The moat, imo, is mostly the tooling on top of the model. ChatGPT's thinking and deep research modes are still superior to the competition. But as the models themselves get more and more efficient to run, you won't necessarily need to rent them or rent a data center to run them. Alibaba's Qwen mixture of experts models are living proof that you can have GPT levels of raw inference on a gaming computer right now. How are these AI firms going to adapt once someone is able to run about 90% of raw OpenAI capability on a quad core laptop at 250-300 watts max power consumption?
I think one answer is that they'll have moved farther up the chain; agent training is this year, agent-managing-agents training is next year. The bottom of the chain inference could be Qwen or whatever for certain tasks, but you're going to have a hard and delayed time getting the open models to manage this stuff.
Futures like that are why Anthropic and oAI put out stats like how long the agents can code unattended. The dream is "infinite time".
