coindesk.com
In November of 2024, I brought together five founders from CoinFund’s portfolio for a roundtable discussion on taking a decentralized approach to implementing AI’s technology stack. This stack is not trivial: it starts with a foundation of GPU compute aggregation, runs into the difficult problems of decentralized AI training and cost effective inference, and raises questions about decentralized data acquisition and AI consumer product development.
What was clear from this discussion was that our decentralized AI (deAI) founders are not just selling AI to Web3; they are pushing forward the state-of-art of AI itself. Moreover, it dawned on me that, in the near future, the best bet for retail investors to get financial exposure to frontier AI will not be through public stocks or private placements – but through digital assets. Leaving the conversation feeling inspired about the innovation and the growing financial opportunity, I told the group that we would see an explosion of growth in the coming months, calling it “deAI Summer 2025.”
Eight months later, it’s finally here.
We are now in a world where proof-of-concept models have already been pre-trained and post-trained on decentralized networks. It’s very likely that 100 billion parameter pre-training on decentralized networks will be demonstrated this year. Some companies believe we can get to a frontier intelligence model on decentralized networks using post-training and reinforcement learning. Even Jack Clark, the co-founder of Anthropic, wrote about decentralized training on his personal blog.
AI development is at an inflection point. As large companies like OpenAI, Microsoft and Google are looking to capture the consumer as a top line goal, more users are employing AI tools in their workflows and personal lives. But with centralized corporations at the helm, there is more private data at risk than ever before. It’s one of the reasons that deAI has emerged as a key category for Web3, and has the potential to fundamentally change the way models are built and owned.
According to data from analytics platform Kaito, in the past 12 months, deAI has taken up over 30% of mindshare across crypto. Companies building distributed compute protocols, agent frameworks, and decentralized marketplaces are dominating this intersection. Since there are few public equities focused on direct exposure to frontier AI technology – and most retail investors do not have access to the private rounds of frontier labs like OpenAI or Anthropic – deAI cryptoassets are emerging as a highly sought asset class for investments into AI companies.
While there’s a lot of hype surrounding deAI, at CoinFund, we’ve worked to cut through the noise and ask ourselves which teams are solving the most impactful and most difficult problems. Among those is the issue of whether it is possible to train large, competitive models on decentralized networks.
Training very large models seems to require high-end, memory-rich GPUs with fast communication bandwidth, as training nodes continually process throughput that exceeds many multiples of the size of the entire internet. Historically, developers in this field have made tremendous efforts to lower these bandwidth requirements, but have not found an effective approach.
For example, previous attempts at compressing this bandwidth have derailed training processes, as lossy compression caused models to fail to converge. Everyone assumed there was not a viable solution to lowering bandwidth requirements to the point that a training process could take place on regular GPU hardware on slow internet connections. Accordingly, most experts agreed that decentralized training was a dead end.
We disagreed.
Our thesis
CoinFund’s thesis on decentralized AI is that Web3 networks will enable AI models to be crowd-resourced, trained, owned, and maintained as open and valuable public goods. These networks will aggregate record amounts of compute, compete on the frontier of AI, and create massive innovation in AI broadly by solving problems like bandwidth optimization, error detection, fault tolerance, distribution, sustainable open source and sharding of models.
Additionally, these networks – while open – will also have business models. For example, decentralized AI models will live in the network and users will pay the network for inference on them. These networks will incentivize and align users by compensating them for funding and training models and giving them a share in the network.
In CoinFund’s portfolio, we’ve backed decentralized training companies like GenSyn, Prime Intellect and Pluralis. These companies have made astounding progress in decentralized pre-training and post-training. GenSyn has aggregated over 85,000 participants to its Swarm RL testnet. Prime Intellect has trained a 32 billion parameter model with distributed compute. Pluralis recently launched a landmark paper, showing that lossless compression can enable large scale model pre-training on decentralized networks at scale.
We’ve also backed data management platform Perle, which recently published a paper on human-annotation in machine learning, and model fine-tuning platform Bagel recently published its framework for zero knowledge fine-tuning. CoinFund’s portfolio company Giza recently debuted ARMA, an AI agent that’s captivated DeFi, by helping users optimize their stablecoin yield across different protocols.
There’s an abundance of activity at the intersection of Web3 and AI, and as founders take into account how they scale their products in the coming months, the main goal should be onboarding real customers and reaching product-market fit. Companies focusing on inference should be selling inference competitively into Web 2.0 and generating revenue. Early mover AI training networks will aggregate and add more customers.
By the end of 2025, we expect multi-hundred billion parameter models in a decentralized manner. That would have been impossible just 18 months ago! If we continue down this path, open-source AI companies may one day overtake the dominant players in machine learning today. If not, our everyday digital experiences may exist in the hands of just a few corporations on the West Coast of the United States.