These energy consumption and environmental issues caused by the development of data and AI are an issue that continues to receive attention. This issue is also accompanied by the blockchain industry's search for alternatives to "Proof of Work (PoW)" - traditional mining is too energy-consuming and does not generate actual value. At the same time, AI is developing rapidly, especially after 2022, the demand for computing power and chips in the field of AI, such as large models, reasoning services, and distributed training, has increased sharply.

This article focuses on the decentralized AI ecosystem, which is regarded as a new field, and analyzes how energy consumption and environmental issues coexist with the development of the AI ecosystem, how the two affect each other, and what kind of situation will it eventually present?

The core concept of the decentralized AI ecosystem can be simply summarized in one sentence: Let tens of millions of GPUs, NPUs, and TPUs around the world not only be used to calculate hashes, but also to train models, run inferences, and generate zero-knowledge proofs required for privacy protection. That is, turn the originally "wasted" resources (computing, storage, communication, network, etc.) into a force to promote the development of AI.

Currently, the three pillars of decentralized AI have been initially formed by the distributed training network, inference network, and GPU computing power market.

Three Pillars of Decentralized AI

Distributed Training Network

Distributed Training Network Training a large model requires coordinating hundreds or thousands of nodes, each of which handles gradient calculations, parameter synchronization, data distribution, etc. The goal is to break the data center's monopoly on large model training and allow anyone to contribute computing power and benefit from it.

Distributed Inference Network

Once the model training is completed, the traditional approach is to deploy the model to a centralized cloud server, such as AWS or Google Cloud. But decentralized AI requires a new way: to distribute the inference tasks to nodes around the world for execution, just like edge computing, which is closer to users, faster in response and lower in cost. This model is particularly suitable for applications that want to reduce latency and improve accessibility, such as chatbots, image recognition, and speech-to-text.

GPU computing power market

Whether training or inference, one thing is needed: computing power. In reality, a large number of consumer-grade devices (such as laptops in your hands), small mining farms and even idle game consoles are "sleeping". Therefore, GPU computing power trading platforms came into being. They are like Uber, scheduling idle resources and providing them to those who need them. These platforms not only provide developers with a cheap and flexible source of computing power, but also open up new revenue channels for hardware owners. Distributed training and inference networks rely on the distributed GPU computing market as their underlying infrastructure layer.

Why is decentralized AI a trend

Decentralized AI means more efficient use of global computing resources, stronger decentralization capabilities, and the natural advantages of privacy and security, which will naturally become the trend of the future.

We no longer need to build mining farms to run meaningless hash operations, but can invest these resources in truly valuable AI tasks. Just likeyour laptop automatically joins a training network when you sleep at night, and continues to use it normally during the day, and can also get some token rewards.

In terms ofstronger decentralization capabilities,traditional AI training and reasoning are highly dependent on a few technology giants, and decentralized AI networks break this monopoly pattern, allowing more people to participate and form true "crowdsourcing intelligence".

And due to the widespread application of privacy technologies such as Zero-knowledge proof (ZKP) technology, the protection of raw dataduring model training or inference is particularly important for sensitive scenarios such as finance and medical care.

As we mentioned before, decentralized AI and decentralized physical infrastructure networks (DePIN) also face the same challenges: communication problems between nodes, synchronization problems of model parameters, and compatibility problems between heterogeneous devices.

DePINbringsefficient energy utilization

DePIN (decentralized physical infrastructure network) is about sharing resources - aggregating idle resources and realizing market-oriented operation according to effective mechanisms. For example, decentralized cloud computing is to aggregate idle computing resources through blockchain, smart contracts and other technologies to realize the market-oriented operation of cloud computing resources. This sharing method avoids the waste of resources, thus playing a certain role in the efficient use of energy and environmental protection.

The other side of decentralization is to tilt towards clean energy and renewable energy nodes through policies or incentives, thereby stimulating more resources from renewable energy. For example, for data centers built with nuclear power, since they use clean energy, decentralized cloud computing platforms can "green mark" their computing nodes on the platform, tilt more incentives towards them, and invisibly encourage more computing power providers to adopt cloud computing power based on clean energy construction, and computing power demanders to give priority to purchasing "green computing resources".

The decentralized AI ecosystem is built on a decentralized physical infrastructure network. This is the relatively balanced forward trend of energy and AI ecosystem development that we can see at present. On the one hand, we need to build more physical facilities that rely on clean energy and renewable energy. On the other hand, we need DePIN to organize and optimize resources to make them operate more efficiently.