DeepSeek vs. ChatGPT: Fact check on the Chinese AI language model

AI is a highly complex topic that even many technicians do not understand. Added to this is reporting that is often inaccurate for reasons of necessary simplification or a need for attention. The result is statements that paint a distorted picture of DeepSeek. The most important facts in plain language.

Introduction

DeepSeek is a Chinese company. The DeepSeek-R1 language model what recently published by this company. It is said to be just as good and in parts even better than OpenAI's language model o1 ("ChatGPT").

This led to company values of AI companies such as Nvidia being at rock bottom. Even the data protection organization noyb reported this in their newsletter on 30.01.2025

It is often presented as if DeepSeek is significantly more efficient than ChatGPT. This is true in the relevant aspects, but less so in others.

Then you read headlines like this:

This gives the impression that the Chinese language model is not secure because user data may be misused.

Most of the statements of this kind circulating in the public domain are not fully accurate.

Facts about DeepSeek

Data protection & data security

DeepSeek is the name of a Chinese company. If DeepSeek is used instead as a name for a language model, a distinction must be made between two variants:

1. Cloud version, also referred to as "app
2. Open source language model, can be used completely independently

The reporting above that "DeepSeek" probably misuses user data can only refer to the cloud version ("App"). Because the open-source language model can be downloaded and run locally, completely without internet connection. Without such a connection, bad user data cannot wander off to China.

ChatGPT is not entirely secure, after all American intelligence laws allow US authorities and US intelligence agencies to access data from others. The EU-US Data Protection Agreement DPF what never much worth and only a formality. It's based on a presidential decree by Joe Biden. With the dissolution of an important committee by Trump it will be further devalued. The presidential decree could just as easily be declared null and void by Donald Trump.

OpenAI also has a lot of fun collecting your data. Even if your ChatGPT data is not used for AI training, it might be used for other purposes! For example, to evaluate the OpenAI AI, which then makes you more and more dependent (price increases have already been announced).

Efficiency

It is allegedly claimed that the training of DeepSeek-R1 cost around $6 million. The actual costs were higher, as this number does not refer to the total costs and also refers to the basic model DeepSea Explorer V3. A sum of $100 million what reported for ChatGPT.

That is definitely true:

• DeepSeek-R1 can be operated on significantly cheaper hardware than ChatGPT
• DeepSeek-R1 responds much faster than ChatGPT because it is much "smaller
• The operation of DeepSeek-R1 therefore consumes significantly less energy

Why is DeepSeek-R1 smaller than ChatGPT? According to DeepSeek, R1 is a 685B model, i.e. it consists of 685 billion neuron connections. ChatGPT is probably similar in size. So what now?

DeepSeek-R1 works so like the human brain: When you speak, predominantly the language center is activated. When speaking, therefore only a few of all your neurons in the brain fire. Technically this is achieved by DeepSeek-R1 through a so-called Mixture of Experts architecture. This architecture has been common knowledge for a long time. It what already used by Mistral.

Because DeepSeek-R1 is open source, it can be downloaded and operated by the user. To run DeepSeek-R1 on your own hardware, you need a server that costs around 30,000 euros. Many companies can afford this. ChatGPT, on the other hand, you don't want to run on your own hardware, apart from the fact that you can't because OpenAI doesn't want to and therefore hasn't released the model.

Quality

In various benchmarks, DeepSeek-R1 performs just as well as OpenAI o1. This is despite the fact that R1 is much more efficient and smaller than ChatGPT. Users report that R1 is just as good as ChatGPT, while others see R1 ahead.

The Chinese censors have deleted or distorted some facts from the model. As a result, the quality of some political questions is poor.

However, a general chatbot is the worst possible use case for a corporate AI. In this respect, it almost doesn't matter that some political facts in R1 are questionable. With standard procedures such as fine-tuning or RAG, text applications can be operated very well with R1. Other use cases can be realized even better with R1. These include:

• Consult company knowledge (consult your documents),
• Fulfilling the obligations arising from the Supply Chain Act, for example by evaluating supplier documents,
• Creating summaries, translations or language simplifications.
• Programming wizard

Above all, the last application case is highly interesting. So it what possible for the author to program within a few hours what would have taken several weeks otherwise. And mostly on the sofa at the unfriendly tablet and without microphone usage. It's just a difference whether one programs with AI in a playful way for 30 minutes or if one achieves the same result without fun after two days at best.

Further advantages of DeepSeek-R1 over ChatGPT

OpenAI regularly releases or introduces a new sub-version of ChatGPT. These variants differ in their response to your question. Consistency is not given here. Without consistency, no Reliability when automating processes.