Is DeepSeek the next revolution in generative AI?

On January 20, DeepSeek, an AI company based in China, announced the R1 of its AI model. This caused a massive reaction around the world, both in the media and in the stock markets. While some have categorized the reaction as overkill or hype, the market’s reaction clearly shows that DeepSeek needs further investigation and validation.

This article aims to share what we know about DeepSeek currently and what we would advise SHI customers to do in the immediate future, pending further research.

What is DeepSeek?

DeepSeek’s chatbot provides functionality similar to ChatGPT, Claude, Copilot, or Perplexity. What is significant about DeepSeek’s announcement is that it claims to have created this model at a fraction of the cost required to create similar models by leveraging different training techniques.

AI models are typically created, or “trained”, by feeding an algorithm massive amounts of data. This allows the models to learn patterns and thereby make accurate predictions that appear in the form of results. This traditional process involves data scientists taking advantage of large clusters of computers that have hundreds, often thousands, of powerful GPUs. The process is typically iterative, involving multiple runs and verification to achieve the level of accuracy desired.

These resources and expertise are very expensive and in high demand, resulting in billions of dollars in costs to create these models. DeepSeek claims to have employed unique training techniques and lower-cost hardware to develop its model, which reduced development costs to approximately $6 million.* The company also claims to have leveraged a technique called inferencing to reduce the compute required to execute the models.

*These are DeepSeek’s own figures and have not been validated by independent sources.

Claimed DeepSeek breakthroughs

DeepSeek has reported advances in several areas:

Cost efficiency

According to the announcement, which still must be validated as noted above, DeepSeek developed its R1 model at 3-5% of the cost of other models.

Novel approach to model training

DeepSeek leveraged pure reinforcement learning combined with distillation reasoning, as well as other techniques such as a multi-stage training pipeline.

Hardware limitations

According to its own announcement, DeepSeek primarily uses H800 GPUs — which are much more memory-constrained than their trade-restricted H100 counterparts — as well as cluster scaling.

Performance parity

The DeepSeek announcement claims to match or exceed OpenAI’s o1 model on many benchmarks.

Impact of the DeepSeek announcement  

While the breakthrough and cost reductions are significant, there are a few relevant points that need further discussion and research:

• The $6M cost, while remarkable, still must be validated to ensure it includes ALL the costs associated with the model development to accurately be compared to other model development costs.
• While the DeepSeek announcement is significant for the AI community and will likely influence future model development, much more detail is required to understand the full costs and impacts.
• With respect to the impact on tech stocks due to its disruptive nature, the DeepSeek announcement has the potential to somewhat curb demand for high-end GPUs. But DeepSeek still required thousands of GPUs to accomplish their task, and GPUs will still be required to execute the model at scale. History has also shown that demand for compute power has continued to scale regardless of technological advancements of the applications that run on them.
• There are geopolitical and economic impacts that cannot be ignored, but will not be commented on further in this blog post.
• Finally, most of our customers are not involved in foundational model development and training, which is what the $6M announcement involved. However, many customers do perform model fine-tuning, and literally all customers are using their models for inferencing (getting the results from the model). The DeepSeek announcement does, in fact, have potential impacts on these activities as well.

SHI’s recommendations for customers  

Like many other models, DeepSeek is open-source and readily available. However, like any model, it serves specific use cases and isn’t a universal solution. Additionally, like ChatGPT, Claude, etc., using the DeepSeek model from their website carries security implications, including, but not necessarily limited to, the company’s ties to the Chinese government.

DeepSeek openly admits to collecting user data on their website beyond that of other chatbot websites, and all models use data input by users to train themselves and thus become more accurate. For these reasons, SHI strongly recommends that customers not leverage any publicly available chatbots where any proprietary data could be captured or exposed.

For DeepSeek specifically, additional research, testing, and experimentation are required before implementation is considered. Additionally, many models require additional information — typically company data — to produce the results customers require, especially if customer-specific answers are required. All this is a long way of saying that proper model selection and implementation are complicated and typically require expertise.

For these reasons, SHI has installed the DeepSeek model in our secure Advanced Gen AI lab and is performing testing including benchmarking and validation of results to corroborate the training, but also to confirm there is no training bias given the various restrictions in China.

The DeepSeek announcement is exciting and has the potential to revolutionize AI model development. However, more information is needed to verify everything claimed.

Caution is recommended, given the source.

For more information, SHI customers are invited to reach out to our AI experts via your account team.

As we continue to research this significant development in the AI world, we will publish additional blogs. We will share the results of our own testing and share reactions from additional experts within SHI.

Cory Peters, Michael Hicks, and Russ Cantwell also contributed to this article.