How DeepSeek’s LLMs Slash Training Costs and Reshape China’s Compute Landscape

DeepSeek has released three versions of its large language model (LLM) family: the base model V3, the reinforcement‑learning‑enhanced R1‑Zero, and the generalized‑reasoning R1. All three claim strong performance, but the standout is the dramatically reduced training cost.

The V3 model was trained on 2,048 H800 GPUs for roughly two months, consuming about 2.788 million GPU‑hours. At an estimated $0.20 per GPU‑hour, the total training expenditure was approximately $557,600 USD. By contrast, a comparable model such as Llama‑3‑405B required about 30.8 million GPU‑hours, translating to $0.6‑1 billion in training costs—over eleven times higher than DeepSeek V3.

The R1 model builds on V3’s foundation by incorporating large‑scale reinforcement learning and multi‑stage training pipelines, further boosting inference capabilities while potentially keeping costs lower than traditional approaches.

These efficiency gains have significant implications for China’s compute industry. Lower training budgets reduce entry barriers for AI developers, stimulate demand for high‑performance GPU clusters, and may prompt a restructuring of the domestic hardware and cloud‑service supply chain to accommodate the surge in cost‑effective AI workloads.