DeepSeek bursts the Agent track bubble, DeFAI may become a new growth point for Web3 AI

Author: Kevin, the Researcher at BlockBooster

TLDR:

• The emergence of DeepSeek has broken the computing power moat, and computing power optimization led by open source models has become a new direction;

• DeepSeek is beneficial to the model layer and application layer in the upstream and downstream of the industry, and has a negative impact on the computing power protocol in the infrastructure;

• DeepSeek's favorableness inadvertently pierces the last bubble of the Agent track, and DeFAI is most likely to give birth to newborns;

• The zero-sum game of project financing is expected to come to an end, and a new financing method of community launch + a small number of VCs may become the norm.

The impact caused by DeepSeek will have a profound impact on the upstream and downstream of the AI ​​industry this year. DeepSeek successfully allowed home consumer graphics cards to complete the big model training tasks that could only be undertaken by a large number of high-end GPUs. The first moat around the development of AI, computing power, began to collapse. When algorithmic efficiency ran wildly at a rate of 68% per year, and hardware performance followed the linear climb of Moore's Law, the deep-rooted valuation model in the past three years was no longer applicable. The next chapter of the company will be opened by the open source model.

Although the AI ​​protocol of Web3 is completely different from Web2, it is inevitably affected by DeepSeek, which will give birth to new use cases for the upstream and downstream of Web3 AI: infrastructure layer, middleware layer, model layer and application layer. .

Clarify the collaboration relationship between upstream and downstream

agreements

Through the analysis of technical architecture, functional positioning and actual use cases, I divided the entire ecosystem into: infrastructure layer, middleware layer, model layer, application layer, and sorted out their dependencies:

Infrastructure layer

The infrastructure layer provides decentralized underlying resources (computing power, storage, L1), among which computing power protocols include: Render, Akash, io.net, etc.; storage protocols include: Arweave, Filecoin, Storj, etc.; L1 includes: NEAR, Olas, Fetch.ai, etc.

The computing power layer protocol supports model training, inference and framework operation; the storage protocol saves training data, model parameters and on-chain interaction records; L1 optimizes data transmission efficiency through special nodes to reduce latency.

Middleware layer

The middleware layer is a bridge connecting infrastructure and upper-level applications, providing framework development tools, data services and privacy protection. The data annotation protocols include: Grass, Masa, Vana, etc.; development framework protocols include: Eliza, ARC, Swarms, etc.; privacy The computing protocols include: Phala et al.

The data service layer provides fuel for model training, the development framework relies on the computing power and storage of the infrastructure layer, and the privacy computing layer protects the security of data in training/inference.

Model layer

The model layer is used for model development, training and distribution, with the open source model training platform: Bittensor.

The model layer relies on the computing power of the infrastructure layer and the data of the middleware layer; the model is deployed on the chain through the development framework; the model market delivers training results to the application layer.

Application layer

The application layer is an AI product for end users, among which the agents include: GOAT, AIXBT, etc.; the DeFAI protocols include: Griffain, Buzz, etc.

The application layer calls the pre-trained model of the model layer; it relies on the privacy computing of the middleware layer; complex applications require real-time computing power of the infrastructure layer.

DeepSeek may have a negative impact on decentralized computing power

According to the sampling survey, about 70% of Web3 AI projects actually call OpenAI or centralized cloud platforms, only 15% of projects use decentralized GPUs (such as Bittensor subnet model), and the remaining 15% are hybrid architectures (sensitive data locally processed , general tasks to the cloud).

The actual usage rate of the decentralized computing power protocol is much lower than expected and does not match its actual market value. There are three reasons for the low usage rate: Web2 developers continue to use the original tool chain when they migrate to Web3; the decentralized GPU platform has not yet achieved price advantages; some projects evade data compliance review in the name of "decentralization", which is actually calculated. The power still relies on centralized cloud.

AWS/GCP accounts for 90%+ of AI computing power, compared with that of Akash's equivalent computing power is only 0.2% of AWS. The moat of centralized cloud platform includes: cluster management, RDMA high-speed network, elastic scaling; the decentralized cloud platform has improved web3 versions of the above technologies, but the defects that cannot be improved include: the delay problem of distributed node communication is centralized. 6 times the cloud; tool chain split: PyTorch/TensorFlow does not natively support decentralized scheduling.

DeepSeek reduces computing power consumption by 50% through Sparse Training, and dynamic model pruning realizes consumer-level GPU training for 10 billion parameter models. The market's expectations for high-end GPU demand in the short term have been significantly reduced, and the market potential of edge computing has been revalued. As shown in the figure above, before the emergence of DeepSeek, most of the protocols and applications in the industry used platforms such as AWS, and only a few use cases were deployed in decentralized GPU networks. Such use cases valued the latter in consumer computing power. Price advantage and no attention to the impact of delays.

This situation may worsen further with the emergence of DeepSeek. DeepSeek has released the limitations of long-tail developers, and low-cost and efficient inference models will be popularized at an unprecedented speed. In fact, the above-mentioned centralized cloud platforms and many countries have begun to deploy DeepSeek. The significant reduction in inference costs will lead to a large number of Front-end applications, these applications have huge demands for consumer-grade GPUs. Faced with the huge market that is coming, centralized cloud platforms will launch a new round of user competition, not only competing with leading platforms, but also competing with countless small centralized cloud platforms. The most direct way to compete is to reduce prices. It can be foreseeable that the price of 4090 on the centralized platform will be reduced, which is a catastrophe for the computing power platform of Web3. When the price is not the only moat of the latter and the computing power platforms in the industry are forced to lower prices, the result is that io.net , Render, and Akash cannot bear it. The price war will destroy the latter's only remaining valuation ceiling, and the death spiral caused by decline in returns and user loss may give the decentralized computing power protocol a new direction.

The specific significance of DeepSeek to the industry's upstream and

downstream agreements

As shown in the figure, I think DeepSeek will have different impacts on the infrastructure layer, model layer, and application layer, from a positive perspective:

The application layer will benefit from a significant reduction in inference costs. More applications can use low costs to ensure that the Agent application is online for a long time and complete tasks in real time;

At the same time, low-cost model overhead such as DeepSeek can allow the DeFAI protocol to form a more complex SWARM. Thousands of agents are used in a use case, and the division of labor of each agent will be very subtle and clear, which can greatly improve the user experience and avoid User input is erroneously disassembled and executed by the model;

Application layer developers can fine-tune the model and feed DeFi-related AI applications prices, on-chain data and analysis, and protocol governance data without having to pay high license fees.

After the birth of DeepSeek, the existence significance of the open source model layer was proven. Opening high-end models to long-tail developers can stimulate a wide range of development booms;

The computing power wall built around high-end GPUs in the past three years has been completely broken. Developers have more choices and establish a direction for open source models. In the future, AI models will no longer compete with computing power but algorithms, and the transformation of beliefs will become The cornerstone of confidence for open source model developers;

Specific subnets around DeepSeek will emerge one after another, and the model parameters under the same computing power will increase, and more developers will join the open source community.

In terms of negative effects:

The objective use delay of computing power protocols in the infrastructure cannot be optimized;

Moreover, the hybrid network composed of A100 and 4090 has higher requirements for coordination algorithms, which is not the advantage of a decentralized platform.

DeepSeek bursts the last bubble of the Agent track, DeFAI may give birth

to newborns, and the industry's financing methods are ushering in a transformation

Agent is the last hope for AI in the industry. The emergence of DeepSeek liberates the limitations of computing power and depicts the future expectations of the explosion of applications. This was a huge benefit to the Agent track, but due to the strong connection between the industry, US stocks and Federal Reserve policies, the remaining bubble was burst, and the track's market value fell to the bottom.

In the wave of integration between AI and the industry, technological breakthroughs and market games have always been with us. The chain reaction caused by Nvidia's market value fluctuations is like a mirror that reflects the deep dilemma of AI narrative in the industry: From On-chain Agent to DeFAI engine, under the seemingly complete ecological map, it covers the weak technological infrastructure, hollowing out value logic, and capital. The cruel reality that dominates. The seemingly prosperous on-chain ecosystem is hidden hidden diseases: a large number of high FDV tokens compete for limited liquidity, old assets rely on FOMO emotions to survive, developers are trapped in PVP intravolume to consume innovation potential. When incremental capital and user growth hit the ceiling, the entire industry fell into the "innovator's dilemma" - both eager for breakthrough narratives and difficult to get rid of the shackles of path dependence. This tearing state provides a historic opportunity for AI Agent: it is not only an upgrade of the technology toolbox, but also a reconstruction of the value creation paradigm.

In the past year, more and more teams in the industry have found that the traditional financing model is ineffective - the routine of giving VCs a small share, high control of the market, etc. is no longer sustainable. VC pockets are tightened, retail investors refuse to take over, and the threshold for big souvenirs is high. Under the triple pressure, a new gameplay that is more suitable for the bear market is rising: the joint top KOL + a small number of VCs, a large proportion of community launches, and a low market value starts coldly.

Innovators represented by Soon and Pump Fun are opening up new paths through "community launch" - joint head KOL endorsement, distributing 40%-60% of tokens directly to the community, at a valuation of as low as $10 million FDV Water level starts the project, achieving multi-million dollar financing. This model builds a consensus FOMO through KOL's influence, allowing the team to lock in profits in advance, and at the same time exchanges high liquidity for market depth. Although it gives up the advantages of short-term control, it can repurchase tokens in bear markets at a low price through a compliant market. In essence, this is the paradigm migration of power structure: from the VC-led drum-passing game (institutional takeover - sold above - retail investors pay), to a transparent game of community consensus pricing, the project party and the community form a new symbiosis in liquidity premium. relation. When the industry enters a revolutionary cycle of transparency and projects that are obsessed with traditional control logic may become a afterimage of the era under the wave of power migration.

The short-term pain in the market just confirms the irreversibility of the long-term technology trend. When AI Agent reduces the cost of on-chain interaction by two orders of magnitude, and when the adaptive model continues to optimize the funding efficiency of the DeFi protocol, the industry is expected to usher in the long-awaited Massive Adoption. This change does not rely on concept hype or capital to ripen, but is rooted in technological penetration of real needs - just like the electric revolution has not stagnated by the bankruptcy of light bulb companies, Agent will eventually become a real gold race after the bubble bursts road. And DeFAI may be the fertile ground for nurturing newborns. When low-cost reasoning becomes daily, we may soon see the birth of use cases where hundreds of agents are combined into a Swarm. Under equivalent computing power, a significant increase in model parameters can ensure that the agents in the open source model era can be fine-tuned more fully, and even in the face of user complex input instructions, they can be split into tasks that a single agent can fully execute. Optimizing operations on the chain for each Agent may promote an increase in overall DeFi protocol activity and increased liquidity. More complex DeFi products led by DeFAI will appear, and this is where new opportunities emerge after the bursting of the last round of bubble.