Author: YBB Capital Researcher Zeke

I. AI is inevitable, but Crypto x AI?

At the beginning of 2025, the DeepSeek created by Fantom Quantitative brought a "nuclear bomb" to the AI circle. A Chinese AI model that only uses 2,048 NVIDIA H 800 GPUs and costs $5.58 million to train (about one-tenth of Meta's cost) directly matched GPT-4 and Llama 3.1 in benchmark tests such as MMLU and GPQA, and even slightly surpassed these top models from Silicon Valley in complex reasoning and Chinese semantic understanding. The US has been blocking China's chips for years, but on the computing power sandbox of DeepSeek, it has encountered a dramatic deconstruction. Under the siege, Chinese AI has finally developed a technology path that is in line with the national conditions and can match the cutting edge, a set of open-source, low-cost, and homogeneous combinations that directly penetrate the computing power moat of the United States.

Chinese technology products that are always slightly inferior in performance to the US usually give me a stereotypical impression of being cheap and plagiarized, and I believe this is also the intuitive perception of most people about the products of Chinese Internet companies. But this DeepSeek is indeed different. I won't talk about whether the user experience is better than ChatGPT, which is more subjective. Just from the tense reactions from the US political arena and tech giants frequently mentioning this matter, China is no longer playing the role of a technology follower, and the global shock caused by this is also very strong.

Although my wallet was hit first, to some extent this is due to the misjudgment of the traditional AI development. But I still want to talk about my views, that is, the impact of DeepSeek on Crypto.

1. NVIDIA is the biggest victim in this event, one is the questioning of AI computing power demand, and the other is that NVIDIA's unified computing architecture "CUDA" has been bypassed. Friends familiar with the AI field should not be unfamiliar with CUDA, which is one of the important cornerstones driving the development of modern AI. When large model developers use NVIDIA GPUs, they generally do R&D based on CUDA. Using CUDA requires less from developers, because some functions are already encapsulated in CUDA, and there is no need to care about too many details when using them, but it will certainly lose some execution efficiency.

Since CUDA is a general-purpose programming framework, it will cause some flexibility to be lost when training models. DeepSeek's approach is to directly use PTX (an intermediate instruction set framework designed by NVIDIA for GPUs) to bypass the hardware limitations on training speed, which can shorten the training time. Other model training takes 10 days, while DeepSeek can complete it in 5 days. This also means that if DeepSeek intends to adapt to China's domestic GPUs in the future, it will be more at home in hardware adaptation, and NVIDIA's AI chip throne will also be shaken.

In addition to the stock price decline that will have a strong impact on the crypto circle, which is strongly associated with the US stock market, I personally think that in the long run, it is a good thing for decentralized computing power projects. First, more personal GPUs will be able to contribute their surplus computing power in the future, and secondly, if the small and beautiful open-source model path of DeepSeek can succeed, it will force a large number of AI companies to open source, and the demand for local deployment and secondary development computing power will become more and more strong. Looking at the hardware requirements of DeepSeek R 1 from the minimum parameter 1.5 B to the maximum parameter 70 B, from the lowest NVIDIA GeForce GTX 1660 Super to the 40 series, 50 series, and even the professional-level A 100 and H 800 GPUs, they all have the opportunity to contribute surplus computing power again. For the currently relatively transparent decentralized computing power projects with somewhat mediocre computing power supply, this may be an opportunity to turn the tide (of course, the premise is that the latency is low enough).

2. AI framework projects were the hottest emerging track in Crypto before DeepSeek dropped its "nuclear bomb", and it was also the last track I wrote about before the Spring Festival. Now, under the bombardment of DeepSeek, they are almost approaching zero. After all, someone can compete with OpenAI with less than $6 million, while our leading projects with tens of billions of FDV have not yet produced some AI agents that can be called truly practical.

Since the era of Mingwen, we have been pursuing asset-ization in an almost obsessive way. The current crypto circle is already quite open to asset-ization. AI framework projects that are completely off-chain only need an open-source Github code repository and a social media account to issue tokens. The cost of "tokenizing the library" is that they must be able to accept the dimensional attack of traditional AI companies one day.

In the golden age of AI development, the traditional Internet companies will not only have this DeepSeek as their trump card. The development of AI between China and the US will only accelerate. The key is how Crypto should combine with which directions in the upstream and downstream of AI to highlight the advantages of decentralization, and not be killed by a mysterious AOE. Roughly dividing the Crypto x AI technology stack, we can divide it into the computing power layer, data layer, middleware layer, and application layer. In the current layers, I always feel that the necessity of Crypto does not exist. If we consider privacy and security from the perspective of the future, it may be a good angle, after all, the replacement or assistance of human work by AI agents has become a reality today, and how to ensure the privacy of the data processed by AI is a problem that traditional Internet companies cannot solve. Furthermore, if an AI agent has the right to make payments, how to ensure the security of the wallet will also become a problem. Using the blockchain as a compliance audit layer for AI models should be our main development direction in the future.

On the other hand, what should the incentives be directed towards? In addition to being able to catalyze the computing power layer and model sharing, incentives can actually teach AI how to interact with the virtual world. Unlike the natural possession of decades of Internet language data from around the world during LLM training, teaching AI how to act correctly actually requires constant labeling of the correct positions by humans. Just like teaching a visual model to recognize what is an animal and what is a car, this is not something that can be done by outsourcing companies to find a group of college students. Fostering a decentralized network where countless individuals teach AI is also a direction. This has been discussed in more detail in my previous articles. And what else can be done around incentives? Combining with DePin to teach intelligent agents to interact with reality, incentivizing AI to acquire attention, incentivizing AI-related secondary creation (such as the incentive model of Bittensor is a good example), and automatic adjustment of Token incentive mechanisms (this is derived from a question I raised in my previous article: When a decentralized project becomes overwhelmingly large and enters the mainstream, what should be done about deflation and inflation? Should it rely on simple code rules, or listen to the project team of only a few or dozens of people? Or those spiritual leaders? Oh, yes, we also have governance tokens. But governance tokens are meaningless before solving the witch problem. Democratic voting can never be reflected in governance proposals, after all, a16z can veto the approval votes of a large community with just a few wallets, so what's the point of voting?).

We really can't train a DeepSeek-like model in the blockchain like traditional Internet companies by aggregating a batch of top AI talents and renting a large-scale GPU cluster. The meaning of Crypto's existence is to endow another field with irreplaceable decentralized characteristics, just as we have endowed financial freedom in the past. AI is an inevitable narrative for humanity, but what role can Crypto play in it?

3. This is the first time I have mentioned Wordcoin in an article, and even now I find the crypto-utopian project initiated by Sam Altman a bit absurd. Whether to register your iris means you need to make a choice between state surveillance and corporate surveillance, a situation that seems to be like being in the Matrix and having to choose between the red pill and the blue pill.

However, the concept of a universal basic income or inclusive finance may no longer be a joke at this stage. DeepSeek, which can locally deploy models that rival the most advanced large models, has already seen intelligent agents appear in hospitals and government agencies in China. According to a McKinsey report forecast for 2024, up to 50% of jobs could be replaced by AI within the next six years. Future versions of Wordcoin may even be issued by the government, and if this process continues to intensify, related Tokens for inclusive finance may also emerge and be repeatedly hyped, which would coincide with Trump's term, so it is highly possible that this crypto president will issue a similar Token.

4. According to recent statements by Elon Musk, AI may take over 25 years of Nobel Prizes. I believe that using blockchain fundraising (even contributing computing power, storage, methods and other resources) to drive AI research would be more interesting and effective than the current DeSci, and I may call it Decentralized AI Science or DeAIS.

II. Meme Coin is no longer Meme

Previously, when we analyzed Meme Coin, we talked about subculture, community consensus, and the viral effect. But now, sitting in front of GMGN, I analyze conspiracy groups, head addresses, and Dev's rat holes every day. When a CA is posted in various on-chain meme groups, it's time to charge. Today's Meme is more absurd than ever before, and in the current Pump.fun, you can't even find a Token that can let you sleep soundly, as the K-line may just plummet while you're in the bathroom.

The continuous simplification of asset issuance thresholds and the high anonymity of blockchains have fueled this crazy casino culture, allowing unknown teams to turn the crypto market into an ATM. The evolution of Meme is also becoming more and more arbitrary, not only the aforementioned tokenization of memes, but anything, any event, any person, or even any AI can be tokenized. Without a cultural core or consensus cohesion, the so-called leading projects can be completely forgotten in just a few weeks. The celebrity coin craze starting from Trump only lasted a month, and billions of dollars flowed out of SOL after a tweet from President Milei, marking the beginning of the Meme recession. Milei's response was simple: delete the tweet and reply to the crowd: I didn't know.

The rapid development of AI has already taken away too much attention from this world, making it difficult for tech-driven development to make progress. Retail investors who have given up on value investing can only gamble in the pig-slaughtering arena to be among the lucky few. The increasingly scarce liquidity is repeatedly extracted, reflected in the daily red K-lines in CEXs and DEXs, as well as the disdain of traditional capital and outsiders for altcoins.

III. Carving a boat cannot retrieve a lost sword

The cycle law is clearly no longer effective, and all attempts to retrieve a lost sword by carving a boat are futile. BTC going bullish does not mean altcoins will go bullish, but BTC going bearish means altcoins will definitely go bearish. Our perception of altcoins must be renewed - the altcoin market is no longer one that can be supported by a white paper alone. A large project listed on a top CEX must be mature enough to support its own token price.

Looking back at the token growth over the past seven years, in the 2017 market, there were less than 2,000 tokens listed, while in 2024 the number of listed tokens has reached nearly 25,000 (data from CoinGecko, including delisted tokens). The exponential inflation of token numbers is an irreversible evolution of the blockchain's low-entropy value system towards a high-entropy noise system. While in 2017 each token carried the ideal of "disrupting the world", by 2024 tokens have evolved into chips for liquidity exit. The emergence of more tokens has not brought more innovation and implementation, but the high valuations of star projects have exponentially increased the market's demand for liquidity.

As mentioned above, without the recognition of the outside world, retail investors cannot support the valuations of these projects. The listing of most altcoin projects is often the historical high point, and Binance is the last stop. The crypto market needs a revolution, and star projects should live up to their huge fundraising. Bybit's attempt to publicly disclose project financial reports may be a solution. In my personal opinion, the market needs a deep bear market to reshape the valuation system and listing standards of today's altcoin projects.

IV. Bewilderment

I once saw a glimmer of hope on Ton, believing that the beginning of crypto consumer-level applications had arrived, but that brief glimmer faded with the ebb of the Tap to Earn craze. The liquidity mining derived from DeFi five years ago brought the crypto market to its most glorious peak, and five years later, DeFi remains the only successful domain.

The topics of discussion among insiders these days are very simple: have you bought BTC? Have you shorted? Can you give me a CA? Everyone is bewildered, we can't find a right direction anymore, and the current state of the crypto market is that you can't sleep soundly if you buy any token other than BTC. Diamond Hands is not a complimentary term today, and if you don't buy BTC, it's more like a synonym for a fool.

When I open the various chain media on my phone, it's like reading the New York Times and gossip media, reflecting the fact that the hopes of this circle are largely pinned on policies and eyeballs. From a VC's perspective, we should perhaps only invest in tool-type products in the future, with asset issuance platforms becoming a shovel seller and rent collector, just to survive.

Conclusion

This is clearly not the situation we want to see. Although crypto seems to be lost in a fog of direction at the moment, the success of DeepSeek proves that innovation is still the most effective path to break through the predicament. Crypto currently has the best policy environment, attention, capital, and solid infrastructure in history, and there will be many altcoin ETFs in the near future to inject another round of liquidity into the crypto market. We are clearly in the mainstream, but trapped in our own fortress. Perhaps the recession of Meme Coin is hiding a turning point, and the future of humanity may not be just AI.