DeepSeek's Low-Cost Large Model Training Explained

DeepSeek's level of cost control has become legendary in the industry:

  • V3 training cost: approximately $5.5 million
  • R1 GPU usage fee: approximately $294,000

During the same period, U.S. companies budgeted $100 million to $1 billion for training models of similar scale. That's at least an order of magnitude difference.

Cost-Saving Secrets

1. FP8 Mixed-Precision Training

Reducing training precision from FP32/FP16 to FP8 cuts memory bandwidth requirements in half. Since most computational bottlenecks are tied to memory bandwidth, this is a very precise move.

2. Strategic Use of Spot Instances

Computing costs are reduced by 70%. Spot instances can be reclaimed at any time, but DeepSeek's training framework includes sufficient checkpointing and fault tolerance to handle this.

3. MoE Architecture Naturally Saves Compute

The 671B-parameter model activates only 37B parameters per inference. The parameter count ensures capacity and capability, while sparse activation keeps costs manageable.

New Technology: mHC

In January of this year, DeepSeek publicly disclosed a new method—Manifold-Constrained Hyper-Connections (mHC). The core paper is co-authored by founder Liang Wenfeng.

What this method does: it creates multiple information streams within the model, with strict mathematical constraints at each mixing step to ensure total information conservation. Test results show that mHC enables stable training at parameter scales from 3B to 27B, while unconstrained versions are often unstable. The training overhead increases by only 6-7%, which is nearly negligible for large models.

Some analysts have described the mHC paper as a "stunning breakthrough" that could become the foundational technology for DeepSeek's next-generation flagship model.

OpenAI's CEO himself has acknowledged that DeepSeek's R1 operating cost is 20 to 50 times cheaper than OpenAI's comparable model.

Sources: CocoLoop, Computerworld reports, South China Morning Post analysis