GPUs for on-prem AI / NVIDIA
NVIDIA B200 180GB for LLMs
Blackwell in HGX/DGX B200 8-GPU systems (about 14.3 kW per DGX), adding FP4 at 9 PFLOPS dense per GPU. Per-GPU values derived from NVIDIA 8-GPU totals (1,440GB, 64 TB/s; 36 PF BF16 and 72 PF FP8 sparse, halved). Normally sold as full systems. Price is an indicative 2026 street range for one unit, not a quote.
SXM | NVLink 1800 GB/s | CUDA | indicative price $35K-50K
Which models fit, and how many cards
Planning estimates. 4-bit column: one user at 8K context. FP8 column: 10 users at once at 16K context.
| Model | Size | 4-bit: cards | 4-bit: tok/s | FP8, 10 users: cards | FP8: tok/s per user |
|---|---|---|---|---|---|
| Command A+ (05-2026) | 218B / 25B | 1 | 183 | 2 | 153 |
| Step-3.7-Flash | 198B / 11B | 1 | 308 | 2 | 224 |
| Qwen3.8-Flash-Next | 180B / 6B | 1 | 408 | 2 | 282 |
| Mistral Medium 3.5 128B | 128B | 1 | 60 | 1 | 32 |
| Qwen3.5-122B-A10B | 125B / 10B | 1 | 326 | 1 | 219 |
| Ling-3.0-flash | 124B / 5.1B | 1 | 435 | 1 | 332 |
| NVIDIA Nemotron 3 Super 120B-A12B | 120B / 12B | 1 | 297 | 1 | 203 |
| Mistral Small 4 119B A6B (2603) | 119B / 6.5B | 1 | 396 | 1 | 280 |
| gpt-oss-120b | 117B / 5.1B | 1 | 432 | 1 | 298 |
| Command A (03-2025) | 111B | 1 | 69 | 1 | 38 |
| Llama 4 Scout 17B-16E Instruct | 109B / 17B | 1 | 235 | 1 | 118 |
| Qwen3-Coder-Next | 79.7B / 3B | 1 | 505 | 1 | 381 |
| Llama 3.3 70B Instruct | 70B | 1 | 102 | 1 | 53 |
| Qwen3.6-35B-A3B | 36B / 3B | 1 | 505 | 1 | 389 |
| Qwen3-32B | 32.8B | 1 | 184 | 1 | 94 |
| Olmo 3.1 32B Think | 32B | 1 | 187 | 1 | 104 |
| GLM-4.7-Flash | 31.2B / 3B | 1 | 500 | 1 | 337 |
| Gemma 4 31B IT | 30.7B | 1 | 191 | 1 | 98 |
| Qwen3-30B-A3B-Instruct-2507 | 30.5B / 3.3B | 1 | 482 | 1 | 280 |
| Granite 4.2 30B | 30B | 1 | 195 | 1 | 99 |
| NVIDIA Nemotron 3.5 Lightning 30B-A3B | 30B / 3B | 1 | 508 | 1 | 416 |
| Qwen3.8-27B | 27.8B | 1 | 210 | 1 | 127 |
| Gemma 3 27B IT | 27B | 1 | 211 | 1 | 120 |
| Gemma 4 26B A4B IT | 25.2B / 3.8B | 1 | 466 | 1 | 301 |
| Mistral Small 3.2 24B Instruct 2506 | 24B | 1 | 229 | 1 | 126 |
| Magistral Small 1.2 (2509) | 24B | 1 | 229 | 1 | 126 |
| Devstral Small 2 24B Instruct 2512 | 24B | 1 | 229 | 1 | 126 |
| gpt-oss-20b | 21B / 3.6B | 1 | 482 | 1 | 358 |
| Phi-4 | 14B | 1 | 310 | 1 | 161 |
| Gemma 4 12B IT | 11.95B | 1 | 338 | 1 | 203 |
| Qwen3.5-9B | 9.7B | 1 | 378 | 1 | 263 |
| Qwen3-VL-8B-Instruct | 8.8B | 1 | 386 | 1 | 216 |
| Granite 4.2 8B | 8B | 1 | 399 | 1 | 217 |
| Llama 3.1 8B Instruct | 8B | 1 | 401 | 1 | 232 |
| MiniMax-M3 | 428B / 23B | 2 | 211 | 4 | 161 |
| Qwen3.5-397B-A17B | 403B / 17B | 2 | 239 | 4 | 185 |
| Llama 4 Maverick 17B-128E Instruct | 400B / 17B | 2 | 250 | 4 | 183 |
| Trinity Large Thinking | 398B / 13B | 2 | 256 | 4 | 182 |
| GLM-5.3-Flash | 320B / 18B | 2 | 253 | 2 | 190 |
| MiMo-V2.6-Flash-RL | 309B / 15B | 2 | 265 | 2 | 201 |
| Hy3 | 295B / 21B | 2 | 199 | 4 | 133 |
| MiMo-V2.6-Pro-RL | 1020B / 42B | 4 | 160 | 8 | 131 |
| Kimi-K2.6 | 1000B / 32B | 4 | 185 | 8 | 150 |
| Kimi-K2.7-Code | 1000B / 32B | 4 | 185 | 8 | 150 |
| Muse Spark Open 816B-A42B | 816B / 42B | 4 | 169 | 8 | 138 |
| DeepSeek-V4.1-Flash-1017 | 748B / 16B | 4 | 261 | 8 | 217 |
| GLM-5.3 | 744B / 40B | 4 | 155 | 8 | 124 |
| Mistral Large 3 675B Instruct 2512 | 675B / 41B | 4 | 172 | 8 | 141 |
| DeepSeek-V3.2 | 671B / 37B | 4 | 177 | 8 | 145 |
| DeepSeek-V4-Pro-0813 | 1600B / 49B | 8 | 157 | 16 (multi-node) | 92 |
| Kimi-K3 | 2800B / 104B | 16 (multi-node) | 66 | 24 (multi-node) | 58 |
| Qwen3.8-2.4T-A95B | 2446B / 95B | 16 (multi-node) | 67 | 16 (multi-node) | 62 |
Design a deployment on NVIDIA B200 180GB
Preloaded with this hardware. Add your ERP, documents and users.
Questions
What is the largest model that fits on one NVIDIA B200 180GB?
Of the models we track, Command A+ (05-2026) (218B total, 25B active) fits on a single 180 GB card in 4-bit with one user at 8K context. More users or longer context need memory for the KV cache, so leave headroom.
How fast is NVIDIA B200 180GB for LLM inference?
Generation speed is mostly set by memory bandwidth: 8000 GB/s here. Each token reads the active weights once, so a model with 8 GB of active weights tops out near 1000 tokens per second per stream before efficiency losses; batching many users raises total throughput.
Is NVIDIA B200 180GB a good choice for on-prem AI?
Blackwell in HGX/DGX B200 8-GPU systems (about 14.3 kW per DGX), adding FP4 at 9 PFLOPS dense per GPU. Per-GPU values derived from NVIDIA 8-GPU totals (1,440GB, 64 TB/s; 36 PF BF16 and 72 PF FP8 sparse, halved). Normally sold as full systems. Price is an indicative 2026 street range for one unit, not a quote.