So, let's get straight to the point. On Friday, a Chinese company unveiled a new artificial intelligence model. Within hours, Wall Street panicked. The Nasdaq slipped, semiconductor stocks came under pressure, and nearly $392 billion was wiped from the valuations of the two largest American AI companies.
Then what happened?
The very same Chinese company announced it was suspending new subscriptions because it couldn't handle the surge in demand.
The company is called Moonshot AI. It's based in Beijing, was founded in 2023 by Yang Zhilin, and counts Alibaba among its investors. Its new model is called Kimi K3, and it contains 2.8 trillion parameters.
So what does that actually mean?
Put simply, parameters are the numerical settings that allow an AI model to recognize patterns and generate answers. They reflect the size and complexity of the system. But here's an important point: bigger doesn't automatically mean smarter.
What really caught the attention of American companies wasn't the size of the model. It was the fact that Kimi K3 is open source.
In other words, Moonshot AI is making the model's core weights publicly available, allowing anyone to download it, inspect it, and build on top of it. OpenAI and Anthropic take the exact opposite approach. They keep their models closed and monetize them through subscriptions and usage fees.
And that's where the problem begins for them.
Because if a model that's almost as capable is freely available, how do you justify premium pricing?
The market reacted immediately. The Nasdaq fell about 1%, with semiconductor stocks, including Nvidia, taking the biggest hit.
Meanwhile, in the private markets, where companies are valued before going public, the damage was even greater. Anthropic lost roughly $232 billion in valuation, while OpenAI lost around $160 billion.
Combined, that's nearly $392 billion erased in just a few days.
Yes, you read that correctly.
The market's reasoning was straightforward.
If AI becomes cheaper and more accessible, then demand for expensive chips should decline.
So investors sold semiconductor stocks.
We've seen this movie before.
Back in early 2025, DeepSeek triggered almost the exact same reaction.
Same story.
Same panic.
And once again, it appears the market may have overreacted.
Within just 48 hours, demand for Kimi K3 exceeded the company's own expectations and pushed its computing infrastructure to the limit.
Moonshot AI was forced to suspend new subscriptions, prioritize existing users, and split its offering into two separate plans, including one dedicated exclusively to programming.
The company even joked publicly:
"Kimi K3 received far more love than we expected, and our GPUs are definitely feeling it."
You might be wondering:
"Why didn't they just add more servers?"
Because they couldn't.
U.S. export restrictions on Nvidia's most advanced AI chips have made computing power the biggest bottleneck for Chinese AI companies.
And this is where things become even more interesting.
According to Omdia, Kimi K3 is extremely demanding in terms of computing resources.
Its 2.8 trillion parameters require more than 1.5 terabytes of HBM memory, along with massive GPU clusters, storage capacity, and high performance networking.
So what does that actually mean?
Very few organizations in the world can realistically run this model on their own, even though it's open source.
Open source doesn't mean free.
It simply means you're given the model.
You still have to pay for the hardware, electricity, and infrastructure.
In fact, Kimi K3 also consumes roughly 67% more tokens per task than GPT-5.6 Sol.
So in practice, the "cheap" model may not be nearly as inexpensive as it first appears.
And this brings us to the most important takeaway from the entire story.
Investor Gavin Baker explains it well.
Kimi doesn't need to be the cheapest model to disrupt the market.
It only needs to get close.
Because that puts pressure on the profit margins of companies developing frontier AI models.
At the same time, the cheaper AI becomes, the more people use it.
And the more AI gets used, the more infrastructure the world needs.
In other words, value shifts.
It moves away from the companies building the models and toward those selling the chips, electricity, cloud infrastructure, and software that power them.
And here's something worth paying attention to.
Nobody is cutting back on investment.
OpenAI and Nvidia have announced plans involving at least 10 gigawatts of AI infrastructure.
Anthropic has secured agreements totaling up to 10 gigawatts through Amazon, Google, and Broadcom.
And remember AMD's partnership with Microsoft that we discussed in yesterday's newsletter.
They're all pieces of the same puzzle.
Microsoft, in particular, is in a very strong position.
Through Azure and Foundry, it can switch between AI models without losing customers.
So even if model prices fall while usage keeps rising, Microsoft still comes out ahead.
