On January 27, 2025, a stock market rout among computer hardware makers, including Micron, AMD, and most notably Nvidia (which lost 17% in a day!) was attributed to DeepSeek-R1, a new model from a mysterious Chinese company. There has since been a lot of (mostly terrible) commentary on the event, and many questions like “Who is DeepSeek?” “What makes DeepSeek different?” and “Is DeepSeek safe to use?”
So here’s an explainer to catch you up!
Who is DeepSeek?
DeepSeek is a Chinese artificial intelligence company that develops a variety of generative AI models. They also offer direct access to their models for free or very low cost, and have an app available in multiple app stores for easy use. You can find out more about them on their website or Wikipedia page.
What Can DeepSeek Do?
Like other AI companies, DeepSeek offers a variety of models that can generate results from input images or text. In particular, the DeepSeek-R1 model is a large language model – just like those from Google, Microsoft, OpenAI, and Anthropic – which generates text responses based on text prompts. It can be used for the same things all of those other models can be used for – learning, coding, writing, planning, and more.
What Makes DeepSeek Different?
The DeepSeek-R1 model is special in three ways – how it was made, how it works, and how it can be used.
Most of the major model providers, like OpenAI, emphasize the importance of scale in training their models. They claim you need enormous server farms running incredible quantities of calculations to train AI systems as powerful as theirs. What DeepSeek found was that – this isn’t necessarily the case! While the exact cost and resources used to train their model are still disputed, what’s clear is that they didn’t require the same scale of hardware as Google or OpenAI, and the total training cost was likely in the 7-8 figure range. DeepSeek has published multiple papers on how they accomplished this, which include several clever improvements on existing techniques.
This leads to how it works. DeepSeek-R1 uses a technique called “chain-of-thought”, where it’s initial response to a user prompt is an internal dialogue “thinking” through the problem, after which it provides a “user-facing” answer based on those thoughts. In reality, the user can see the internal dialogue just as clearly as the answer, but the internal dialogue helps the model work through a solution before presenting it. This is the same capability as OpenAI’s latest o1 model – and DeepSeek-R1’s implementation appears to deliver better results than o1 in many standard AI tests.
But what really sets it apart is how it can be used. While its performance is on par with OpenAI’s o1, o1 is a closed model, meaning that it is only available directly from OpenAI’s website, app, or API. DeepSeek-R1 is far more open and can be used in multiple ways. You can interact with the model via DeepSeek’s website or app, use a version run by a third-party provider, or run it directly on your own computer. While open models aren’t new (I’ve written about them multiple times previously), having an o1-class model available to run on your own computer is new.
How Do I Access DeepSeek? Is DeepSeek Safe To Use?
You can access DeepSeek-R1 in a number of ways. First, you can use it via DeepSeek’s website or app. A number of third-party providers also serve the model, including Groq and Replicate. You can also run it on your own computer using a tool like Ollama (use it with Open WebUI for the best experience) or LM Studio.
The “safety” of the model depends on how you interact with it. If you use DeepSeek’s website or app, or other third-party providers, you have to send your data to them, and it’s unclear what their data management policies are and how honestly they stick to them. You want to be thoughtful about sharing confidential or personal information and aware of the potential consequences of misuse.
If you run the model locally, on your own computer, it’s perfectly safe. Nothing you share leaves your computer. There are no viruses or malware that will damage your computer or affect your data.
Is DeepSeek open source?
The short answer is: somewhat. DeepSeek has released the models, weights, and white papers describing the training process for the systems. This allows anyone with an internet connection to run a copy of the models, as well as learn from DeepSeek’s work. However, DeepSeek has not released the datasets they used for training. Strictly this means that these models aren’t fully open-source, although they meet most of the criteria.
Why did DeepSeek affect Nvidia?
This isn’t entirely clear. DeepSeek had already been releasing quality models for several months, and they were highly regarded in the generative AI community. The usage of their product had grown significantly with the recent release of their app in January 2025; it had become very popular in multiple app stores and so increased awareness of their work. It’s unclear what the specific tipping point was that brought it to the attention of major market makers.
The common explanation for the impact on Nvidia was DeepSeek’s training process. Nvidia has been the world’s primary supplier of the GPU technology used for training and running generative AI. With American labs focused on scale (i.e. lots of hardware) to improve their models, Nvidia was considered one of the most valuable companies on the planet. However, with DeepSeek showing that great models can be made with smaller amounts of hardware, this challenged the valuation of Nvidia as a company.
What complicates the story is that it wasn’t just Nvidia stock that dropped on January 27, 2025 – other computer hardware companies like AMD and Micron also saw their stock price drop, despite their minor involvement in AI to date. Perhaps traders were just covering their bases?
Another complicating factor is that, the following day, the US announced a plan for significant tariffs on semiconductors originating in Taiwan (where AMD, Nvidia, and a host of others manufacture the chips central to their products). If it had come a day earlier, this would have been a clear explanation for the dip in computer hardware stocks. But a day afterward? That would suggest insider trading.
So, were some whales insider-trading? Or were out-of-the-know traders covering their bases? It’s unclear, and it’s unlikely we’ll get a final explanation.
What Does DeepSeek Mean For My Business?
For most businesses, the importance of this development is relatively small. While it’s great to see continued advancement in the availability of powerful, open models, there are some limitations that would prevent immediate adoption of DeepSeek systems.
- Data security is a major concern. Deepseek is based in China, and it is unclear where they are serving their models from. Western companies serving the models may also not meet the privacy or security requirements of many companies – at least not for work involving sensitive information.
- The best models are large enough that they must be run on expensive hardware, and few companies are willing to undertake the infrastructure and maintenance necessary for that at this stage. (Smaller models can be run on consumer hardware, but they’re not quite as good as the big models.)
- While thinking models like DeepSeek-R1 are great for solving challenging problems, they’re not necessarily required for day-to-day business tasks like customer service, programming, email and project management, knowledge discovery, and more. Much smaller and equally-powerful models are already available for this sort of work.
If your company is using generative AI for research tasks that require multi-step thinking, and you are experimenting with a variety of models, DeepSeek-R1 is a great option to try. Otherwise, I’d recommend staying with the models and providers that you’re already using.
Become your company’s AI expert in under 30 minutes a month by signing up for the Executive Summary newsletter by AI For Business.
If you liked this post, have future posts delivered straight to your inbox by subscribing to the AI For Business newsletter. Thank you!