Last January, the market sold off hard after a Chinese lab launched its frontier model, DeepSeek. Following its launch, it erased trillions of dollars in market cap. Right now, the same thing is happening again.

China’s Moonshot AI just released Kimi K3. Moonshot is valued at roughly $31.5B while leading U.S. AI labs are being discussed at valuations approaching or exceeding $1T.

The simple but flawed argument goes as follows:

Right now, the most expensive part of developing AI is training. OpenAI and others spend hundreds of millions on compute to create better AI models. This requires massive data centers with thousands of GPUs and entire power plants.

But DeepSeek and now Kimi are shaking up the entire industry.

They developed state-of-the-art models that are not only better but also cheaper than almost any other model right now. Their models beat the latest models from leading AI companies such as OpenAI and Anthropic. And they cost up to 70% less to run than their competitors.

The cost per token has been falling fast.

Since ChatGPT launched in 2022, we’ve seen roughly a 1,000x decrease in the price per token. This is how technology works. Everyone expected this to happen. DeepSeek and Kimi simply accelerated the trend even further.

But isn’t that a good thing? I would say yes.

Whenever you lower the cost of a good, new applications emerge that were previously not economically feasible. This is known as Jevons Paradox.

Think of mobile phones. They used to cost tens of thousands of dollars. Now you can get a smartphone for a few hundred dollars.

In the long run, this will probably increase the total demand for AI computation. But Wall Street doesn’t like it when industries are shaken up and the current status quo is questioned.

So what’s the fuss about? Why is everyone panicking?

The reason for the panic is simple. DeepSeek and Kimi showed that you don’t need to be a huge tech company to compete in AI.

You don’t need billion-dollar data centers. You don’t need a team of thousands of highly paid AI engineers. Around 75% of Nvidia employees are reportedly millionaires due to their salaries and stock-based compensation.

The result is that investors and shareholders are once again questioning and scrutinizing the current level of immense CAPEX spending. Billions upon billions have been invested in infrastructure and personnel. Many more billions are planned. But DeepSeek and Kimi showed that a lot is possible with far less.

That’s why Wall Street is panicking.

Nobody knows what will happen next. The selloff could easily continue if the market becomes more emotional, which leads to overreactions.

However, in the long run, there is no way around AI.

The best way to think about AI is as a foundational layer that others will build on top of. Think railways, mobile phones, or electricity. It was always clear that AI models themselves would eventually become commoditized.

So what should investors look out for?

There is no real moat around the AI models themselves. As we’ve seen, almost anyone can create their own models for less, often using existing models and technology as a reference. It is now becoming a race to the bottom for research labs.

Of course, the 2 giants, OpenAI and Anthropic, won’t stand still. They will try to capture as much market share as possible by integrating deeper into the broader ecosystem.

But what now? I think there are 2 key points.

1. First, we will still need AI infrastructure. Maybe the short-term hype has been dampened somewhat. But after the internet bubble, nobody argued that we should stop building internet infrastructure.

2. Second, most of the value will be created at the application layer. What does that mean? Basically, the internet created an entirely new pathway for value creation in software.

The next major companies will create completely new categories of applications. I think many of them will be built around AI agents.

Imagine an army of agents doing work for you, connected to your own data, which you can orchestrate for pennies, 24/7. They will replace some white-collar jobs while augmenting many others.

In short, while the short-term hype has been dampened, DeepSeek and Kimi are a net positive. They have shown everyone that incredible AI models can be built even with limited resources.

Fewer resources spent on training and inference means more resources can be directed toward new use cases. And the biggest beneficiaries will be the companies creating entirely new applications.

The path toward artificial general intelligence just accelerated.

And the last time this happened and the market overreacted, turned out to be a very good buying opportunity. But it’s important to be patient, build your watchlist, and wait for the sell-off to calm down.