Author: Paul Veradittakit Source: VeradiVerdict Translation: Shan Ouba, Golden Finance
DeepSeek was originally a side project of the High-Flyer hedge fund, started by a few thousand Nvidia GPUs. DeepSeek R1 cost more than $1 billion, used 2,000 Nvidia H800 GPUs, and took 55 days to beat the benchmark of OpenAI's o1 model, which required hundreds of billions of dollars and more than 16,000 advanced GPUs to develop, which shocked the world.
DeepSeek R1 has 671 billion parameters, while GPT-4 has 1.76 trillion parameters. OpenAI’s large models may require thousands of GPUs for training and high-end clusters for inference, while DeepSeek 7B and 67B can run on consumer-grade hardware (a handful of A100s or H100s).
DeepSeek-R1 vs. Competitors
In response, Nvidia’s stock price fell 18% ($600 billion loss in market value). The old idea that AI models must be closed source and computationally expensive to succeed is breaking down.
Existing decentralized AI narrative
Demystifying the Crypto x AI stack
The AI x Crypto project believes that crowdsourced, public, decentralized AI will ultimately create better models than centralized AI.
So far, this has not been the case, as the highest performing models have come from closed-source companies such as OpenAI and Anthropic. Crypto x AI companies have adapted to this situation, focusing on infrastructure rather than model building.
For example, GPU marketplaces such as Akash, Render, IoNet, and Exabits have achieved sustainable revenue. Companies that allow users to share network bandwidth, such as Grass and Gradient, have found their niche, providing services such as distributed web scraping to web2 clients. Storage networks such as Arweave, Filecoin, and Ocean have also achieved good results by becoming the building platform for these projects. Supply networks have flourished because they are able to tailor cheaper and more scalable services to off-chain customers.
Messari Report on DePin and AI x Crypto
Now that GPUs and financial resources are no longer a limitation to creating high-quality AI models, web3 AI companies can focus on replicating the effectiveness of DeepSeek while providing new advantages such as modality, user ownership, censorship resistance, privacy, etc.
Pantera provides funding for companies such as SaharaAI and Sentient, believing that these companies can match or exceed the performance of traditional AI companies while remaining competitive by providing other services.
For example, Sahara AI is building a platform where anyone can monetize AI models, datasets, and applications in a collaborative space. Users can manually train models permissionlessly, provide training data, and create custom AI models using no-code tools. They can only meet the needs of all these stakeholders (AI developers, users, resource providers) because everything is tied to their native Sahara blockchain.
The future of AI will be built by Web3 infrastructure
DeepSeek has proven that high-performance AI models do not have to be built in a closed-source way or have to bear the computational costs. I am excited to see AI models developed using the full web3 AI stack next year.
I believe that supply-side projects will continue to grow, and consumer-facing projects can use their ability to build networks that invite community participation to start competing with web2 competitors. For example, users are willing to sign up to train web3 models and actively try them out through hackathons and grant programs.
In the near future, I believe the best performing AI models will be built on-chain.
Weatherly
