Author: YBB Capital Researcher Zeke
I. AI is Inevitable, But Crypto x AI?
As 2025 begins, the AI community was shaken by a “nuclear bomb” dropped by DeepSeek, developed by Fantom Quant. A Chinese AI model, trained using only 2048 NVIDIA H800 GPUs at a cost of $5.58 million (about one-tenth of Meta’s cost), directly rivaled the likes of GPT-4 and Llama 3.1 in benchmark tests such as MMLU and GPQA. It even slightly outperformed these Silicon Valley giants in areas like complex reasoning and Chinese semantic understanding. Despite the years of chip blockades imposed by the US on China, DeepSeek’s computational prowess dramatically deconstructed the US’s monopoly, as China developed a technology path suited to its own national context but still competitive with top-tier global models. This open-source, low-cost, and homogeneous approach penetrated the US’s computational defense with striking success.
For years, Chinese technology products, which were often seen as a step behind in performance, have been stereotyped as cheap and derivative. I believe this was the intuitive perception that many had of Chinese internet companies. However, DeepSeek is different. Without getting into the subjective experience of whether it surpasses ChatGPT, the fact that this event has been frequently mentioned in US political and tech circles speaks volumes. It’s clear that China has now transitioned from a technology follower to a challenger, and the global ripple effect is profound.
Although the first impact is on my wallet, which can be attributed to a misjudgment of traditional AI development, I still want to share my perspective, particularly on the impact of DeepSeek on the crypto industry.
1.NVIDIA is the biggest loser in this event. First, the demand for AI computational power has been questioned, and second, NVIDIA’s unified hardware and software computing architecture, CUDA, was bypassed. If you’re familiar with the AI field, you would know that CUDA is a key cornerstone in driving modern AI development. When large model developers use NVIDIA GPUs, they generally rely on CUDA for their work. Using CUDA lowers the technical requirements for developers since it has pre-packaged functions, meaning developers don’t have to worry about too many details, but this comes at the cost of execution efficiency.
Since CUDA is a general-purpose programming framework, it limits flexibility during model training. DeepSeek’s solution was to directly use PTX (NVIDIA’s intermediate instruction set framework designed for GPUs) to bypass the hardware limitations on training speed, thus shortening the training duration. While other models require 10 days for training, DeepSeek can complete it in 5 days. This also means that if DeepSeek plans to adapt to China’s domestic GPUs in the future, the hardware adaptation process will be more seamless, potentially shaking NVIDIA’s dominance in the AI chip market. (This paragraph is sourced from a report by Korea’s Future Asset Securities on DeepSeek’s training process.)
Aside from the potential drop in NVIDIA’s stock price, which would severely impact the crypto market, which is closely tied to US stocks, I personally believe that in the long run, this will actually benefit decentralized computing projects. First, more personal GPUs will be able to contribute their spare computing power. Second, if DeepSeek’s compact, open-source model approach succeeds, it will force many AI companies to open-source their models, leading to increased demand for local deployment and secondary development computing power. Looking at DeepSeek R1’s hardware requirements, ranging from a minimum of 1.5B parameters to a maximum of 70B, GPUs from the NVIDIA GeForce GTX 1660 Super to the 40 and 50 series, and even professional-level A100 and H800 GPUs, will all have opportunities to contribute surplus computing power. For decentralized computing projects that are currently somewhat underutilized, this could be a potential breakthrough — provided the latency is low enough.
2. AI Framework Projects: The Hot Crypto Track Before DeepSeek’s “Nuclear Bomb”
Before DeepSeek dropped its “nuclear bomb,” AI framework projects were the hottest emerging track in the crypto space, and they were the last topic I covered before the Chinese New Year. However, in the wake of DeepSeek’s breakthrough, most of these projects are now rapidly trending toward zero. After all, DeepSeek achieved parity with OpenAI for less than $6 million in costs, while our leading projects, with FDV in the billions, have yet to produce anything that can truly be considered a practical AI Agent.
Since the inception of blockchain, there has been an almost obsessive pursuit of assetization. Currently, the crypto space has become highly tolerant of assetization. For AI framework projects that are not even on-chain, they only need an open-source GitHub repository and a social account to issue tokens. This “library-based coin” approach carries the inevitable risk of being wiped out by a “two-dimensional foil” attack from traditional AI companies one day.
In the golden age of AI development, traditional internet companies are unlikely to stop with DeepSeek as their sole weapon. AI development in the US-China rivalry will only accelerate, and the key question is how Crypto can combine with the upstream and downstream of AI to highlight the decentralized advantages without being struck down by an unexpected AOE. Broadly, we can categorize the Crypto x AI tech stack into four layers: computing power, data, middleware, and application. In the current layered structure, I fail to see the necessity of Crypto’s involvement. However, from a future perspective, privacy and security could be strong angles, as AI agents have already become a reality in replacing or assisting human jobs. Ensuring the privacy of AI-handled work and personal data could be a challenge traditional internet companies cannot solve. Further, if an AI agent has payment permissions, ensuring the security of wallets will become an issue. Using blockchain as the compliance and auditing layer for AI models could be the key direction for future development.
On the other hand, incentives also play a crucial role. Aside from stimulating computing power and model sharing, incentives could teach AI how to interact with the virtual world. Unlike LLMs, which have decades of global internet data at their disposal, teaching AI the right actions requires continuous human labeling. For instance, teaching a vision model to recognize animals versus cars isn’t something that can be outsourced to a group of college students. To create an AI agent that can interact with the virtual world, a large decentralized network of individuals will be needed to teach the AI. This is one potential direction. In my past articles, I discussed this in greater detail. What else can incentives drive? Combining with DePin to teach AI agents to interact with the physical world, incentivizing AI to gain attention, incentivizing second creation of AI (Bittensor’s incentivization model is a great example), or having token incentive mechanisms automatically adjusted by AI — these are all fascinating possibilities. This leads to another question from my previous article: When a decentralized project grows massive and enters the mainstream, how should deflation and inflation be managed? Should it rely on simple code rules, a few people in the project team, or those key individuals? And, of course, we have governance tokens. However, governance tokens are meaningless before solving the “witch problem.” Democratic voting will never be reflected in governance proposals, as a16z can veto a large community’s votes with just a few wallets, so what’s the point of voting?
We certainly cannot gather a group of high-end AI talents like traditional internet companies, nor can we purchase or rent massive GPU clusters to train. Trying to replicate DeepSeek within blockchain is a pipe dream. The role of Crypto is to bring an irreplaceable decentralized characteristic to another field, just as we once brought financial freedom to the world. AI is the inevitable narrative of humanity, but the crucial question is: What role can Crypto play in this?
3. Wordcoin: A First Discussion on a Cryptographic Utopia Project
This is the first time I’ve mentioned Wordcoin in my articles. This cryptographic utopia project, initiated by Sam Altman, seems absurd to me even now. The decision to record one’s iris feels like choosing between state surveillance and corporate surveillance, something akin to the red pill and blue pill choice in The Matrix.
However, the idea of universal basic income, or financial inclusivity, no longer seems like a joke at this stage. DeepSeek’s AI agents, capable of local deployment to rival top-tier models, have already begun appearing in Chinese hospitals and government institutions. According to McKinsey’s 2024 forecast, up to 50% of jobs could be replaced by AI within the next six years. The future version of Wordcoin could even be distributed by the government. If this trend intensifies, related tokens for universal finance might emerge and be repeatedly hyped. Given a five- or six-year window, this could align with Trump’s presidency. Would a crypto president issue such a token? I think it’s highly likely.
4. Elon Musk and the Future of AI Research Funding
In light of Elon Musk’s recent statements, AI might dominate the Nobel Prize for the next 25 years. I believe that blockchain-based fundraising (even contributing computing power, storage, methods, and other resources) to promote AI research will be more interesting and effective than the current decentralized science (DeSci) movement. Perhaps I can call this Decentralized AI Science, or DeAIS.
