Wow, @arena is made by someone from Taiwan?

Recently, the AI coding leaderboard is really worth watching 👀
But I think the focus is no longer on "who is number one."
What truly matters is: the top AI models are becoming less and less rare. 🧠⚡️
In the past, everyone thought AI would be a winner-takes-all game:
GPT-4 far ahead,
other models can only catch up.
But now, looking at leaderboards like Arena, the top models are getting more crowded. Claude, OpenAI, Google, GLM, Qwen, Kimi, and various open-source and closed-source models are all within the same capability range. The Elo score gap is narrowing, indicating that model capabilities are rapidly standardizing.
This is very similar to water and electricity 🚰
You turn on the tap, and you don’t really care which company supplies the water.
What you care about is:
- Is it cheap?
- Is it stable?
- Will it cut out?
- Can it integrate into your workflow?
AI models are heading in this direction too.
As the capability gap narrows, the market’s re-pricing isn’t about "who is the smartest," but about:
🧩 who can integrate into the workflow
💰 who has the lowest inference cost
🔒 who can meet enterprise compliance and security
📊 who has data feedback and user retention
🛠 who can turn models into products, not just demos
This is especially true for coding models.
Engineers don’t necessarily choose the "top-ranked" model in the end.
They choose the one that is most stable, cheapest, most familiar with their codebase, and least likely to suddenly break.
That’s why when I look at the Arena leaderboard, my first focus isn’t the ranking, but the structural changes.
The more crowded the top twenty, the thinner the moat around the models themselves.
Value shifts toward products, data, distribution, computing costs, and enterprise deployment capabilities. 🏗️
And there’s one more interesting thing:
Arena, the global AI evaluation infrastructure, was co-founded by Wei-Lin Chiang, who graduated from NTU’s Computer Science department, and later did research at UC Berkeley, creating the Chatbot Arena blind testing system.
In the past, the strongest narratives in AI were usually about chips, servers, and supply chains.
But Arena reminds us:
It’s not just about building AI hardware infrastructure.
It’s also about participating in AI trust infrastructure. 🌏
The most important issues in the AI industry in the future might not be:
"Who has the strongest model?"
but rather:
"Who has the qualification to define what is strong?"
"Who can become the credit rating agency in the model world?"
"Who can make the market believe these AI rankings are real?"
Two years ago, the strongest models themselves were the moat.
But in the next phase, what’s truly valuable might be:
- how models are evaluated,
- how they are deployed,
- how they are trusted,
- how they are used long-term by enterprises.
The AI war is shifting from "model capability" to "infrastructure." 🚀