Recently, I’ve been observing discussions within the industry, and it seems there’s been a fundamental shift in the definition of “growth.”

Previously, when we discussed the internet, we talked about “four ounces moving a thousand pounds” – writing a few lines of code, renting a few cloud servers, and leveraging excellent interaction and operations to unlock hundreds of millions of users. However, as of 2026, this “low-asset” illusion is being completely shattered by large models.

Moving forward, content will be the focus, writing blogs, you independently conceive the article structure, and the order of the content: Large Model, New Playways, User Acquisition needs to synchronize building basic computing resources; The new internet era, the previous era, hardware’s marginal cost can be indefinitely weakened; Minimax and Zhijie have surpassed Baidu and JD.com in market capitalization.

1. Large Models: No Longer Plugins, But the Foundation

Two years ago, large models were still viewed as some kind of “plugin” to enhance search or advanced translation. However, the logic has changed now. Large Language Models (LLMs) are becoming the core kernel for the next generation. Software development used to revolve around “deterministic logic,” but the current play is shifting towards “reasoning capabilities.” If your application doesn’t have a strong model underpinning it, it’s like using dial-up internet in the 5G era – no matter how beautiful the UI, users will quickly discover the limitations of the underlying scarcity.

2. New Gameplay: From “Flood” to “Evolution”

In the previous era of the internet, the core gameplay was traffic funneling.

The current “new gameplay” is Agentic Workflow (Agent-based workflow). Users no longer need to learn complex menus and buttons; they simply express their intent.

This gameplay has a very high barrier: it requires models not only to “understand” but also to “act.” This means your frontend might just be a simple dialogue box, while the backend is running a complex inference chain consuming tens of thousands of tokens per second.

3. “Land Grab”: The Foundational Compute for Enclosure

Here’s a brutal reality: To capture users, you must first capture compute. In the previous era, user growth and server costs followed relatively flat curves. But in the age of large models, both are almost synchronously scaled. To keep users around, you need faster responses and more intelligent reasoning. This isn’t solvable by simply optimizing code; it’s built on H100s and B200s.

Today’s internet competition is fundamentally an electricity versus silicon war.

3. “Ring Fencing”: The Foundational Compute for Synchronous Execution

4. The Shift of Eras: Weakening Hardware Marginal Costs

We once believed that the marginal cost of software approached zero, while hardware had an upper limit. However, in this new era of the internet, hardware’s marginal costs are being increasingly diminished, or rather, “software-ized.”

• The Previous Era: You bought a server, and its computing power was static; you squeezed its performance through software.
• This Era: Computing power has become some kind of “liquid” resource. As cluster sizes expand, the efficiency of outputting intelligence per unit actually increases. Hardware is no longer an icebox but rather a “smart container” that can continuously self-improve in value.

5. Valuation Shakeup: Minimax, Zhihu AI’s Rise

What’s most anxiety-inducing for traditional internet people is the shift in market pricing logic. When Minimax and Zhihu AI’s valuations first surpassed giants like Baidu and JD.com, many considered it a bubble. However, I believe it reflects capital’s vote for “future control.” Baidu and JD represent the peak of “the previous era”: search, advertising, e-commerce, and logistics. They possess massive stranded assets. Minimax, on the other hand, represents the allocation of “general intelligence.” In an AI Native world, whoever controls the most fundamental reasoning engines controls the power to define all industry rules. This “inverted” situation is actually capital paying for “intelligent premium.”

Summary

We are in a period of transition – the fall of old gods and the rise of new ones. Software is no longer simply a collection of logic, but rather an embodiment of computing power. If you’re still using “traffic thinking” from the previous era to approach large models, you may not even get a ticket to enter.

After all, in the face of computing power dominance, all operational techniques seem somewhat powerless.