Skip to content
Longterm Wiki
Index
Fact·f_TCKCfktktg·Fact

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

Verdictconfirmed95%
1 check · 4/30/2026

1 → confirmed

Our claim

entire record
Subject
DeepSeek
Property
Description
Value
DeepSeek-V3 training cost approximately $5.58M using 2,788K H800 GPU hours over ~2 months on 2,048 H800 GPUs
As Of
December 2024
Notes
Dramatically lower than comparable frontier models; pre-training on 14.8T tokens

Source evidence

1 src · 1 check
confirmed95%primaryHaiku 4.5 · 4/30/2026

NoteThe source directly confirms all key elements of the claim: (1) Total GPU hours: 2.788M H800 hours (source states '2.788M' in abstract and Table 1 shows 2788K); (2) Training cost: $5.576M (source states this in abstract and Table 1, which rounds to approximately $5.58M as claimed); (3) Duration: ~2 months (source states 'less than two months' for pre-training stage); (4) GPU count: 2048 H800 GPUs (source explicitly states this); (5) Pre-training tokens: 14.8T (confirmed in abstract and text). Minor rounding difference ($5.576M vs $5.58M) is within acceptable tolerance and represents the same value in different precision levels.

Case № f_TCKCfktktgFiled 4/30/2026Confidence 95%