Will the AI arms race end? The future winners may not be just one company, but the entire AI ecosystem!

Let’s talk about recent thoughts on AI. Yesterday, Google released its financial report: the expected revenue was 117.0 billion yuan, and the actual revenue was 119.8 billion yuan—beating expectations—but the stock price fell. The main reason is that AI spending is too high, and cash flow turned negative.

So going forward, this AI arms race may slow down. As for the stock decline, shareholders don’t agree to increase AI spending—after all, a listed company needs ROI (return on investment).

Therefore, a shift may happen in the future AI arms race: it won’t be about stopping investment; instead of “unlimited burning of money,” it will move into a phase that focuses more on efficiency and returns.

Then why are people trying so hard to estimate and drive compute power right now?

Because the AI industry has a very important theory: Scaling Law.

In fact, researchers have known for a long time that the more compute power you invest into AI, the larger the number of AI parameters, and the larger the training data, the “smarter” the AI becomes—and it’s a linear relationship!

That’s why major AI companies have started to pile up compute power: the more money someone throws in, the smarter their AI is, and they can lead the industry.

But right now, AI is still in a very early stage. It hasn’t reached the point where it can evolve itself—meaning, when it encounters a problem and realizes it won’t be able to handle it, it proactively seeks knowledge, modifies its own capabilities, and learns in a way that it will permanently apply next time.

No current large model can do that. At most, it can only achieve “limited self-improvement.”

For example, humans have a model write 1,000 pieces of code, then let the AI run them to find problems, analyze the reasons for failure, retrain on the data, and improve its abilities. But this kind of limited self-improvement is still in the experimental stage; it requires humans to keep updating the model and releasing new versions. In other words, humans are still driving AI evolution.

True “pure online” self-evolution is still far from being achieved. That’s why we see major companies releasing a new version every so often—those new versions are AI models that have been optimized and evolved by humans.

Why can’t this be done now?

First, at its core, a model is a huge collection of parameters. Those parameters are called model weights. A model can’t change its own weights—it can only remember context, and then, based on the predefined weights and the data it was trained on, generate your answer.

Second, AI currently can’t determine what counts as “real knowledge.” For example, in philosophy, medicine, and economics, many things are right when argued one way and also right when argued the other way. Like in classic debate questions: “A person cannot step into the same river twice.” If you tell AI to answer that, what would it say? Even humans don’t have a standard answer. If you force the AI to give you one standard answer, it still can’t truly judge.

Also, if I say Xiaomi is good so the stock can be bought, while others say Xiaomi is trash so anyone who buys the stock will get stuck—this AI also can’t judge.

Third, it may end up learning mistakes more and more. If an AI could modify its own weights, then if something is wrong but it keeps insisting it’s right, it would end up going further and further down the wrong path, and ultimately the model might collapse.

Fourth, the cost is too high. For AI to learn, it needs lots of GPUs and electricity. Constantly replacing models and changing parameters makes the economic cost unsustainable.

So this is the current “bottleneck.” That’s why the top AI industry today is researching team-based collaboration models. For example, form a team so that it creates a closed loop: some teams work on research models, some on reading models, some on experiment models, some on programming models, and some on evaluation models—forming an end-to-end cycle.

So earlier I thought AI’s future would be one dominant player, where the winner takes all—which might not hold. In the future, it’s very likely to become more diversified. And if the AI models from different vendors can be connected, it could very possibly become a super AI intelligent agent. Because each company may train on different data, but if all the data can be integrated, AI capability could increase exponentially.

As AI model parameters keep growing, AI capability, parameters, compute power, and training data volume are proportional. But with changes in model parameters, AI might suddenly “emerge” a new capability—like suddenly “getting it.” For example, if you’ve been learning math abilities, one day you suddenly solve advanced data equations.

And currently, each company’s AI models actually have their own strengths. For instance, Claude is better at programming, GPT has stronger math ability, and Gemini is excellent at video. So in the end, if there is a front-end coordinator—or if the model capabilities are connected—each model will have a chance to survive.

If you treat each AI like a person, it’s clear that teamwork matters more than needing one person to be a jack-of-all-trades.

Based on the thoughts above, I think it’s now possible to focus on AI large-model companies. In the future, these companies all have opportunities. And since the AI infrastructure has already been laid out for the most part, the key is to look for which large models are trading at undervalued levels and move decisively.

Because the next round of AI competition may not be a war between a single model.

It’s more likely to be an AI ecosystem war.

And the real winner may not be one AI, but rather a network of super-intelligent agents made up of multiple AIs.