I. AI Is Inevitable, But Crypto x AI?

At the beginning of 2025, DeepSeek, created by the quantitative firm Huanfang, dropped a "nuclear bomb" in the AI community. A Chinese AI model that used only 2,048 NVIDIA H800 GPUs and had a training cost of $5.58 million (about one-tenth of Meta's) achieved results comparable to GPT-4o and Llama 3.1 in benchmark tests like MMLU and GPQA. It even slightly surpassed these top Silicon Valley models in complex reasoning and Chinese semantic understanding. Despite the United States' years-long chip embargo against China, DeepSeek's computational sandbox dramatically deconstructed this blockade. Under siege, Chinese AI has finally developed a technology path that is both suitable for national conditions and competitive with the cutting-edge. An open-source, low-cost, and homogeneous combination has directly breached the U.S.'s computational moat.

II. DeepSeek Innovation and Endogenous Dilemmas: Navigating Crypto in the Fog

Chinese tech products that always seem to lag slightly behind the U.S. in performance have long given me the stereotypical impression of being cheap and derivative. I believe this is also the intuitive perception most people have of Chinese internet companies' products. But DeepSeek is truly different this time. Setting aside the subjective experience of whether it is better than ChatGPT, the nervous reactions from the U.S. political and tech giants show that China is no longer playing the role of a technological follower, and the resulting global shock is quite strong.

Although my wallet was the first to be affected, which was partly due to a misjudgment of traditional AI development, I still want to share my thoughts on DeepSeek's impact on Crypto.

1. NVIDIA was the most severely impacted in this incident. Firstly, the demand for AI computing power was questioned, and secondly, NVIDIA's unified computing architecture "CUDA" was bypassed. Friends familiar with the AI field should be no stranger to CUDA, which is one of the important cornerstones of modern AI development. When large model developers use NVIDIA GPUs, they generally conduct research based on CUDA. Using CUDA has lower requirements for developers because it encapsulates some functions, allowing users to ignore many details. However, this inevitably leads to a loss of execution efficiency.

Since CUDA is a general-purpose programming framework, it can result in some flexibility loss during model training. DeepSeek's approach is to directly use PTX (the intermediate instruction set framework designed by NVIDIA for GPUs) to bypass hardware limitations on training speed, thereby shortening training time. While other models may take 10 days to train, DeepSeek can complete it in just 5 days. This also means that if DeepSeek intends to adapt to domestic Chinese GPUs in the future, it will be more adept at hardware integration, potentially shaking NVIDIA's throne in AI chips. (This paragraph is based on an analysis of DeepSeek's training process by Mirae Asset Securities in South Korea.)

In addition to the significant impact on the cryptocurrency market due to stock price declines, I personally believe that in the long run, this is actually a good thing for decentralized computing projects. First, more individual GPUs will be able to continue contributing their remaining computing power in the future. Second, if DeepSeek's small and beautiful open-source model path succeeds, it will force many AI companies to open-source their models. The demand for computing power for local deployment and secondary development will become increasingly strong. Considering the hardware requirements of DeepSeek R1, with the smallest parameter of 1.5B to the largest parameter of 70B, GPUs ranging from the lowest NVIDIA GeForce GTX 1660 Super to the 40 and 50 series, and even professional A100 and H800 GPUs, will all have the opportunity to contribute their spare computing power again. For decentralized computing projects that currently have a low profile and somewhat marginal computing power supply, this might be a chance to turn things around (of course, provided that latency can be sufficiently low).

1. AI framework projects were the hottest emerging track in Crypto before DeepSeek dropped its "nuclear bomb," and also the last track I wrote about before the Spring Festival. Now, under DeepSeek's bombardment, they are almost all heading towards zero. After all, others can compete with OpenAI at a cost of less than $6 million, while our leading projects with a fully diluted valuation of billions have yet to produce AI Agents of practical value.

Since the inscription era, we have been almost obsessively pursuing assetization. The current cryptocurrency market's tolerance for assetization is already quite open. Completely off-chain AI framework projects only need an open-source GitHub code repository and a social account to issue tokens. The price of "tokenizing a repository" is that one must be prepared for the day when traditional AI companies launch a dimension-reducing strike.

In the golden age of AI development, traditional internet companies will not only have DeepSeek as their big move. The development of AI between China and the U.S. will only accelerate. The key question is which directions Crypto should combine with in the upstream and downstream of AI to highlight the advantages of decentralization and not be wiped out by an unexpected AOE attack. Roughly dividing the technical stack of Crypto x AI, we can categorize it into the computing layer, data layer, middleware layer, and application layer. In these current layers, I always feel that the necessity of Crypto's existence is lacking. If we consider the future, privacy and security might be a good angle. After all, AI Agents replacing or assisting human work has become a reality today. How to ensure the privacy of work-related and personal data processed by AI might be a problem that traditional internet companies cannot solve. Furthermore, if an AI Agent has the authority to make payments, ensuring the security of wallets will also become an issue. Using blockchain as the compliance audit layer for AI models should be our main future development direction.

On the other hand, where should the incentives be directed? In addition to fostering the computing layer and model sharing, incentives can actually teach AI how to interact with the virtual world. Unlike LLM training, which naturally has decades of global internet data, teaching AI to act correctly requires continuous human labeling of correct actions. For example, teaching a visual model to recognize animals and cars is not something that can be done by outsourcing to a group of college students. Creating an AI Agent that can interact with the virtual world requires a complex decentralized network, with countless individuals teaching AI. This is also a direction I have discussed in more detail in my past articles. What else can be done around incentives? Combining with DePin to teach agents how to interact with the real world, incentivizing AI to gain attention, incentivizing secondary creations related to AI (for example, Bittensor's incentive method is a good example), and automating token incentive mechanisms through AI (this point is derived from a question I raised in my previous articles. The original wording is as follows: When a decentralized project becomes incredibly large and enters the mainstream, what should be done about deflation and inflation? Should it rely on simple rule-based code, or follow a project team of a few or dozens of people? Or those key figures? Oh, right, we also have governance tokens. However, governance tokens are meaningless until the Sybil problem is solved. Democratic voting can never be reflected in governance proposals, after all, a16z can veto a large community's affirmative vote with just a few wallets. So what's the point of voting? And so on.)

We indeed cannot aggregate a group of top AI talents like traditional internet companies, purchase or rent a large-scale GPU cluster for training, and recreating DeepSeek within the blockchain is naturally a pipe dream. The significance of Crypto is to endow another field with an irreplaceable decentralized characteristic, just as we once did with financial freedom. AI is an inevitable narrative for humanity, but what role can Crypto play in it?

1. This is the first time I have mentioned Wordcoin in an article. This crypto utopia project initiated by Sam Altman still seems somewhat absurd even now. Whether to record iris data means you have to choose between national surveillance and corporate surveillance. It feels like being in The Matrix, choosing to take the red pill or the blue pill.

However, the concept of universal income or inclusive finance may no longer be a joke at this stage. The ability to locally deploy cutting-edge large models like DeepSeek has already brought agents into Chinese hospitals and government units. According to McKinsey's 2024 forecast report, up to 50% of jobs could be replaced by AI in as soon as six years. Future versions of Wordcoin might even be distributed by governments. If this process continues to accelerate, related tokens of the inclusive finance concept may also emerge in large numbers and be repeatedly hyped. With a five to six-year window, it coincides with Trump's term. So, would this crypto president issue a similar token? I think it's highly possible.

1. Based on Elon Musk's recent statements, AI might sweep the Nobel Prizes for the next 25 years. Therefore, I believe that raising funds through blockchain (or contributing computing power, storage, and other resources) to promote AI research is more interesting and effective than the current DeSci. Perhaps I can call it Decentralized AI Science, or DeAIS.

III. Meme Coin No Longer Memes

Once, when we analyzed Meme Coins, we discussed subcultures, community consensus, and propagation effects. But now, sitting in front of GMGN every day, what I analyze are conspiracy groups, leading addresses, and the Dev's front-running. When a CA is sent to various pump groups, that's the moment to charge. Today's Memes are more outrageous than ever. In the current stage of Pump.fun, forget about finding a Token that you can sleep soundly with; you might just go to the bathroom, and the price chart will plummet. The continuous simplification of asset issuance thresholds and the high anonymity of blockchain have fueled this crazy casino culture. An unknown amateur team can use the crypto market as an ATM. The evolution of Memes has become increasingly arbitrary, not just the aforementioned "tokenizing a repository," but any event, any person, or even any AI can be turned into a coin. Without a cultural core or consensus, so-called leading projects can be completely forgotten in just a few weeks. The celebrity coin trend that started with Trump only lasted a month, and with President Milei's single tweet, hundreds of millions of dollars flowed out of Sol again. Memes began to recede. President Milei's reaction was simple: delete the tweet and reply to everyone, "I didn't know."

The rapid development of AI has already taken too much attention from the world, leaving little room for the progress of technologists. Retail investors who abandon value investing can only gamble on being one of the few lucky ones in a scam. The increasingly scarce liquidity is repeatedly extracted, reflected in reality as the daily red K-lines on Cex and Dex, and the traditional capital and outsiders' disdain for altcoins.

IV. Carving on the Boat Cannot Retrieve the Lost Sword

The cycle law has clearly failed. All the approaches of marking the boat to find the lost sword are ineffective. A bull run in BTC does not necessarily mean a bull run in altcoins, but a bear run in BTC will definitely lead to a bear run in altcoins. Our understanding of altcoins must be renewed. The altcoin market is no longer a market that can be supported by a white paper. A large project listed on a top Cex must be mature enough to support its own coin price.

DeepSeek Innovation and Endogenous Dilemmas: Navigating Crypto in the Fog

Looking back at the token growth over the past seven years, there were fewer than 2,000 listed tokens in the 2017 market, while in 2024, the number of listed tokens approached 25,000 (data from CoinGeko, including delisted tokens). The exponential expansion of tokens is essentially the irreversible evolution of the blockchain's low-entropy value system into a high-entropy noise system. When each token in 2017 still carried the ideal of "changing the world," tokens in 2024 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 recognition from the outside world, retail investors cannot support the valuations of these projects. For most altcoin projects, the moment of listing is often the historical peak, and Binance is the final stop. The crypto market needs a revolution. Star projects should live up to their huge funding rounds. Bybit's trial of project financial statement disclosure might be a solution. But in my humble opinion, the market needs a deep bear run to reshape the valuation system and listing standards of today's altcoin projects.

V. Confusion

I once saw a glimmer of light on Ton, thinking that the beginning of Crypto consumer-level applications had arrived. But this fleeting light died out with the fading trend of Tap to Earn. Five years ago, the liquidity mining derived from DeFi brought the crypto market to an incredibly glorious peak. Five years later, the only successful field we have is still DeFi.

Nowadays, the topics I discuss with people in the circle are very simple: Have you bought BTC? Have you shorted it? Give me a CA. Everyone is confused. We can no longer find a correct direction. The reality in the crypto market is that, apart from BTC, buying any other token will not let you sleep soundly. Diamond Hands is no longer a compliment today. If you don't buy BTC, it is more like a synonym for a fool.

When I open various chain media on my phone, it feels like I am reading The New York Times and gossip media. All these phenomena reflect that the hope of this circle has mostly been pinned on policy and attention. From a VC perspective, we might only invest in tool-based products in the future, becoming sellers of asset issuance platforms and rent collectors to survive.

Conclusion

This is clearly not the status quo we are willing to see. Despite the current Crypto market seemingly being lost in a fog of direction, DeepSeek's success has proven that innovation remains the most effective path to break through seemingly insurmountable challenges. Crypto currently enjoys the best policy environment, attention, funding, and solid infrastructure in history. In the near future, there will be many more altcoin ETFs that can inject liquidity into the crypto market again. We are clearly in the mainstream, yet trapped in our own siege. Perhaps behind the current receding tide of Memes lies a turning point, and the future of humanity may not be dominated solely by AI.

I. AI Is Inevitable, But Crypto x AI?

At the beginning of 2025, DeepSeek, created by the quantitative firm Huanfang, dropped a "nuclear bomb" in the AI community. A Chinese AI model that used only 2,048 NVIDIA H800 GPUs and had a training cost of $5.58 million (about one-tenth of Meta's) achieved results comparable to GPT-4o and Llama 3.1 in benchmark tests like MMLU and GPQA. It even slightly surpassed these top Silicon Valley models in complex reasoning and Chinese semantic understanding. Despite the United States' years-long chip embargo against China, DeepSeek's computational sandbox dramatically deconstructed this blockade. Under siege, Chinese AI has finally developed a technology path that is both suitable for national conditions and competitive with the cutting-edge. An open-source, low-cost, and homogeneous combination has directly breached the U.S.'s computational moat.

II. DeepSeek Innovation and Endogenous Dilemmas: Navigating Crypto in the Fog

Chinese tech products that always seem to lag slightly behind the U.S. in performance have long given me the stereotypical impression of being cheap and derivative. I believe this is also the intuitive perception most people have of Chinese internet companies' products. But DeepSeek is truly different this time. Setting aside the subjective experience of whether it is better than ChatGPT, the nervous reactions from the U.S. political and tech giants show that China is no longer playing the role of a technological follower, and the resulting global shock is quite strong.

Although my wallet was the first to be affected, which was partly due to a misjudgment of traditional AI development, I still want to share my thoughts on DeepSeek's impact on Crypto.

1. NVIDIA was the most severely impacted in this incident. Firstly, the demand for AI computing power was questioned, and secondly, NVIDIA's unified computing architecture "CUDA" was bypassed. Friends familiar with the AI field should be no stranger to CUDA, which is one of the important cornerstones of modern AI development. When large model developers use NVIDIA GPUs, they generally conduct research based on CUDA. Using CUDA has lower requirements for developers because it encapsulates some functions, allowing users to ignore many details. However, this inevitably leads to a loss of execution efficiency.

Since CUDA is a general-purpose programming framework, it can result in some flexibility loss during model training. DeepSeek's approach is to directly use PTX (the intermediate instruction set framework designed by NVIDIA for GPUs) to bypass hardware limitations on training speed, thereby shortening training time. While other models may take 10 days to train, DeepSeek can complete it in just 5 days. This also means that if DeepSeek intends to adapt to domestic Chinese GPUs in the future, it will be more adept at hardware integration, potentially shaking NVIDIA's throne in AI chips. (This paragraph is based on an analysis of DeepSeek's training process by Mirae Asset Securities in South Korea.)

In addition to the significant impact on the cryptocurrency market due to stock price declines, I personally believe that in the long run, this is actually a good thing for decentralized computing projects. First, more individual GPUs will be able to continue contributing their remaining computing power in the future. Second, if DeepSeek's small and beautiful open-source model path succeeds, it will force many AI companies to open-source their models. The demand for computing power for local deployment and secondary development will become increasingly strong. Considering the hardware requirements of DeepSeek R1, with the smallest parameter of 1.5B to the largest parameter of 70B, GPUs ranging from the lowest NVIDIA GeForce GTX 1660 Super to the 40 and 50 series, and even professional A100 and H800 GPUs, will all have the opportunity to contribute their spare computing power again. For decentralized computing projects that currently have a low profile and somewhat marginal computing power supply, this might be a chance to turn things around (of course, provided that latency can be sufficiently low).

1. AI framework projects were the hottest emerging track in Crypto before DeepSeek dropped its "nuclear bomb," and also the last track I wrote about before the Spring Festival. Now, under DeepSeek's bombardment, they are almost all heading towards zero. After all, others can compete with OpenAI at a cost of less than $6 million, while our leading projects with a fully diluted valuation of billions have yet to produce AI Agents of practical value.

Since the inscription era, we have been almost obsessively pursuing assetization. The current cryptocurrency market's tolerance for assetization is already quite open. Completely off-chain AI framework projects only need an open-source GitHub code repository and a social account to issue tokens. The price of "tokenizing a repository" is that one must be prepared for the day when traditional AI companies launch a dimension-reducing strike.

In the golden age of AI development, traditional internet companies will not only have DeepSeek as their big move. The development of AI between China and the U.S. will only accelerate. The key question is which directions Crypto should combine with in the upstream and downstream of AI to highlight the advantages of decentralization and not be wiped out by an unexpected AOE attack. Roughly dividing the technical stack of Crypto x AI, we can categorize it into the computing layer, data layer, middleware layer, and application layer. In these current layers, I always feel that the necessity of Crypto's existence is lacking. If we consider the future, privacy and security might be a good angle. After all, AI Agents replacing or assisting human work has become a reality today. How to ensure the privacy of work-related and personal data processed by AI might be a problem that traditional internet companies cannot solve. Furthermore, if an AI Agent has the authority to make payments, ensuring the security of wallets will also become an issue. Using blockchain as the compliance audit layer for AI models should be our main future development direction.

On the other hand, where should the incentives be directed? In addition to fostering the computing layer and model sharing, incentives can actually teach AI how to interact with the virtual world. Unlike LLM training, which naturally has decades of global internet data, teaching AI to act correctly requires continuous human labeling of correct actions. For example, teaching a visual model to recognize animals and cars is not something that can be done by outsourcing to a group of college students. Creating an AI Agent that can interact with the virtual world requires a complex decentralized network, with countless individuals teaching AI. This is also a direction I have discussed in more detail in my past articles. What else can be done around incentives? Combining with DePin to teach agents how to interact with the real world, incentivizing AI to gain attention, incentivizing secondary creations related to AI (for example, Bittensor's incentive method is a good example), and automating token incentive mechanisms through AI (this point is derived from a question I raised in my previous articles. The original wording is as follows: When a decentralized project becomes incredibly large and enters the mainstream, what should be done about deflation and inflation? Should it rely on simple rule-based code, or follow a project team of a few or dozens of people? Or those key figures? Oh, right, we also have governance tokens. However, governance tokens are meaningless until the Sybil problem is solved. Democratic voting can never be reflected in governance proposals, after all, a16z can veto a large community's affirmative vote with just a few wallets. So what's the point of voting? And so on.)

We indeed cannot aggregate a group of top AI talents like traditional internet companies, purchase or rent a large-scale GPU cluster for training, and recreating DeepSeek within the blockchain is naturally a pipe dream. The significance of Crypto is to endow another field with an irreplaceable decentralized characteristic, just as we once did with financial freedom. AI is an inevitable narrative for humanity, but what role can Crypto play in it?

1. This is the first time I have mentioned Wordcoin in an article. This crypto utopia project initiated by Sam Altman still seems somewhat absurd even now. Whether to record iris data means you have to choose between national surveillance and corporate surveillance. It feels like being in The Matrix, choosing to take the red pill or the blue pill.

However, the concept of universal income or inclusive finance may no longer be a joke at this stage. The ability to locally deploy cutting-edge large models like DeepSeek has already brought agents into Chinese hospitals and government units. According to McKinsey's 2024 forecast report, up to 50% of jobs could be replaced by AI in as soon as six years. Future versions of Wordcoin might even be distributed by governments. If this process continues to accelerate, related tokens of the inclusive finance concept may also emerge in large numbers and be repeatedly hyped. With a five to six-year window, it coincides with Trump's term. So, would this crypto president issue a similar token? I think it's highly possible.

1. Based on Elon Musk's recent statements, AI might sweep the Nobel Prizes for the next 25 years. Therefore, I believe that raising funds through blockchain (or contributing computing power, storage, and other resources) to promote AI research is more interesting and effective than the current DeSci. Perhaps I can call it Decentralized AI Science, or DeAIS.

III. Meme Coin No Longer Memes

Once, when we analyzed Meme Coins, we discussed subcultures, community consensus, and propagation effects. But now, sitting in front of GMGN every day, what I analyze are conspiracy groups, leading addresses, and the Dev's front-running. When a CA is sent to various pump groups, that's the moment to charge. Today's Memes are more outrageous than ever. In the current stage of Pump.fun, forget about finding a Token that you can sleep soundly with; you might just go to the bathroom, and the price chart will plummet. The continuous simplification of asset issuance thresholds and the high anonymity of blockchain have fueled this crazy casino culture. An unknown amateur team can use the crypto market as an ATM. The evolution of Memes has become increasingly arbitrary, not just the aforementioned "tokenizing a repository," but any event, any person, or even any AI can be turned into a coin. Without a cultural core or consensus, so-called leading projects can be completely forgotten in just a few weeks. The celebrity coin trend that started with Trump only lasted a month, and with President Milei's single tweet, hundreds of millions of dollars flowed out of Sol again. Memes began to recede. President Milei's reaction was simple: delete the tweet and reply to everyone, "I didn't know."

The rapid development of AI has already taken too much attention from the world, leaving little room for the progress of technologists. Retail investors who abandon value investing can only gamble on being one of the few lucky ones in a scam. The increasingly scarce liquidity is repeatedly extracted, reflected in reality as the daily red K-lines on Cex and Dex, and the traditional capital and outsiders' disdain for altcoins.

IV. Carving on the Boat Cannot Retrieve the Lost Sword

The cycle law has clearly failed. All the approaches of marking the boat to find the lost sword are ineffective. A bull run in BTC does not necessarily mean a bull run in altcoins, but a bear run in BTC will definitely lead to a bear run in altcoins. Our understanding of altcoins must be renewed. The altcoin market is no longer a market that can be supported by a white paper. A large project listed on a top Cex must be mature enough to support its own coin price.

DeepSeek Innovation and Endogenous Dilemmas: Navigating Crypto in the Fog

Looking back at the token growth over the past seven years, there were fewer than 2,000 listed tokens in the 2017 market, while in 2024, the number of listed tokens approached 25,000 (data from CoinGeko, including delisted tokens). The exponential expansion of tokens is essentially the irreversible evolution of the blockchain's low-entropy value system into a high-entropy noise system. When each token in 2017 still carried the ideal of "changing the world," tokens in 2024 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 recognition from the outside world, retail investors cannot support the valuations of these projects. For most altcoin projects, the moment of listing is often the historical peak, and Binance is the final stop. The crypto market needs a revolution. Star projects should live up to their huge funding rounds. Bybit's trial of project financial statement disclosure might be a solution. But in my humble opinion, the market needs a deep bear run to reshape the valuation system and listing standards of today's altcoin projects.

V. Confusion

I once saw a glimmer of light on Ton, thinking that the beginning of Crypto consumer-level applications had arrived. But this fleeting light died out with the fading trend of Tap to Earn. Five years ago, the liquidity mining derived from DeFi brought the crypto market to an incredibly glorious peak. Five years later, the only successful field we have is still DeFi.

Nowadays, the topics I discuss with people in the circle are very simple: Have you bought BTC? Have you shorted it? Give me a CA. Everyone is confused. We can no longer find a correct direction. The reality in the crypto market is that, apart from BTC, buying any other token will not let you sleep soundly. Diamond Hands is no longer a compliment today. If you don't buy BTC, it is more like a synonym for a fool.

When I open various chain media on my phone, it feels like I am reading The New York Times and gossip media. All these phenomena reflect that the hope of this circle has mostly been pinned on policy and attention. From a VC perspective, we might only invest in tool-based products in the future, becoming sellers of asset issuance platforms and rent collectors to survive.

Conclusion

This is clearly not the status quo we are willing to see. Despite the current Crypto market seemingly being lost in a fog of direction, DeepSeek's success has proven that innovation remains the most effective path to break through seemingly insurmountable challenges. Crypto currently enjoys the best policy environment, attention, funding, and solid infrastructure in history. In the near future, there will be many more altcoin ETFs that can inject liquidity into the crypto market again. We are clearly in the mainstream, yet trapped in our own siege. Perhaps behind the current receding tide of Memes lies a turning point, and the future of humanity may not be dominated solely by AI.