fact F-0045 Verified

DePue revises his GPU-count and cost estimates upward on the same day

Date: 2026-11-12
Topics: blackwell
File: facts/F-0045-depue-revises-his-gpu-count-and-cost-estimates-upward-on-the.md

Report: first version gave 2,000–3,000 Blackwell GPUs and $150–200M for the largest cluster; revised to best estimates of ~4,500 (Kimi K3) and ~3,000 (DeepSeek-V4-Pro), ~$200–300M for the largest cluster, >$0.5B across five runs, with ranges as wide as 1,000–9,500. Stated reason: Blackwell training runs disclosed so far ran less efficiently than Nvidia benchmarks. The X lead post still shows the original figures. Several news write-ups relay the original figures.

Sources

S-0045 · Unemployment Labs (report) · 2026-11-12 · link (full)
S-0044 · X (post + thread) · 2026-11-12 · link (full)

Referenced by (2)

source S-0044 Will DePue, X thread: "I have a major result to share" (Chinese labs on Blackwell)
claim C-0046 DePue: the largest runs used thousands of Blackwell GPUs; over half a billion dollars, likely billions in total