I think Anthropic and OpenAI have found product-market fit

Frontier LLMs, agents, and “product–market fit”

• Many agree that coding agents + harnesses are the first clear PMF: they materially speed up routine code work, enable more experiments/POCs, and are now “daily drivers” for some developers.
• Others argue the PMF moment actually arrived a year or more ago; what’s new is just pricing and packaging (enterprise plans, agents).
• Several stress that “working” ≠ “good”: maintainability, subtle bugs, and code quality remain open problems, especially with large agentic changes.

Pricing, margins, and hardware economics

• Commenters debate whether current per‑token pricing yields healthy margins or is still subsidy:
  • One camp says inference is already profitable vs energy/compute, with training as the big fixed cost.
  • Another doubts this, citing huge datacenter capex, GPU depreciation, and rumors that “all you can eat” seats can consume thousands of dollars in API-equivalent usage.
• There is sharp disagreement over claims that labs must “make back $5–10T in 5 years”; several people call those numbers arithmetically and financially wrong or at least highly speculative.

Enterprise demand, token budgets, and governance

• Multiple reports that large orgs are:
  • Being pushed off flat-rate “max” plans onto API-style consumption billing.
  • Starting to get serious about token budgets; some teams are ordered to cut usage by ~30% month‑over‑month.
  • In some places, AI use is being mandated and even performance‑measured (token leaderboards), raising questions about artificial vs organic demand.
• Uber quotes about blowing through AI budgets are widely discussed; some see them as evidence usage isn’t justified by ROI, others as simple mis-forecasting after November’s capability jump.

Open‑weight, Chinese models, and on‑prem

• Strong theme: GLM, DeepSeek, Qwen, Kimi, etc. are “good enough” for many coding and office tasks at a fraction of the cost.
• For individuals and small shops, cheap open‑weights hosted by third parties or on local GPUs are already displacing $200/mo frontier subscriptions.
• Enterprise adoption of Chinese models is seen as constrained by compliance, geopolitics, and lack of local infra, but they’re expected to dominate lower‑tier and non‑US/EU markets.

Demand limits and trillion‑dollar valuations

• Many doubt the total addressable market needed to support current capex and valuations:
  • Even optimistic “5–20% of every knowledge worker’s salary goes to tokens” scenarios look tight.
  • Consumer willingness to pay $20–$60/month is questioned; most people can live with free or “good enough” models.
• Some foresee a future “AI winter” or at least a painful correction once subsidies end, token prices rise, and enterprises rigorously measure ROI. Others think small, highly leveraged businesses plus long‑run hardware and algorithmic efficiency gains could eventually justify large, but not necessarily current, valuations.