How long will it take for DeepSeek to defeat OpenAI?

Written by
Caleb Hayes
Updated on:July-10th-2025
Recommendation

Can the domestic AI big model DeepSeek surpass OpenAI? Brother K deeply analyzes the current AI competition landscape.

Core content:
1. Performance comparison between DeepSeek and the top American AI big model
2. The contribution of DeepSeek's technological innovation to narrowing the AI ​​gap between China and the United States
3. The development trend of AI technology and the challenges faced by DeepSeek

Yang Fangxian
Founder of 53AI/Most Valuable Expert of Tencent Cloud (TVP)

Recently, Brother K attended several industry conferences in the field of AI and found that many people generally have an optimistic mood of "the boat has passed through thousands of mountains". I understand that this atmosphere is largely due to the outstanding performance of the domestic large model DeepSeek.


It is undeniable that DeepSeek has reversed the situation where China's AI lags behind the United States by 2 to 3 years with its technological innovation, but has it really reached the stage of "China's AI surpassing the United States in all aspects"? Has our DeepSeek really completely defeated OpenAI? Brother K will share his observations and thoughts.


01

Know yourself and know your enemy: How far is DeepSeek from OpenAI?


1. There is still a gap with the first-tier large models in the United States


From a comprehensive evaluation perspective, DeepSeek has outperformed Meta's Llama series in some tasks. For example, in multiple benchmarks of natural language processing, DeepSeek's open source model has demonstrated impressive performance. However, if it is compared with OpenAI's GPT-4, Anthropic's Claude 3, and Google's Gemini 2.0 Flash, the gap between them is still obvious.


Taking Gemini 2.0 Flash as an example, this model not only performs well in reasoning tasks, but also keeps costs lower. According to data released at the 2024 Google I/O conference, the response speed of Gemini 2.0 Flash has doubled compared to before.


The training cost is also significantly lower than the previous generation, which makes it the "most cost-effective". In addition to the significant improvement in overall cost performance, Gemini 2.0 Flash supports full-modal data processing capabilities with deep integration of video, audio, and text, which makes DeepSeek, which is currently mainly limited to the text field and has not yet made efforts in multi-modal scenarios, slightly inferior.


2. Shortened the technological generation gap with OpenAI


But having said that, DeepSeek's technological innovation and breakthroughs are indeed something we should be proud of. In the past, OpenAI's GPT series almost led the world with a generational crushing attitude, especially the technological leap from GPT-3 to GPT-4, which left other players far behind. But DeepSeek came from behind and achieved the reasoning ability of the benchmark GPT-4o1 at a lower cost with the technical route of "algorithm optimization instead of computing power stacking".


More importantly, DeepSeek has chosen the open source direction. This strategy not only allows DeepSeek's own technology to iterate quickly, but also better promotes the development of the entire AI technology ecosystem.


3. It won’t be another OpenAI that defeats OpenAI


DeepSeek has indeed made many technological innovations, such as using a hybrid expert model (MoE) architecture to split a large model into multiple "experts", with division of labor and collaboration during training and on-demand calls during inference, thereby improving efficiency; it also introduced a multi-head potential attention mechanism (MLA) to optimize the efficiency of long text processing and reduce video memory usage, etc. But another fact that cannot be ignored is that these technologies still rely on the Transformer architecture, and are innovated and optimized on this basis. This approach of exchanging training efficiency for manual optimization is essentially still a technology-following route.


However, with the development of AI technology, the limitations of the Transformer architecture have gradually become apparent. The industry generally believes that when the parameter scale of the model reaches a certain level, the marginal benefits of performance improvement may gradually decrease. For DeepSeek, the space for improving performance through technical optimization based on the existing architecture has become smaller and smaller. In other words, on this technical route, the possibility of defeating OpenAI with OpenAI's "tricks" is extremely low .


Historical experience shows that disruptive innovation often comes from breaking through the existing technological paradigm . Just like Apple, which completely changed the landscape of the mobile phone industry by redefining the way mobile phones interact, the AI ​​field also needs such disruptive innovation, rather than incremental improvements under the existing architecture. AI master Yann LeCun has also pointed out that future breakthroughs in large models may come from a completely new computing paradigm. If DeepSeek or other large models want to truly surpass OpenAI fundamentally, they may have to find a new way in terms of technology and work hard.


In fact, Chinese AI companies have begun to try different technical routes. For example, MiniMax, one of the "Six Little Tigers" of domestic large models, proposed a "linear attention" mechanism, which has performed particularly well in the field of long texts and is worth looking forward to. MiniMax is also taking the open source route.




02

DeepSeek rewrites the course of AI history


1. Becoming the “Open Source Leader” in the AI ​​Industry


At present, DeepSeek is the undisputed leader in the field of AI open source. OpenAI has turned to closed source since GPT-3. Although Meta and Mistral insist on open source, they have always been unable to compete with the GPT series. It was not until the emergence of DeepSeek that the strength of the "open source faction" was truly strengthened.


DeepSeek's open source models, such as DeepSeek-V3 and DeepSeek-R1, have been widely welcomed by developers on the Hugging Face platform. DeepSeek-R1 quickly received 10,000 likes and became the "most popular large model among nearly 1.5 million models" on the platform. The same is true on GitHub. Within less than three months of its launch, DeepSeek's open source project attracted more than 100,000 forks and stars, becoming one of the most popular AI open source projects during the same period.


Through feedback from the open source community, DeepSeek has formed a closed loop of rapid iteration, which has also enabled the "long tail effect" to be demonstrated, attracting a large number of developers. These developers may each develop applications with a small audience, but when they come together they have formed a rich and complete application ecosystem.


2. Reduced the cost of large model training


DeepSeek's FlashMLA technology improves the H800 GPU's inference efficiency by 30% and increases the memory bandwidth to 3000GB/s. The training cost is about 40% lower than that of the traditional Transformer model. What's more commendable is that DeepSeek has open-sourced its technology and innovation, which is a great boon for small and medium-sized enterprises with limited resources, allowing them to call high-performance models at a very low cost, allowing more companies to participate in the training and application of large models.


It is no exaggeration to say that DeepSeek, with its innovations in engineering and algorithms, has triggered a "universal revolution" of low-cost training of large models, which has greatly benefited the entire AI industry. For this reason alone, DeepSeek is absolutely worthy of the praise and love of users around the world.


3. Upgraded experience of top-notch Chatbot


Another eye-catching feature of DeepSeek is its significant breakthrough in C-end user experience. The Chatbot it launched not only supports networking functions, but also makes the reasoning process transparent through the "Chain of Thought" (CoT), allowing users to intuitively see how the model gets the answer step by step. This transparent experience not only enhances users' trust in AI, but also conforms to the "user-centric" design concept, and improves user satisfaction and loyalty by satisfying users' curiosity and desire for control.


DeepSeek's reasoning visualization user experience, amazing generation effects, and free use strategy are a perfect combination that instantly attracted a large number of C-end users. Without any marketing promotion, it created the AI ​​application myth of "over 100 million users in 7 days after launch" worldwide.




03

DeepSeek reshapes the AI ​​industry landscape in China and the United States


1. Shake the brand mind of ChatGPT, the world's number one ChatBot


ChatGPT has long been regarded as a synonym for AI technology, firmly occupying the "brand mind" of the "world's first" Chatbot and being regarded as irreplaceable. However, its high cost of use (the basic version is priced at US$20 per month) has deterred many users. The launch of DeepSeek's free model has dealt a heavy blow to ChatGPT, and its zero-cost strategy has quickly attracted a large number of ChatGPT users to migrate.


According to Bloomberg data, DeepSeek has been the most downloaded app in 140 markets around the world, while ChatGPT has seen a significant decline in usage. OpenAI founder Sam Altman publicly admitted that he was under "a lot of pressure" and even wanted to "meet with DeepSeek's managers." DeepSeek has proven with facts and data that ChatGPT is not necessarily the "number one" in the world of ChatBots, nor will it be the "only" choice for users.


2. Impact on closed-source large model companies


Kodak missed the digital era because it stuck to film; Nokia was disrupted by the market because it rejected smartphones; these have become classic cases in business history of failure due to failure to adapt to trends and unwillingness to change. Today, many closed-source large model companies are also experiencing the difficult choice of "change or not change".


Faced with the success of DeepSeek's open source, some closed-source large model companies have to re-examine their strategies. The impact of the open source model on them is not only reflected in the technical level, but also in the market level. Technically, the mystery of closed-source technology has been broken, and even the user experience has been surpassed by the open-source DeepSeek. In this way, users' acceptance and trust in open source technology will only increase. Faced with this embarrassing situation, how should closed-source companies that are still "half-covered" deal with it? Don't forget that users will vote with their feet.


Looking at the market level, in the recent earnings season, many closed-source AI companies mentioned the impact of the open source trend on their business and said they would accelerate the open source process. Some companies that acted quickly have already taken the lead. For example, OpenAI has announced that some large models will be open sourced, Anthropic has also stated that it will open source some core modules, and Google has also followed up to accelerate the Gemini open source plan.


The "group stress response"-style self-rescue of closed-source big model companies is not out of conscience, but because of the pressure of closed-source models. To a certain extent, DeepSeek has single-handedly tilted the balance of open source and closed-source games towards the latter. From this perspective, it is not an exaggeration to say that DeepSeek has reshaped the AI ​​industry landscape.




04

The direction is the sea of ​​stars


1. DeepSeek aims at AGI


Whether it is for people or companies, if you want to achieve great things, you must dare to try and make mistakes and have confidence. For DeepSeek, its biggest confidence is its parent company, Magic Quant. DeepSeek is backed by Magic Quant, which is rich and powerful, with sufficient capital reserves, and does not need to be affected by small problems such as "daily necessities" or short-term interests. They have enough strategic determination to focus on the basic research of artificial intelligence and aim at the more ambitious field of AGI.


DeepSeek announced in February this year that it had established an AGI exploration team and planned to open source five code bases to promote AGI research. These all demonstrate their determination and strategic ambition in the field of AGI. If you keep thinking about it, there will be a response. Moreover, they have strong financial resources and a long-term vision. What they need most may be just time.


2. Innovation and open source dual engine drive


If we "upgrade" the situation, DeepSeek's future goal may not be a short-term competition with AI giants such as OpenAI, but also a geopolitical game from a higher perspective. Today, computing power is national strength, and leading AI technology means comprehensive leadership in the entire field of science and technology, production and manufacturing in the future. DeepSeek insists on the dual-engine drive of "innovation + open source" in order to maintain its leading position in this key field related to "national destiny" and to secure a good position.


Facts have also proved that DeepSeek's strategy is correct. This model not only helps to break the US's AI hegemony and technological monopoly, but also injects new vitality into the global AI ecosystem, and ultimately helps China's AI leap from a "rule acceptor" to a "standard co-governor", which is of far-reaching significance and great impact.


Having written this, Brother K wants to express his respect for DeepSeek and also say that in this era, it is difficult for any field to have an "ivory tower". Only garage culture and community co-creation spirit can inspire people. Only in this way can the life of enterprises or individuals have more resilience and breakthroughs. The rise of DeepSeek is not the end, but just the prologue. Let us look forward to more "DeepSeek" stories that belong to us.