I think this is mostly right. But there’s something missing here: agentic tuning. Lately, OpenAI and Anthropic have focused on models that are well trained to use tools, not just to do well on standard benchmarks like GPQA or MMLU. You’re probably right that the most connected platform will come out ahead. But they’ll need to keep their models sharp at using these tools. And with more and more MCP tools, we might end up with context windows packed full of just tool calls. That could be a problem.