Why DeepSeek Is Outpacing ChatGPT: Cost, Performance, and Local Deployment

DeepSeek has become the hottest topic of the Chinese New Year period, reaching 40 million global downloads in less than a month and surpassing ChatGPT as the fastest‑growing AI application, while being fully open source.

DeepSeek vs. ChatGPT

To assess DeepSeek’s capabilities, the simplest method is to compare it directly with OpenAI’s ChatGPT.

Model Classification

ChatGPT offers two large models: the 4‑OMNI model for general users and the O1 model for professional, deep‑thinking tasks. DeepSeek’s counterparts are the V3 model and the R1 model, respectively.

Practical Performance

In real‑world use, DeepSeek’s performance matches or exceeds ChatGPT, especially in solving mathematics, physics, and reasoning problems, and it responds faster. For example, DeepSeek correctly handles the classic “Which is larger, 1.11 or 1.9?” query with a clear reasoning process.

DeepSeek not only provides the correct answer but also shows detailed solution steps and verification, demonstrating superior Chinese language understanding compared to ChatGPT.

Cost

The cost of AI models consists of hardware expenses and training expenses. While model algorithms are largely based on public research, the dominant cost lies in the hardware—especially GPUs. Nvidia’s high‑end B200 GPU can cost up to $40,000 per unit, and limited supply forces companies to compete for chips.

Nvidia’s CUDA (Compute Unified Device Architecture) provides a parallel computing platform that lets developers harness GPU power for AI workloads. More CUDA cores translate to greater AI compute capability.

Because Nvidia controls CUDA, AI model costs remain high, and US export restrictions on Nvidia chips hinder China’s AI development. OpenAI, with unrestricted access, reportedly used 50,000 Nvidia GPUs to train ChatGPT, whereas DeepSeek, a Chinese company, operates with far fewer resources.

DeepSeek’s Impact

Despite limited hardware, DeepSeek achieves performance comparable to ChatGPT, even surpassing it in certain areas. It bypasses CUDA limitations by using innovative training techniques that directly leverage GPU resources, allowing the use of lower‑end chips and reducing training time.

DeepSeek’s approach also enables training on Huawei chips, cutting service costs to about one‑tenth of ChatGPT’s price, offering high cost‑effectiveness and broad application prospects.

The success of DeepSeek has shaken the AI market: Nvidia’s stock fell 17% in a single day, erasing nearly $600 billion in market value, and major cloud providers such as Nvidia, Microsoft, and Amazon announced integration of DeepSeek models, marking a “Sputnik moment” for the AI industry.

Local Deployment of the R1 Model

To avoid network attacks, reduce latency, protect sensitive data, and enable flexible fine‑tuning, the R1 model can be deployed locally.

Install Ollama

Visit https://ollama.com/, download the installer, and install it on your machine.

Pull DeepSeek R1 Model

Open Ollama’s model library at https://ollama.com/library/deepseek-r1 and select the DeepSeek‑R1 model.

GPU Requirements

Typical desktop computers can run the 1.5B or 7B versions. After copying the provided command, paste it into a terminal; when the download reaches 100%, the model is stored locally.

Run

ollama list

to view installed models. The R1 model can now be used for question‑answering via the command line.

GUI Packaging

For a more user‑friendly interface, use a GUI such as Cherry Studio. After installing, select Ollama as the model service and configure the local model details.

After configuring the model, you can start using DeepSeek locally.