DeepSeek Faces Server Capacity Challenges Amid Rapid Growth and Cyber Attacks

Why on Earth is the DeepSeek Server Always "Busy"?

While DeepSeek offers tremendous savings over OpenAI, Google, and Anthropic API, it is experiencing severe performance issues. OpenAI sparked the AI revolution with ChatGPT, and DeepSeek emerged as the second household name in the space. In no time, the DeepSeek app topped the iPhone App Store's Top Charts, quickly reaching millions of downloads on Google Play Store. Initially, it was all about how good the model was, how it crashed the market, and how China had outperformed the United States. Now that the hype has cooled down - users are running into a problem. Many users now report that instead of answering queries, the model frequently responds with: "Server is busy, please try again". "I relate to DeepSeek at a spiritual level because I, too, get an exhausted, busy server after answering two questions," jested a user named Soumya on X. While this isn't an issue for users downloading the model locally, it is certainly frustrating for the majority of people using the app. For instance, Stefano Gogioso, a computer science lecturer at the University of Oxford, said he only has two judgements regarding DeepSeek. One, it excels in the chain of thought reasoning, and the other - the server is busy, try again later. Moreover, several industry professionals are also questioning whether DeepSeek has what it takes to run the model in the cloud. Dylan Patel, founder and CEO of semiconductor analysis firm Semi Analysis, said that DeepSeek is likely "bleeding out money". "DeepSeek doesn't have any capacity to actually serve the model," he said. "The ability to use it is like non-existent now. So many people are trying to use it, they don't have the GPUs to serve it [on cloud]," he added. DeepSeek reportedly had over 22 million daily active users on its platform in January, second only to ChatGPT. Are these too many for a relatively new AI player to handle? This isn't just about the chatbot application; even DeepSeek's API is facing a similar fate, if not worse. While it offers tremendous savings over OpenAI, Google, and Anthropic API, it is experiencing severe performance issues. For instance, DeepSeek's status shows that its API experienced a downtime for 10 continuous days. More recently, on February 6, it was reported that DeepSeek temporarily suspended API top-ups for developers. Note that DeepSeek may only have access to a portion of the GPUs, and its parent company, High Flyer, uses the rest for hedge fund activities. This is primarily what the company does, and DeepSeek is said to be a mere side project. Earlier, one of SemiAnalysis's research claims that DeepSeek has access to 50,000 GPUs. However, Patel said that even if one believes their research, DeepSeek is 'nowhere' close to the GPU volumes needed to serve the model at scale. In contrast, Patel said that OpenAI and Microsoft have hundreds of thousands of GPUs between them to serve their high-performance models. An analysis from Omdia, a tech consulting firm, estimated that in 2024, Microsoft bought 485,000 units of NVIDIA's Hopper chips, whereas Meta bought 224,000 chips. Amazon and Google bought 196,00 and 169,00 Hopper chips, respectively. That is the scale at which these companies operate. These outages could also be attributed to the brutal cyber attacks that DeepSeek has faced. As per the South China Morning Post, DeepSeek faced a DDoS (distributed denial of service) attack, with over 230 million requests per second. Earlier, when DeepSeek-R1 was launched, the Chinese lab also briefly stopped new sign-ups for the platform due to a large-scale cyber attack. Thanks to these issues, DeepSeek is likely limiting input from users. According to an individual from the Chinese state media, these attacks were reportedly from the United States. Ultimately, what matters to the user is how it mitigates these issues. After all, everyone loves DeepSeek. Thanks to its open-source nature, alternative platforms to use DeepSeek are available in multiple forms. The easiest way to move away from DeepSeek's native app and still use the model for free is through OpenRouter. This free-to-use platform lets you access DeepSeek R1 and the V3 models through a chat interface. The other alternative is using DeepSeek on Perplexity. The AI search engine platform allows users to choose DeepSeek's R1 model for responses. The free version of Perplexity offers five queries per day with DeepSeek R1. If you're sceptical of DeepSeek's data collection policies, Perplexity hosts the model on servers in the United States. Lastly, the best option for using DeepSeek for free without any server-related issues is to download the model locally. Multiple sources on the web provide step-by-step instructions on how to do so. For example, the following guide from DataCamp helps you download and set up DeepSeek on Ollama, a platform that lets you run AI models locally on your system.

Why on Earth is the DeepSeek Server Always 'Busy'?

While DeepSeek offers tremendous savings over OpenAI, Google, and Anthropic API, it is experiencing severe performance issues. OpenAI sparked the AI revolution with ChatGPT, and DeepSeek emerged as the second household name in the space. In no time, the DeepSeek app topped the iPhone App Store's Top Charts, quickly reaching millions of downloads on Google Play Store. Initially, it was all about how good the model was, how it crashed the market, and how China had outperformed the United States. Now that the hype has cooled down - users are running into a problem. Many users now report that instead of answering queries, the model frequently responds with: "Server is busy, please try again". "I relate to DeepSeek at a spiritual level because I, too, get an exhausted, busy server after answering two questions," jested a user named Soumya on X. While this isn't an issue for users downloading the model locally, it is certainly frustrating for the majority of people using the app. For instance, Stefano Gogioso, a computer science lecturer at the University of Oxford, said he only has two judgements regarding DeepSeek. One, it excels in the chain of thought reasoning, and the other - the server is busy, try again later. Moreover, several industry professionals are also questioning whether DeepSeek has what it takes to run the model in the cloud. Dylan Patel, founder and CEO of semiconductor analysis firm Semi Analysis, said that DeepSeek is likely "bleeding out money". "DeepSeek doesn't have any capacity to actually serve the model," he said. "The ability to use it is like non-existent now. So many people are trying to use it, they don't have the GPUs to serve it [on cloud]," he added. DeepSeek reportedly had over 22 million daily active users on its platform in January, second only to ChatGPT. Are these too many for a relatively new AI player to handle? This isn't just about the chatbot application; even DeepSeek's API is facing a similar fate, if not worse. While it offers tremendous savings over OpenAI, Google, and Anthropic API, it is experiencing severe performance issues. For instance, DeepSeek's status shows that its API experienced a downtime for 10 continuous days. More recently, on February 6, it was reported that DeepSeek temporarily suspended API top-ups for developers. Note that DeepSeek may only have access to a portion of the GPUs, and its parent company, High Flyer, uses the rest for hedge fund activities. This is primarily what the company does, and DeepSeek is said to be a mere side project. Earlier, one of SemiAnalysis's research claims that DeepSeek has access to 50,000 GPUs. However, Patel said that even if one believes their research, DeepSeek is 'nowhere' close to the GPU volumes needed to serve the model at scale. In contrast, Patel said that OpenAI and Microsoft have hundreds of thousands of GPUs between them to serve their high-performance models. An analysis from Omdia, a tech consulting firm, estimated that in 2024, Microsoft bought 485,000 units of NVIDIA's Hopper chips, whereas Meta bought 224,000 chips. Amazon and Google bought 196,00 and 169,00 Hopper chips, respectively. That is the scale at which these companies operate. These outages could also be attributed to the brutal cyber attacks that DeepSeek has faced. As per the South China Morning Post, DeepSeek faced a DDoS (distributed denial of service) attack, with over 230 million requests per second. Earlier, when DeepSeek-R1 was launched, the Chinese lab also briefly stopped new sign-ups for the platform due to a large-scale cyber attack. Thanks to these issues, DeepSeek is likely limiting input from users. According to an individual from the Chinese state media, these attacks were reportedly from the United States. Ultimately, what matters to the user is how it mitigates these issues. After all, everyone loves DeepSeek. Thanks to its open-source nature, alternative platforms to use DeepSeek are available in multiple forms. The easiest way to move away from DeepSeek's native app and still use the model for free is through OpenRouter. This free-to-use platform lets you access DeepSeek R1 and the V3 models through a chat interface. The other alternative is using DeepSeek on Perplexity. The AI search engine platform allows users to choose DeepSeek's R1 model for responses. The free version of Perplexity offers five queries per day with DeepSeek R1. If you're sceptical of DeepSeek's data collection policies, Perplexity hosts the model on servers in the United States. Lastly, the best option for using DeepSeek for free without any server-related issues is to download the model locally. Multiple sources on the web provide step-by-step instructions on how to do so. For example, the following guide from DataCamp helps you download and set up DeepSeek on Ollama, a platform that lets you run AI models locally on your system.