Capital · out
lean compute — training costs cited in low-single-digit $M; no confirmed large external multi-year compute commitment (export-locked, though China has granted conditional H200 access, and a domestic-chip manufacturing effort is underway)
Sourced figures 1
- Wikipedia — DeepSeek V3 training ~$5.576M — the whole point is doing it cheap
the working
Threads 1
A claim is one metric we assert about a node — the value is the summary, this page is the receipt. Sourced from kestrel; supersedes-history not yet shown. Back to DeepSeek.