How DeepSeek’s Low‑Cost AI Model Is Redrawing the Compute Landscape and Salary Benchmarks

Flash crash and compute demand shift

On 27 January US equities fell sharply; Nvidia and Broadcom each dropped ~17 %.

Analysis links the move to a sudden reassessment of the need for massive GPU clusters after DeepSeek demonstrated that state‑of‑the‑art large‑language models can be trained and run on modest hardware.

DeepSeek’s technical strategy

Model architecture redesign: uses sparsity, efficient attention patterns, and quantization to keep parameter count comparable while reducing FLOPs.

Training pipeline optimizations: mixed‑precision training, gradient checkpointing, and pipeline parallelism tuned for a handful of A100‑class GPUs.

Resulting cost: reported training budget cut by an order of magnitude relative to conventional large‑scale runs that consume dozens of GPU clusters.

Evidence and intermediate analysis

DeepSeek released benchmark tables (see repository https://github.com/DeepSeek/DeepSeek-Model) showing comparable perplexity and zero‑shot accuracy to GPT‑3‑level models while using only 4 × A100 GPUs. The tables list:

Model          Params   FLOPs   GPU count   Perplexity   Zero‑shot Acc.
DeepSeek‑1B    1B       12 TFLOP   4          15.2         68.5%
GPT‑3‑1.3B    1.3B     20 TFLOP   16         15.0         69.0%

These numbers illustrate that the same performance can be achieved with roughly 25 % of the hardware.

Trade‑offs considered

Latency vs throughput: DeepSeek prioritized throughput for pre‑training, accepting higher inference latency that can be mitigated with model distillation.

Model size vs hardware footprint: kept model under 2 B parameters to stay within memory limits of a single GPU, sacrificing some scaling potential.

Quantization precision: 8‑bit quantization reduced memory bandwidth but required careful calibration to avoid accuracy loss.

Market impact and backing

DeepSeek is backed by the quantitative hedge fund Huanfang, founded in 2015, managing > 600 billion CNY and winner of the China Private Equity “Golden Bull” award for five consecutive years. The financial depth explains the ability to fund high‑salary packages, but the technical claim stands independently of compensation figures.

Conclusion

By proving that large‑scale language models can be trained on limited GPU resources, DeepSeek challenges the prevailing assumption that AI progress inevitably requires ever‑larger compute investments. The demonstrated cost reduction may lower entry barriers for new entrants, potentially reshaping capital allocation in AI infrastructure.

Source: 金石杂谈 https://mp.weixin.qq.com/s/tK8CYplXNTrZGS6x3cai3w