inference costs

Bilingual

No, it doesn't cost Anthropic $5k per Claude Code user
2 months ago
- Forbes claimed Claude Code Max users cost $5,000 in compute, but this is based on retail API prices, not actual inference costs.
- Actual compute costs are roughly 10% of API prices, as shown by open-weight models on OpenRouter (e.g., Qwen 3.5 at $0.39/MTok input vs Opus $5/MTok).
- Heavy users cost Anthropic ~$500 per month, not $5,000—a $300 loss, not $4,800.
- Most users are break-even or profitable; average Claude Code user consumes $180/month API-equivalent, costing $18 to serve.
- Cursor faces the real $5,000 cost because they pay retail API prices for Anthropic models.
- The narrative that AI inference is a money pit is misleading; it discourages competition and justifies high API markups.
AI GPUs probably live longer than three years
2 months ago
- The claim that AI GPUs last only three years under high load originates from an anonymous source paid by Tegus, incentivized to sound confident rather than provide accurate data.
- Anectodal evidence suggests longer lifespans: Google's TPUs last eight years, AWS has never retired A100 servers, and ex-crypto GPUs remain functional.
- Hard data from supercomputers like Titan shows over 90% GPU survival rates at six years under proper cooling conditions.
- GPUs may have a short economic lifespan due to newer, more efficient models, but this does not support the argument that inference costs will rise post-bubble.
- GPUs are only part of datacenter costs; replacing them doesn't require building entirely new facilities.
- The belief that AI GPUs last only 1-2 years is popular among skeptics but lacks solid evidence and is contradicted by public communications and statistical data.
GLM 5.2 and the coming AI margin collapse
2 months ago
- DeepSeek's R1 model caused a market reaction due to low training costs, but AI economics are more complex.
- Training costs are fixed, while inference costs scale with demand and have high gross margins for frontier labs.
- GLM 5.2 is a strong open-weights competitor to models like Opus and GPT, with near-equivalent quality.
- Drawbacks of GLM 5.2 include slower speed for interactive use, lack of vision support, and poor web search capabilities.
- Migration to open-weights models like GLM 5.2 is easy due to OpenAI/Anthropic-compatible endpoints, with low switching costs.
- Cost savings are significant: GLM 5.2 is priced at ~$4.40/MTok, less than 20% of Opus, and may become even cheaper with optimizations.
- Enterprise concerns about data privacy with providers like Z.ai can be mitigated by using other providers or on-premises hosting.
- Inference margin collapse could reshape the AI industry, benefiting those who leverage lower costs.

inference costs

No, it doesn't cost Anthropic $5k per Claude Code user

AI GPUs probably live longer than three years

GLM 5.2 and the coming AI margin collapse