A few days ago, I had a video call with an old classmate who works on large model architecture in Silicon Valley. He has been losing sleep over the shortage of a few thousand H100s recently, and he’s lost a lot of hair.

I half-jokingly suggested to him: “Why not try the hottest decentralized computing (DePIN) on the chain right now? The price is only one-third of AWS, with tens of thousands of nodes distributed globally, promoting ‘computing power equity.’”

He finished listening, first stunned for three seconds, then let out a nearly self-deprecating bitter laugh: “Buddy, are you kidding me? The model weights I’m running are built from billions of dollars, they are the lifeblood of the company. You want me to chop up this data and throw it onto those homemade graphics cards that can't even guarantee stability and might drop out at any moment? That's like having a precise surgery that requires a sterile operating room, and you take me to an outdoor stall that doesn't even have a sanitized kitchen knife.”

This smile directly tore down the current DePIN track's most self-indulgent fig leaf.

1. Precision and Efficiency: Algorithm science does not believe in 'distributed sentiment.'

Currently, there is a popular narrative in the circle: as long as you connect all the idle graphics cards in the world, you can piece together a 'world computer' that surpasses supercomputing centers.@MidnightNetwork

This sounds exhilarating, but in the eyes of real AI engineers, it is simply a pipe dream.

The current public chain architecture is essentially using extremely inefficient redundancy to achieve so-called 'decentralization.' Large model training requires extremely low-latency InfiniBand interconnects and microsecond-level parameter synchronization. However, the current Web3 computing power network has nodes distributed in basements in Seattle and apartments in Shanghai, separated by thousands of kilometers of physical distance and uncontrollable network jitter.

This has led to an extremely absurd situation: you spend 100 yuan on the chain to buy computing power, and as a result, 80 yuan is consumed in the communication pull between nodes. This efficiency, in the time-sensitive AI competition, is not a boost but a suicide.

2. Security and Trust: Who dares to hand over the 'brain' to a black box?

The training data and model weights of a large model are the most expensive commercial secrets of this era. In traditional centralized data centers, there are physical firewalls and legal contracts to back them up. But in the current decentralized network, due to the lack of effective Trusted Execution Environments (TEE) and computing power integrity verification, you have no idea whether the node that took your task is helping you calculate logic or secretly cloning your weights.

Currently, the DePIN track is mostly still caught up in the numerical game of 'how many thousands of graphics cards I have,' yet no one is addressing the most core issue: how to safely and efficiently complete a task with a hundred billion-dollar AI model on unfamiliar nodes that do not trust each other?

3. @Akash or @Io.net? Who is stitching this裂痕$NIGHT

While everyone is busy hyping the soaring computing power tokens, the truly rational players have started to pay attention to projects like Akash or Livepeer that attempt to reconstruct the 'computing power customs' from the ground up.

They are no longer obsessed with connecting individual gaming PCs, but instead are turning to the standardized integration of enterprise-grade idle data centers. This may not sound so 'decentralized,' and even a bit dull, but this is the true interface that can connect to the traditional AI industry.

What they are doing is processing those scattered, non-standard 'raw materials' into 'standardized power' that can be directly called by large companies through a rigorous verification protocol and scheduling algorithm.

• For developers: What they feel is a silky experience like cloud services, even unaware of the blockchain behind it.

• For computing power providers: Must pass specific hardware and software certifications, and can't just come in to fleece sheep by pulling a few broadband lines.

Conclusion: Don't talk about lab purity in nightclubs.

The current Web3 computing power track indeed resembles selling high-end microscopes in nightclubs. Retail investors care about when the tokens will be listed on Binance, while those who truly need computing power—AI startups—are standing at the door, coldly watching this group revel in the chaos of latency and data leak risks.

#night I do not deny the future of DePIN; on the contrary, I believe it could be one of the greatest narratives of the next decade. But the logic must be reconstructed: computing power is not a commodity, trust is.