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

Verdictconfirmed98%
1 check · 4/14/2026

The source text directly confirms all key elements of the claim: (1) Total training cost of 2.788M H800 GPU hours matches exactly; (2) The cost calculation of approximately $5.58M is confirmed (source states $5.576M, which rounds to $5.58M); (3) Pre-training on 14.8T tokens is explicitly stated; (4) Training duration of ~2 months ('less than two months') is confirmed; (5) 2,048 H800 GPUs is explicitly mentioned. The claim's description accurately reflects Table 1 and the surrounding text in the technical report. All numerical values align within rounding conventions.

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
confirmed98%primaryHaiku 4.5 · 4/14/2026

NoteThe source text directly confirms all key elements of the claim: (1) Total training cost of 2.788M H800 GPU hours matches exactly; (2) The cost calculation of approximately $5.58M is confirmed (source states $5.576M, which rounds to $5.58M); (3) Pre-training on 14.8T tokens is explicitly stated; (4) Training duration of ~2 months ('less than two months') is confirmed; (5) 2,048 H800 GPUs is explicitly mentioned. The claim's description accurately reflects Table 1 and the surrounding text in the technical report. All numerical values align within rounding conventions.

Case № f_TCKCfktktgFiled 4/14/2026Confidence 98%
Source Check: Description | Longterm Wiki