DeepSeek R1: Why is it so Cheap? - 27/01/2025

DeepSeek’s R1 has been making waves in the AI community, offering performance comparable to industry giants like OpenAI, but at a fraction of the cost. It’s natural to wonder how they can achieve such a feat. Here’s a breakdown of the factors contributing to DeepSeek R1’s remarkably low price.

Hardware Efficiency

H800 Chips

DeepSeek utilizes NVIDIA’s H800 chips, a less expensive and slightly less powerful version of the H100. While this presents a compute disadvantage, DeepSeek compensates with architectural innovations and efficient resource utilization.

Mixture-of-Experts (MoE)

This system activates only the necessary neural networks for specific tasks, significantly reducing computational costs and improving efficiency on the H800s.

Architectural and Training Innovations

Multi-Head Latent Attention (MLA)

This optimizes attention mechanisms by compressing key-value vectors, boosting cache efficiency and reducing memory usage.

FP8 Precision

Using 8-bit floating-point numbers for computations accelerates processing and reduces memory requirements while maintaining accuracy.

Multi-Token Prediction (MTP)

This technique predicts multiple tokens sequentially during training, enhancing both efficiency and the coherence of generated text.

Caching and Subsidies

Aggressive Caching

DeepSeek employs a robust caching system that significantly reduces costs for repetitive queries. Every cache hit translates to lower costs for users.

Potential Subsidies

While not explicitly confirmed, there are indications that DeepSeek might be subsidizing its services, potentially operating at a loss to gain market share and drive adoption.

Other Contributing Factors:

Cheaper Costs in China

Operating in China likely provides DeepSeek with advantages in terms of labor and operational costs compared to companies based in the US.

The Impact of DeepSeek R1

DeepSeek R1’s affordability has sent ripples through the AI industry, potentially altering the landscape of AI development and accessibility. Its success demonstrates that high performance doesn’t necessarily require exorbitant costs, paving the way for a more competitive and inclusive AI ecosystem.