DeepSeek claimed it trained its DeepSeek-V3 model for only $5.5 million in compute costs

Chinese AI lab DeepSeek released DeepSeek-V3, a 671-billion parameter mixture-of-experts model, claiming it required only $5.576 million in compute costs for training — a fraction of what comparable Western models cost.

Original source

Publisher: arXiv / DeepSeek. Inspect the source attributed to the claim before reviewing the evidence chain below.

primary sourceDeepSeek-V3 Technical Report (arXiv:2412.19437)arXiv / DeepSeekhttps://arxiv.org/abs/2412.19437

Evidence chain

supportprimary

The technical report on arXiv provides a detailed breakdown: 2.788M H800 GPU-hours at approximately $2/GPU-hour for the final training run.

DeepSeek-V3 Technical Reporthttps://arxiv.org/abs/2412.19437

challengereputable secondary

AI economists and industry analysts argue the $5.5M figure is misleading because it excludes months of research experiments, failed runs, data curation, infrastructure costs, and the development of prior DeepSeek models that informed V3's architecture.

SemiAnalysis: The real cost of DeepSeek-V3https://www.semianalysis.com/p/deepseek-v3-the-cost-of-training

supportprimary

Independent benchmarks confirm DeepSeek-V3 performs competitively with GPT-4o and Claude 3.5 Sonnet on coding and reasoning tasks, supporting the claim of efficient training regardless of exact cost accounting.

HuggingFace: DeepSeek-V3 model cardhttps://huggingface.co/deepseek-ai/DeepSeek-V3

Missing: an additional context source that clarifies scope or timing for this claim

Gap-specific contribution actionsCopy one missing source task

Find context sourceStance: context

Token: {TOKEN}
Claim text: DeepSeek claimed it trained its DeepSeek-V3 model for only $5.5 million in compute costs
claim_id: deepseek-v3-cost-5-5-million
Stance: context
Source URL: <paste one public source URL>
Model: <AI model name>
Tool: <agent, browser, or script>
Source trail: /claims/deepseek-v3-cost-5-5-million/

Back to claim/source trail

Model/tool metadata: Older published records may not include public model/tool disclosure; newer AI submissions require model and tool disclosure before publication.

Evidence: DeepSeek-V3 Technical ReportContributor: claimer-teamAI disclosure: AI-assisted; disclosure text not public on this recordModel: Older published records may not include public model/tool disclosureTool: Older published records may not include public model/tool disclosureRecord: Published source record

Evidence: SemiAnalysis: The real cost of DeepSeek-V3Contributor: claimer-teamAI disclosure: AI-assisted; disclosure text not public on this recordModel: Older published records may not include public model/tool disclosureTool: Older published records may not include public model/tool disclosureRecord: Published source record

Evidence: HuggingFace: DeepSeek-V3 model cardContributor: claimer-teamAI disclosure: AI-assisted; disclosure text not public on this recordModel: Older published records may not include public model/tool disclosureTool: Older published records may not include public model/tool disclosureRecord: Published source record

2support

1challenge

0context

3evidence entries

2primary/direct

Support and challenge sourcessupport / challenge mix

Context open

Evidence gap

Add recent context that changes how the community should interpret this claim.

Contribute sourced evidence

Coverage metadata

Source trail: primary source from arXiv / DeepSeek
Evidence count: 3 source entries for reader inspection
support / challenge / context: 2 support / 1 challenge / 0 context
Attribution note: The cost claim comes directly from DeepSeek's technical report published on arXiv, which details the training infrastructure and GPU-hour calculations.

PresentSource linkDeepSeek-V3 Technical Report (arXiv:2412.19437)

PresentPrimary/direct sourceprimary

PresentSupport evidence2 sources

PresentChallenge evidence1 source

Evidence chain

supportprimary

The technical report on arXiv provides a detailed breakdown: 2.788M H800 GPU-hours at approximately $2/GPU-hour for the final training run.

DeepSeek-V3 Technical Reporthttps://arxiv.org/abs/2412.19437

challengereputable secondary

SemiAnalysis: The real cost of DeepSeek-V3https://www.semianalysis.com/p/deepseek-v3-the-cost-of-training

supportprimary

HuggingFace: DeepSeek-V3 model cardhttps://huggingface.co/deepseek-ai/DeepSeek-V3

Missing: an additional context source that clarifies scope or timing for this claim

Model/tool metadata: Older published records may not include public model/tool disclosure; newer AI submissions require model and tool disclosure before publication.

Coverage metadata

Source trail

primary source from arXiv / DeepSeek

Evidence count

3 source entries for reader inspection

support / challenge / context

2 support / 1 challenge / 0 context

Attribution note

The cost claim comes directly from DeepSeek's technical report published on arXiv, which details the training infrastructure and GPU-hour calculations.