DeepSeek’s Disruptive Entrance into the GenAI Market

Our recent podcast discussed the impact of DeepSeek, a startup that has quickly made waves in the AI industry with its innovative models. With just over two hundred employees, DeepSeek has introduced models that are not only high-performing but also significantly more cost-effective than many existing alternatives. Let’s dive into the details of their breakthrough and what it means for the future of AI.

DeepSeek’s Groundbreaking Models

Over the last two months, DeepSeek has released two significant models:

• V3: This chat model is notable for its significantly lower training costs. DeepSeek reported it cost around $5 million to train, which is substantially less than what other AI labs have reported ($80m-100m for OpenAI). While this figure excludes auxiliary costs like hardware, data labeling, and human capital, the training is still likely much cheaper than other models.
• R1: This is an open-weights reasoning model that was fine-tuned from V3 using mainly reinforcement learning, similar to OpenAI’s models, and it stands out as a cost and performance-competitive open-source reasoning model. 

 

Lower Costs, Higher Competition

The reduction in cost of training and running an AI model of this performance that is open-weights is a key factor in DeepSeek’s disruption. One million output tokens using DeepSeek R1 on Together AI costs ~$6 (and only $0.55 on Deepseek’s own API currently), compared to ~$60 for OpenAI’s o1, demonstrating a significant cost reduction for inference. This cost efficiency has major implications for businesses looking to implement AI:

• Reduced operational costs: Lower inference costs make it more affordable for companies to embed LLMs into their products, leading to a greater return on investment.
• Increased adoption: Cheaper AI models could accelerate the adoption of AI technologies across various industries.
• Stimulated competition: DeepSeek’s innovations are pushing down prices across the industry, with other players like Google adjusting their prices to stay competitive.

At this point it is worth noting that the trend in cost reduction has since been continued by Google’s release of the top-performing Gemini 2.0 Flash with a cost of $0.4 per one million output tokens.

Reasoning Models vs. Chatbot Models

There are also some key differences between reasoning and chatbot models.

• Chat models work by predicting the next word in a sequence and are good at tasks like grammar, syntax, world knowledge, and translation. However, they can struggle with complex problems that require multi-step solutions.
• Reasoning models are based on chat models, but fine-tuned to produce and evaluate intermediate steps of reasoning to reach a conclusion. They use a reinforcement learning approach to favor a good path or intermediate step.

Reasoning models, such as DeepSeek’s R1 are particularly useful when dealing with complex, multi-step approach and planning problems. They are actively used in coding and research agents. These models will thrive in industries where information is vast and complex such as law, academia, finance and medicine.

Open Source Considerations

DeepSeek’s R1 model is an open-weights model. This means the model architecture, training code, and pre-trained weights are publicly available. Lately, there has been a rise in the adoption of open-weight language models by enterprises. However, there are things to consider:

• Hosting costs: While the model weights are free, hosting it yourself can be expensive, potentially costing more than using a cloud provider’s pay-as-you-go service.
• Data privacy: Hosting your own model can be beneficial for data privacy, as your data doesn’t have to be sent to a third party.
• Security: Open-weights models might have security vulnerabilities, with some models failing cybersecurity tests. However, for some internal applications, like document classification, these vulnerabilities may not be as relevant.
• Interoperability: Businesses should consider interoperability and how easily models can be swapped in and out of their systems, and may want to use smaller models for specific tasks.

 

Potential Use Cases for DeepSeek Models

With this in mind, there are various potential use cases for DeepSeek models:

• Automation and efficiency: DeepSeek’s cost-effective models can significantly reduce the cost of high-throughput, inference-serving use cases.
• Personalization: As chatbots become more common, these models can power user-facing applications at competitive prices.
• Agentic systems: Reasoning models are well-suited for building agentic systems where complex decisions are required.
• Research and productivity: Reasoning models can assist in research and boost internal productivity.
• Marketing: Reasoning models can help create effective marketing campaigns by analyzing large amounts of data and past performance.

 

Risks + Comparisons to other models

DeepSeek’s models offer significant advantages in terms of cost and performance, but there are potential risks and downsides to consider, particularly when compared to more established models like Google Cloud’s. 

While DeepSeek offers open-weights models, the company is based in China, which raises data privacy concerns for some users, especially when using the DeepSeek API, as data may be subject to Chinese government directives. Many users have found questions or topics that DeepSeek refuses to discuss. This is not a risk with Google’s models. 

Furthermore, with such low costs, DeepSeek might not have put as much effort into security as it has into performance. Although DeepSeek’s models are performant, the potential trade-off in security is something businesses should carefully consider, especially if the use case involves sensitive data or public-facing applications.

Conclusion

DeepSeek’s advancements in AI model development have significantly disrupted the GenAI market. Their open-weights reasoning model and lower training costs have made advanced AI more accessible and affordable for businesses. 

While questions surrounding open-weights models, hosting, and security still need to be addressed, the potential for innovation and wider adoption is undeniable. As the industry continues to evolve, DeepSeek’s impact will likely encourage further advancements and lower prices, making AI tools more accessible to everyone.