Fact

DeepSeek — Description: DeepSeek-V3 training cost approximately $5.58M using 2,788K H800 GPU hours over ~2 months on 2,048 H800 GPUs

partial85% confidence

1 evidence check

Last checked: 3/31/2026

The source confirms the GPU hours (2.788M H800 GPU hours) and token count (14.8T tokens) from the claim. However, the source text excerpt does not explicitly state the $5.58M cost figure, the ~2 months duration, or the 2,048 H800 GPUs specification. While these details may appear elsewhere in the full technical report, they are not present in the provided excerpt. The claim's core technical specifications (GPU hours and tokens) are confirmed, but the cost estimate and hardware configuration details cannot be verified from this excerpt alone.

Evidence — 1 source, 1 check

arxiv.org/abs/2412.19437(1 check)

partial85%primaryHaiku 4.5 · 3/31/2026

Found: The source states: 'DeepSeek-V3 requires only 2.788M H800 GPU hours for its full training' and 'We pre-train DeepSeek-V3 on 14.8 trillion diverse and high-quality tokens'

Note: The source confirms the GPU hours (2.788M H800 GPU hours) and token count (14.8T tokens) from the claim. However, the source text excerpt does not explicitly state the $5.58M cost figure, the ~2 months duration, or the 2,048 H800 GPUs specification. While these details may appear elsewhere in the full technical report, they are not present in the provided excerpt. The claim's core technical specifications (GPU hours and tokens) are confirmed, but the cost estimate and hardware configuration details cannot be verified from this excerpt alone.

Debug info

Record type: fact

Record ID: f_TCKCfktktg