Economics of Deploying GLM-5.2 on 8×B200
qubridInc · reddit · 2026-07-08
The Qubrid team shared a deployment configuration analysis for GLM-5.2 (approx. 750B total params/40B activated MoE, 256 experts top-8, DSA+MLA attention, 1M context, MIT license) on 8×B200 nodes. Key findings: MoE decoding is bandwidth-bound rather than compute-bound at medium concurrency, meaning B200 offers only about 1.2x performance per dollar compared to H200 at FP8 (matching the HBM bandwidth ratio, not the 2.3x compute ratio). The real game-changer is NVFP4 (halving weight bytes, which Hopper lacks FP4 tensor cores for). The recommended configuration is NVFP4 + two TP=4 replicas, achieving roughly 2x throughput compared to TP=8. The author also noted the current lack of a comprehensive concurrency sweep table for GLM-5.2 on a documented 8×B200 setup.
More from Infra
- Anthropic S-1 chart fuels positive-ROI narrative on AI infrastructure buildout — BenBajarin · 2026-10-26
- Nearly half of Claude revenue now flows through cloud marketplace commitments, up from 11% — BenBajarin · 2026-10-26
- Anthropic ARR goes vertical after price cuts: tokens up 4x, $100B+ ARR back in play — GavinSBaker · 2026-10-26
- Alex Banks argues power caps GPU supply, real risk is demand cooling first — thealexbanks · 2026-10-26
- Analyst: Ant has the most fungible AI workload of any model provider today — BenBajarin · 2026-10-26
- Anthropic S-1: training infra edge cut per-task cost 40% and doubled revenue growth — BenBajarin · 2026-10-26