

Original Article ↗ Hacker News Discussion ↗

Anthropic’s latest revenue figures show rapid but sharply decelerating growth, prompting questions about whether its AI boom is already hitting structural limits. Commenters debate how much of the business is sustainably profitable versus subsidized by massive compute commitments and investor capital, and whether Anthropic has any durable moat as rivals like OpenAI, Google, and cheaper Chinese or open-weight models close the capability gap. Many predict a shift from “token-maxxing” and experimentation toward strict cost control and model diversification, with future success hinging more on pricing, efficiency, and real-world integration than on frontier model benchmarks alone.

Revenue, Growth, and Profitability

  • Thread centers on Anthropic’s S‑1: reported Q3 revenue around $17B, ~47% QoQ and >1000% YoY, versus earlier rumors the article relied on.
  • Many argue “flattened” is overstated: growth is still very high, just slowing from previous ~150% QoQ.
  • Some note strong gross margins on subscriptions (≈85%) and improved “revenue per dollar of compute,” suggesting inference is already profitable.
  • Others counter that focusing on revenue or compute-only margins is misleading without full costs (training, labor, facilities).

Valuation, Obligations, and Bubble Risk

  • Large pre‑IPO funding and massive multi‑year compute commitments imply Anthropic needs sustained, very high growth to satisfy investors.
  • Debate over whether current or implied future valuations can be justified vs. global revenue benchmarks.
  • Some see this as dot‑com‑style overbuild; others argue it’s standard for high‑capex tech and not unprecedented.
  • Concern that IPOs mainly serve to offload risk to public/retail investors before any AI correction.

Pricing, Model Quality, and Competition

  • Strong praise for Opus 5.5, especially for coding and agentic workflows; some say it “just works” with minimal prompt engineering.
  • Others find OpenAI’s latest models, or competitors like Astra/Sol or Chinese models (DeepSeek, etc.), comparable or better, especially on price and vision.
  • Opus 5.5 is cheaper and more efficient than prior tiers; several users report their Anthropic spend fell because the model gets more done with fewer tokens or allows downgrading to Haiku.
  • Expectation that providers will increasingly compete on price and efficiency; some foresee Jevons‑paradox‑like token growth but only linear revenue.

Moats, Open Models, and Ecosystem

  • Split views on moat:
    • Pro‑moat side: extremely hard to build frontier labs, retain top researchers, and maintain internal processes; few true training labs exist.
    • Skeptical side: intelligence is commoditizing, many “good enough” models (including open and Chinese ones) undercut on cost; enterprises may switch if engineers can’t tell the difference.
  • Some argue compute access and ecosystem/integration (tools, agents, vertical apps) will be the key defensible assets.

Usage Patterns, Tokenmaxxing, and Real-World Impact

  • Reports of enterprises clamping down on “tokenmaxxing” and AI “slop,” as excessive code/PRs increase review and incident load.
  • Many claim the bottleneck is now testing, review, coordination, and business constraints, not code generation speed.
  • Observations that outside software and a few domains, adoption is patchy; many workers revert to old workflows despite available AI tools.

AGI and Long-Term Trajectory

  • AGI talk appears diminished; some think we’re already near it, others insist current LLMs are far from independent, human-level, self‑directed intelligence.
  • Several expect cyclic hype–disillusionment–rediscovery dynamics; others warn some tech waves simply die out.