Comment by credit_guy

I disagree. AI is very cheap. People speculate that OpenAI and Anthropic and Google heavily subsidize the AI they provide. But all the evidence points towards this not being true. You can look at all the independent providers, like Cerebras, Groq, TogetherAI, and dozens others. Some may swim in venture capital money, and can afford to subsidize, but I doubt all can do that. And if they can't, then how do you explain that the cost of the million tokens is so low?

And this is just now. Inference costs are plummeting, because models are becoming more and more efficient. I can get 6 tokens/second on my local Ollama from GPT-OSS-20B using only CPU, and I can get 11 tps from Qwen3-30B. This was unthinkable 6 months ago. I am quite certain I'll get faster speeds 6 months from now and faster still 6 months later. Model architectures are becoming better and better, and models with the same number of parameters are becoming smarter and smarter.