How much memory does a model actually need?
A model's weights (its "brain") have to sit entirely in GPU memory while it runs — a model can't be partially loaded. The math is simple: parameters (in billions) × 1 GB, plus 20% working room for the temporary scratch space the model uses mid-conversation. Below, watch what happens when that number meets real hardware.
GLM-5.2
Z.ai
❌ If you tried: one NVIDIA H200 SXM 141GB (141 GB)
✅ Minimum that actually fits: 1x NVIDIA HGX H200 8-GPU Server (1,128 GB)
Why this combination: One HGX H200 server already holds 1,128 GB — more than the 893 GB GLM-5.2 needs — so a single server is the smallest thing that actually fits it.
DeepSeek V4 (Pro)
DeepSeek
❌ If you tried: one NVIDIA H200 SXM 141GB (141 GB)
✅ Minimum that actually fits: 2x NVIDIA HGX H200 8-GPU Server (2,256 GB)
Why this combination: A single HGX H200 server (1,128 GB) isn't enough for DeepSeek V4 Pro's 1,920 GB requirement. Two servers clustered together give 2,256 GB — enough room, at the lowest price that qualifies. A GB300 rack (20 TB) also fits, with far more headroom, at a much higher price.
Kimi K2.7 (Code)
Moonshot
❌ If you tried: one NVIDIA H200 SXM 141GB (141 GB)
✅ Minimum that actually fits: 2x NVIDIA HGX H200 8-GPU Server (2,256 GB)
Why this combination: One HGX H200 server (1,128 GB) is just short of Kimi K2.7's 1,200 GB requirement. Two servers clustered together give 2,256 GB, comfortably enough, at the lowest price that qualifies. A GB300 rack also fits, with far more headroom, at a much higher price.