2026 Sovereign MoE VRAM & Bare-Metal GPU Repatriation Calculator
Model HBM3e KV-cache residency across DeepSeek-V4.1-Flash, Kimi K2.5, GLM-5.3, and Qwen 3.8 against 24-month NVIDIA HGX B200 / H200 bare-metal payback.
Total HBM3e VRAM
894 GB
Minimum HGX Node
1× 8-GPU HGX H200
CAPEX Break-Even
Month 9.2
HBM3e Residency Allocation (Weights vs. Dynamic KV-Cache vs. 15% CUDA Graph)
Weights: 511G
KV: 295G
15%
24-Mo Cloud API Spend
$684,000
24-Mo Bare-Metal TCO
$311,600
Net 24-Mo Savings
$372,400
Need the Full 8-Bay HGX Rack X-Ray & Boardroom CFO Dossier (PDF)?
Launch the complete interactive workbench on Logic42.ai with instant PDF export.