At least when it comes to AI model training itself, there really isn’t much of a moat.

AI = more GPUs + more data + better training recipes + more engineering optimizations.
This is already industry consensus. AI itself doesn’t require exclusive equipment, materials, and process accumulation the way advanced chip manufacturing does.
Once a model’s capabilities have been publicly demonstrated, it effectively tells the entire industry that this is feasible—so later entrants don’t have to shoulder the full cost of failure from repeating all the breakthroughs and missteps of leading companies, and they can quickly narrow the gap.
The real moat for AI vendors has never been the model’s capabilities themselves. The truly valuable moat lies in the user ecosystem, behavioral data, and developer mindshare.
So even if Kimi, Grok, and DeepSeek have shown that model parameters and capabilities can quickly catch up, I still think it’s hard for them to pose a real challenge to OpenAI or Anthropic.
Once a user ecosystem forms, it becomes an extremely deep moat. Even now, with AI search capabilities so strong, Google’s search ads business is still accelerating in growth—this is still the case even when the capability is clearly lagging behind the market.
Don’t start calling for OpenAI and Anthropic’s valuations to be slashed in half just because other models are developing quickly. OpenAI has nearly 1 billion WAU, and Anthropic holds more than 50% of the enterprise LLM spending share.
Under these circumstances, they also continuously generate high-quality user data. This moat can’t be explained by simply “winning a leaderboard” or making a show of leaderboard performance.