Test a claim Challenge my view Assess a new source Build a brief
The claim you want to test Examples and search work without a key. Find evidence Try a claim The $5.6m claim Training without Nvidia Beyond a benchmark Access through diversion Who received the subsidy?
SAVED EXAMPLE
Evidence for this claim 2 records The disclosed training estimate is not the full cost of developing the model.
All evidence Supports Challenges & limits
Source selectionAll sources Exclude company-authored sources Independent evaluations only Research caseAll five cases DeepSeek-V3 efficiency Pangu domestic training CloudMatrix systems engineering Rerouted H100/H200 access State support evidence gap Independent evaluation describes a source type, not verification of every claim. Filters never rewrite the published findings.
The widely repeated $5.576 million figure is a rental-equivalent training-compute estimate, not an all-in development cost. Evidence & limitations RECORDED QUOTATION OR DATA
The report excludes prior research and ablation experiments from its training-cost estimate.
COUNTEREVIDENCE
The narrow figure remains useful for comparing the disclosed final run if its scope is stated.
STILL UNRESOLVED
R&D, acquisition, depreciation, staff, data, networking, and failed-run costs are undisclosed.
Research confidence: High Section 1, note immediately below Table 1
Full evidence record (opens in a new tab) DeepSeek reports pre-training a 671B-parameter MoE on 14.8 trillion tokens and using 2.788 million H800 GPU-hours across all disclosed training stages. Evidence & limitations RECORDED QUOTATION OR DATA
“2.788M H800 GPU hours” for the full training run.
COUNTEREVIDENCE
The run and benchmarks are author-reported and have not been independently reproduced end to end.
STILL UNRESOLVED
The public report does not establish acquisition dates, electricity use, or total program cost.
Research confidence: High Abstract; section 1; Table 1
Full evidence record (opens in a new tab)