Large models are still competing over parameters, but what truly begins to hit the industry ceiling is another issue: context storage.

As inference length, agent call chains, and long-term memory all grow longer, the real factors that determine experience and cost are not just computing power, but whether "context can be read, written, scheduled, and reused efficiently."
This is also why the market has recently started shifting focus toward infrastructure such as Context Memory, KV Cache, and layered inference storage.
In the next phase of AI competition, it may not be about who can generate more content, but who can keep the model working stably and at low cost over longer tasks.
If training was about competing with GPUs, then in the agent era, it’s about competing with Memory.
That’s also why I discussed with friends in a group about why lobsters are worth playing with—I said Claude Code is all about targeting lobsters. But they said, regarding full context, I was speechless—I should just honestly and diligently raise lobsters.