DeepSeek bursts the AI bubble, is it a blessing or a curse for Crypto AI?

Author: BUBBLE, BlockBeats

In January 2025, the launch of DeepSeek R1 caused a stir in the AI community, fundamentally changing the Crypto AI ecosystem. During the past cycle, Crypto AI primarily revolved around AI Agents, but DeepSeek R1 and its open-source strategies have completely changed the game: with extremely low training costs and groundbreaking adaptive training methods, the vision of a decentralized AI industry is no longer a mere fantasy but an attainable reality. This transformation has far-reaching implications, as the total market capitalization of Crypto AI has significantly shrunk, with many AI tokens experiencing a 70% pullback. But is this truly a crisis? Or does it signify a complete reshuffling of Crypto AI? Is DeepSeek the "terminator" that shatters the Crypto AI narrative, or the "disruptor" that accelerates its entry into a practical era?

The Wild Growth of DeepSeek

The development of DeepSeek can be traced back to 2021. At that time, the quantitative trading hedge fund, Huanfang, began to recruit AI talent on a large scale. It was uncommon for quantitative firms to pivot to AI, and most of the recruits were AI researchers exploring cutting-edge directions, including large models (LLM) and text-to-image models. Although there were rumors that Huanfang was transforming to better utilize the idle GPU resources within the company, the primary reason was likely to seize the high ground in frontier AI technologies like large models.

By the end of 2022, Huanfang had attracted an increasing number of top AI talents, mainly from Tsinghua University and Peking University. Stimulated by ChatGPT, Huanfang's CEO, Liang Wenfeng, was determined to enter the field of general artificial intelligence and established DeepSeek in early 2023. However, with the rapid rise of AI companies like Zhipu, Moonlight, and Baichuan Intelligence, DeepSeek, as a pure research institution lacking a star founder, faced significant difficulties in independent financing. Therefore, Huanfang chose to spin off DeepSeek and fully fund its development. Although this decision carried high risks, DeepSeek was free from the profit commitments or valuation pressures of its financiers. At the same time, it had relatively ample GPU resource reserves, allowing the team to focus on technological breakthroughs. A group of innovative young people could thrive in this fertile ground, making DeepSeek feel more like a research institute than a company.

Just like early OpenAI, no one would have imagined that a company researching robotic hands to solve Rubik's cubes would eventually develop ChatGPT, nor could anyone foresee how Huanfang, a quantitative firm, would use DeepSeek to penetrate the current AI bubble. The former took seven years, while the latter took only two. In November 2023, DeepSeek launched its LLM with 67 billion parameters, achieving performance close to GPT-4. In May 2024, DeepSeek-V2 went live, and in December of the same year, DeepSeek-V3 was released, performing on par with GPT-4o and Claude 3.5 Sonnet in benchmark tests. The rapid technological leap of DeepSeek is not due to the company's financial resources or high academic qualifications, but rather a result of a technological singularity after "ChatGPT impacts the global AI industry," where various singularities accelerate in any fertile ground that satisfies imagination, until the next critical singularity appears.

Finally, in January 2025, DeepSeek accelerated through the singularity, using their first-generation reasoning-capable large model, DeepSeek-R1, to open that door with training costs far lower than ChatGPT-O1 and outstanding performance.

Distributing the Key to the Interstellar Gate to the World with Open Source

Just one day after the release of DeepSeek R1 and the announcement of the open-source model, U.S. President Trump officially announced at a White House press conference the commencement of a $500 billion mega-investment "Stargate" plan. A joint venture named Stargate was established by OpenAI, SoftBank, Oracle, and investment firm MGX to build new AI infrastructure for OpenAI in the United States.

This level of investment is comparable to the "Manhattan Project," aiming to leverage national power to push closed-source AI to its peak through algorithm stacking, monopolizing the AI market to ensure the leading position of the U.S. domestic AI industry. However, at the time of the plan's announcement, it was unlikely they anticipated that just days later, this open-source model across the ocean would not only break down the door but also bring hammers to smash the walls, while simultaneously distributing hammers to others.

As an open-source model that can rival top closed-source models, DeepSeek's new training architecture has triggered a chain reaction, making it difficult for closed-source AI to compete. Closed-source models that cannot keep up with DeepSeek R1 will be directly eliminated by the capital market. Even Marc Andreessen, the founder of A16z, "the investor of OpenAI," has publicly stated the need to focus more on open-source AI rather than emphasizing closed-source AI. In the industry, whether supporting the potential emergence of AGI or viewing AI merely as an upgraded version of the SaaS industry, there is a consensus that the harms of closed-source far outweigh those of open-source, whether it be black-box issues, industrial monopolies, information security, or capital manipulation of attention—each is a dangerously misguided development direction.

Despite some industry insiders questioning the massive dataset required for V3's mixture of experts technology "MoE," suspecting it was distilled from OpenAI's model, and concerns that the reinforcement learning "RL" methods in R1 require substantial hardware resources, leading to doubts about the number of training chips used, these do not diminish the structural reforms it brings to the industry.

The open-sourcing of DeepSeek R1 breaks the commercial logic of OpenAI's closed-source large models in its training architecture, using a self-evolving logic for models to avoid the traditional paradigm's heavy investment in computing power and data labeling. Although training models still feels like opening a blind box, the cost of that blind box has been significantly reduced.

At the AI hardware level, DeepSeek's V3 open-source directly challenges NVIDIA's market dominance. The moat of NVIDIA GPUs largely lies in its underlying parallel computing platform and programming model, CUDA. Its extensive ecosystem and sufficient number of developers make the learning costs of using non-NVIDIA chips for training prohibitively high, while high barriers to purchase and political restrictions have led to a fragmented global AI development.

For us, in the short term, the U.S. stock market's AI sector has significantly shrunk, and the total market capitalization of Crypto AI has nearly been slashed, with the market entering a bear phase. However, in the long term, the most recognized AI industry is moving towards an open-source, transparent, and decentralized development path. From any perspective, the combination of Crypto and AI will become more harmonious.

The Redemption of Crypto AI, Forward! Forward! Relentlessly Forward

During this round of the Crypto AI bubble burst, many AI concept tokens have experienced a 70% pullback, and the Crypto AI market has significantly shrunk. Some jokingly remarked, "With $5.5 million, you can train a large model; why buy any Crypto AI with these AI market caps?" Indeed, Crypto is a market dominated by capital rather than products, and 90% of AI tokens lack practical significance.

However, as the regulatory framework for the crypto market improves, the crypto market remains the most suitable soil for small and medium-sized AI companies to start their ventures. The cost of large models brought by DeepSeek, at 1/100 of ChatGPT O1, along with its model training methods, will lead to an ecological growth that is thousands of times greater than the current market.

In direct terms, what DeepSeek brings to crypto is a decentralized training model, making Depin-type projects more rational, allowing the training process and information feeding to be more transparent, and establishing a more reasonable value reward mechanism for dataset contributors, making it easier for both sides of model training to settle. The surrounding ecological development of the AI industry, which is thousands of times larger, will enhance the richness of the downstream industry of Crypto AI. When enough competitive and creative product narratives appear in the market, as long as one truly breaks through, external capital will naturally flow back into Crypto. The market has long suffered from PVP, and a series of celebrity coins following TrumpCoin have disrupted the originally abundant liquidity and positive feedback balance in the AI market. Therefore, the bubble burst by DeepSeek is actually a greater boon.

Currently, many Crypto AI projects are either quickly integrating DeepSeek or updating their architecture based on it, including ElizaOS, Argo, Myshell, Build, Hyperbolic, Nillion Network, infraX, and more. Some of these projects have directly optimized their products through DeepSeek.

Myshell

In the production flow of chatbots and application plugins, Myshell has integrated V3 and R1, and even the image generation model Janus-Pro. The technical team at Myshell completed the model integration in just half a day. As one of the few projects in the blockchain space that has consistently focused on refining its product and has even gained recognition in Web2AI products while hesitating to issue tokens, the open-source nature of DeepSeek will bring good news to Myshell's users by significantly lowering costs, allowing for more Agent developers to join the already well-developed Myshell.

Argo

Sam Gao, the developer of Argo, integrated DeepSeek into Argo's key functionalities during the initial product design phase. As a workflow system, Argo has embedded LLM as the standard DeepSeek R1 and has delegated the generation of original workflows to DeepSeek R1. Due to the nature of the workflow, the token consumption and context information will be substantial, averaging ">=10k Tokens." Additionally, Argo has incorporated CoT "Chain-of-Thought" into the workflow thinking process. After the open-sourcing of DeepSeek, not only has the cost of workflow products decreased, but Argo can also deploy LLM locally, ensuring user privacy and security.

Before the launch of DeepSeek R1, Argo had already integrated its preliminary training logic, Chain-of-Thought "CoT," into the production process of Argo's Agent Workflow. Particularly for tasks like meme trading and market trend analysis, Argo has customized its workflow using Graph-of-Thought (GoT), a novel approach that constructs reasoning as a graph, where nodes represent "LLM thoughts" and edges indicate the dependencies between these thoughts.

Given that Argo has chosen GoT (the only Crypto AI Workflow currently using this model), it has achieved a more reliable and transparent process. This innovative approach directly impacts the security and trustworthiness of transactions on the Argo platform. Integrating the thought graph (GoT) into Web3 AI agents places Argo at the forefront of AI crypto trading. The structured reasoning of CoT not only enhances the security of financial transactions but also ensures transparent and reliable decision-making, which is crucial in decentralized finance (DeFi).

Notably, Argo's core developer Sam collaborated with Shaw to write a paper titled "EraseAnything: Enabling Concept Erasure in Rectified Flow Transformers," which discusses how to remove unwanted concepts from large-scale text-to-image diffusion models without compromising the overall generative performance of the model, with assistance from DeepSeek researcher Xingchao Liu.

Hyperbolic

Hyperbolic Labs has also taken the lead in announcing the hosting of the DeepSeek-R1 model on its GPU platform, allowing users to rent Hyperbolic GPU resources to run the DeepSeek-R1 model locally or in designated data centers without sending sensitive data to DeepSeek's servers. This method ensures data privacy while leveraging the excellent reasoning performance of the DeepSeek model. Through Hyperbolic's decentralized computing network, users can access the efficient reasoning capabilities of the DeepSeek model at a lower cost, making it a highly competitive solution for startups, super individual entrepreneurs, or even simply efficient AI users.

This round of the bubble burst has indeed dealt a heavy blow to the Crypto AI market, with many AI tokens losing their speculative value. However, fundamentally, DeepSeek is not eliminating Crypto AI but rather forcing the market to evolve rapidly. After DeepSeek R1, the future of Crypto AI will no longer solely rely on speculation but will need to be reconstructed around decentralized AI computing, economic incentive mechanisms for model training, fair distribution of AI resources, and practical products. The real challenge is whether Crypto can leverage the technological revolution brought by DeepSeek to establish a genuinely valuable AI ecosystem, rather than merely creating concepts and speculation.

This is not the end, but an evolution. Crypto AI needs to move forward faster and more aggressively. / Accelerate