jakiro
Author: Hedy Bi
The Hong Kong Web3 Carnival has come to an end. However, the free pulse of Web3 is still beating and continuously penetrating into other industries.Compared with the previous cycle, the logic of the opening of this round of bull market is“Native Innovation narrative” transforms into a model of “mainstream recognition, capital -driven”EssenceThe web3 development stage observed by the author also observed from one“The absolute freedom of closed and niche” has evolved into the stage of “really tolerance”.
Do not jump out of the box to analyze under such logic,Shouzhang’s native innovation narrative can no longer adapt to the current development of Web3.Since the entire Web3 hugging compliance, web3 has been re -focused on the financial sector, driven by the Hong Kong government.The mainstream financial institutions also accelerate the participation of Web3 through RWA and spot ETF.
In this conference, in addition to the mainstream financial institutions entering Web3, we also saw an opportunity to connect to Web2 and Web3 -the DePin track.In particular, the development of the development of large AI models has made the sub -tracks in Depin -the re -distribution of computing power has become a popular concern again.
>
Source: OKG Research
The computing power is bait, but the AI model training is not the best scene of DEPIN
“Blockchain is to build trust through technology, and AI is an industry that needs to be trusted extremely.” Dragonfly Capital’s management partner HASEED QURESHI said at the conference.
DEPIN is not a new track, and it has been proposed a few years ago.It is the outbreak of AI models that the industry has conducted a lot of discussions on computing power and data. According to calculations, the cost of large models calculated increased by 31 times per year.The global GPU shortage, companies like Nvidia have great pricing power at the top of the food chain at the current market demand.The monopoly or decentralization is the reason why the cost of cost has become a heated discussion on the Web3 DEPIN track.
Although the AI model training is the cause, Rome was built in a day,AI big model training is not the best landing scene of Depin at presentEssenceThe requirements for computing power in the production of large models of AI are mainly around two aspects: reasoning and training.In the training session, through feeding a lot of data, a complex neural network model is trained.In the reasoning session, use a large amount of data to reason by using the training model.
>
Source: nvidia
The combination of decentralization and computing power, the difficulty coefficient from training to fine -tuning training to reasoning is decreasing layer by layer.In DEPIN, you can see that more projects in the industry are concentrated in reasoning rather than training.The main reasons for the Nvida GPU cluster used by most enterprises’ AI training are their own powerful parallel computing capabilities and memory bandwidth.Compared with the reasoning link, the requirements for parallel computing power and bandwidth are reduced a lot.And large model training values stability, because once the training is interrupted, training needs to be re -training.If a decentralized computing power is used for GPT on Ethereum, only a single matrix multiplication operation will consume up to $ 10 billion in GAS and take 1 month.
In addition, the author analyzes several popular projects in this conference.A trend that shows the supply side exceeding the demand side, that is, the needs of AI developers who are distributed in the world’s computing power more than AI model training or reasoning tasks around the world.It is not to say that demand does not exist. OPENAI’s founder Sam Altman proposes to raise $ 7 trillion to build a advanced chip factory that exceeds TSMC 10 times the size of TSMC and the production and model training for chips.Stanford University’s research also shows that no matter which language model, when the scale of the training parameters exceeds the critical value of that scale, its performance (such as accuracy) has increased sharply.This is completely opposite to the law of “strong miracles”, which also means that in reality, there are many difficulties to solve the decentralized computing ideas.
The “historical traceability” of the DEPIN track dates back to the “sharing economy”
The concept of DEPIN itself is not difficult to understand, and it can even be traced back to the web2 to review the Internet industry. At least 15 years, web2 players are immersed in gathering personal tangible assets to create a “sharing economy”.If it is re -assigned to the demander directly through the form of the intangible asset (such as the idle server, etc.) directly through the form of Peer to Peer (P2P) or the form of Peer to Business (P2B).A optimization of production relations.This is what DEPIN is going to do.
Because of this, below the DEPIN track, everyone’s enthusiasm is high.In fact, web2 has accumulated for a long time for “re -distribution”, but this time, the intermediary is removed directly.At present, nearly a thousand DEPIN projects, especially the Solana ecology. According to MESSARI, the Solana ecology is in a leading position in the DEPIN infrastructure. This is because the Solana public chain has high infrastructure integration and performance.In terms of regional distribution, it is expected that the top 10 DEPIN will be from Asia from 2024 to 2025.
>
Source: MESSARI
Web3 and AI have a lot of cross points. The computing power, as the universal currency of the future digital world, is the first to pay attention to.But the most reasonable landing scene of decentralized computing power is not the easiest to land.
At the intersection of web3 and AI, in addition to the technical technically overcome difficulties and continuously break through such problems,There are also many other branches, such as AI agents to give creators ownership and small AI model computing power scenes worth exploring, and they will be more landing.The success of the business model and the breakthrough of technology will always be balanced.DEPIN is accelerating this process, and DEPIN’s “hunting journey” will return full.
