AI's Affordability Crisis

Original Article ↗ Hacker News Discussion ↗

AI Costs, Pricing, and Margins

  • Strong disagreement on whether “AI is too expensive.”
    • Some say API prices per capability have dropped ~50x in a few years; others note per‑token prices trending up and hardware/cooling costs exploding.
    • Debate over whether inference is high‑margin (75%+ on API pricing) or still subsidized once depreciation and constant retraining are included.
  • Subscription vs per‑token math is contested:
    • Flat‑rate plans look massively subsidized if fully used; critics argue this ignores typical under‑usage and caching.
    • Others argue API prices may already be “soaked,” making subscription comparisons misleading.

Enterprise Behavior and “Token Panic”

  • Multiple anecdotes of “AI free‑for‑all” followed by sudden clampdowns once token bills arrived.
    • Access to frontier models restricted, approvals and monitoring added, some tools completely shut off.
    • Companies now emphasize ROI and “cheap models by default,” especially for coding.
  • Demand for expensive tokens appears highly elastic: usage drops when budgets and monitoring kick in.

Competition: Chinese, Open, and Local Models

  • Many point to Chinese and open‑weight models (e.g., DeepSeek, Qwen, GLM) as far cheaper and increasingly “good enough,” especially for coding.
  • Some shops report monthly spend in the low hundreds of dollars for whole teams by mixing local models, cheap Chinese APIs, and a small number of Western subscriptions.
  • Concern that US export bans and security policies will push some users away from US models or force domestic “rebranded” training on foreign weights.

Profitability, Capex, and Bubble Risk

  • Huge capex on GPUs and datacenters (~hundreds of billions per year) seen as potentially unsustainable.
    • Comparisons to the dot‑com/fiber and nuclear power economics; training likened to constantly rebuilding a power plant.
  • Debate over reported profitability:
    • Some point to “profitable quarters” and growing revenue.
    • Others argue accounting (one‑time charges, EBITDA excluding depreciation, stock comp, buybacks) obscures true economics.
  • Many expect a shake‑out or “correction” as enterprises discover limited, uneven ROI.

Developer Productivity and Use Cases

  • Many individual devs report massive productivity gains: faster onboarding, code generation, reviews, and data analysis.
  • Others argue that beyond a point, humans must still deeply understand code, eroding net gains, especially for long‑lived systems.
  • Concern that “vibe coders” and students over‑relying on AI will lack fundamentals to review or debug AI‑written code.

Lock‑in, Regulation, and Future Market Structure

  • Skepticism about durable vendor lock‑in: harnesses are thin, models are somewhat fungible, and switching is relatively easy.
  • Counterpoint: real lock‑in may be at enterprise level via contracts, compliance, and proprietary agent platforms.
  • Some foresee LLMs ending up as commodity cloud services (like databases), with hardware/inference efficiency as the main moat.

Social, Cultural, and “Enshittification” Concerns

  • Fears of AI‑driven enshittification: pervasive ads, AI‑generated slop overwhelming forums, and subtle manipulation in “assistant” interfaces.
  • Anxiety that frontier labs’ survival depends on replacing large shares of white‑collar labor, especially software engineers.
  • Others maintain that demand for “intelligence” is effectively infinite and that, despite financial turbulence, LLMs are here to stay and will keep improving.