Hardware shortlist

Compare GPU Servers vs AI Workstations for Your Workload

Choose two to four systems. Compare price, GPU memory, installation needs and the current models that fit each system.

Build a shortlist How systems are tested

Three-quarter supplier render of a 4U OEM multi-GPU rack server — OEM platform reference render showing the rack-server form and external service access.

11-system range

Build a two-to-four system shortlist.

Compare price, GPU configuration, memory, installation requirements and every current catalogue model against each system's available GPU memory.

Default four-system shortlist. Use the controls above to compare another two to four systems.
Decision field	SME workstation Team 32	SME workstation Company 64	SME workstation Studio 96	SME workstation Studio 192	Enterprise PCIe rack Value Rack 128	Enterprise PCIe rack Value Rack 256	Enterprise PCIe rack Enterprise 384	Enterprise PCIe rack Enterprise 768	Enterprise PCIe rack H200 1.1TB	Frontier platform Frontier Native 2.3TB	Frontier platform Frontier Rack 20TB
Price	£7,500 Guide price · complete configured system Price details Complete GPU Servers private AI deployment for the listed configuration and service scope.	£18,500 Guide price · complete configured system Price details Complete GPU Servers private AI deployment for the listed configuration and service scope.	£20,000 Guide price · complete configured system Price details Complete GPU Servers private AI deployment for the listed configuration and service scope.	£42,000 Guide price · complete configured system Price details Complete GPU Servers private AI deployment for the listed configuration and service scope.	£45,765 Guide price · complete configured system Price details Complete GPU Servers private AI deployment for the listed configuration and service scope.	£82,675 Guide price · complete configured system Price details Complete GPU Servers private AI deployment for the listed configuration and service scope.	£99,999 Guide price · complete configured system Price details Complete GPU Servers private AI deployment for the listed configuration and service scope.	£189,999 Guide price · complete configured system Price details Complete GPU Servers private AI deployment for the listed configuration and service scope.	£349,000 Guide price · complete configured system Price details Complete GPU Servers private AI deployment for the listed configuration and service scope.	Pricing on request Configured to order Price details Configured and priced around the selected workload, facility, delivery and support needs.	Pricing on request Configured to order Price details Configured and priced around the selected workload, facility, delivery and support needs.
GPU configuration	1 × NVIDIA GeForce RTX 5090 32GB 1 physical GPU; 32GB per GPU	2 × NVIDIA GeForce RTX 5090 32GB 2 physical GPUs; 32GB per GPU	1 × NVIDIA RTX PRO 6000 Blackwell 96GB 1 physical GPU; 96GB per GPU	2 × NVIDIA RTX PRO 6000 Blackwell 96GB 2 physical GPUs; 96GB per GPU	4 × NVIDIA RTX 5090 AI 32GB 4 physical GPUs; 32GB per GPU	8 × NVIDIA RTX 5090 AI 32GB 8 physical GPUs; 32GB per GPU	4 × NVIDIA RTX PRO 6000 Blackwell 96GB 4 physical GPUs; 96GB per GPU	8 × NVIDIA RTX PRO 6000 Blackwell 96GB 8 physical GPUs; 96GB per GPU	8 × NVIDIA H200 NVL 141GB 8 physical GPUs; 141GB per GPU	8 × NVIDIA B300 288GB 8 physical GPUs; 288GB per GPU	72 × NVIDIA GB300 288GB 72 physical GPUs; 288GB per GPU
Aggregate GPU memory Physical total, not a universal pooled or contiguous memory space.	32GB 32GB is available on one GPU as one physical memory space.	64GB 64GB is the physical total across 2 separate 32GB GPUs. It is not automatically pooled; supported software may divide a model or workload across them, subject to topology and testing.	96GB 96GB is available on one GPU as one physical memory space.	192GB 192GB is the physical total across 2 separate 96GB GPUs. It is not automatically pooled; supported software may divide a model or workload across them, subject to topology and testing.	128GB 128GB is the physical total across 4 separate 32GB GPUs. It is not automatically pooled; supported software may divide a model or workload across them, subject to topology and testing.	256GB 256GB is the physical total across 8 separate 32GB GPUs. It is not automatically pooled; supported software may divide a model or workload across them, subject to topology and testing.	384GB 384GB is the physical total across 4 separate 96GB GPUs. It is not automatically pooled; supported software may divide a model or workload across them, subject to topology and testing.	768GB 768GB is the physical total across 8 separate 96GB GPUs. It is not automatically pooled; supported software may divide a model or workload across them, subject to topology and testing.	1,128GB (1.1TB) 1,128GB is the physical total across eight H200 GPUs. It is not automatically one pooled memory space; supported parallel operation depends on the selected topology and runtime.	2,304GB (2.25TB) 2,304GB is distributed across eight B300 GPUs connected by HGX NVLink and NVSwitch. The fabric enables high-bandwidth multi-GPU work, but software support and the exact workload still determine usable capacity.	20,736GB (20.25TB) 20,736GB is distributed across a rack-scale NVLink and NVSwitch fabric. It is specialist shared infrastructure, not one ordinary GPU memory space.
Form and facility	Workstation A ventilated floor or desk-side position with clear intake and exhaust paths A suitable UK circuit and an agreed UPS decision after measured-load review An acoustic and heat check under the accepted workload A named owner for accounts, updates, backups and incident response	Workstation A ventilated floor or desk-side position with clear intake and exhaust paths A suitable UK circuit and an agreed UPS decision after measured-load review An acoustic and heat check under the accepted workload A named owner for accounts, updates, backups and incident response	Workstation A ventilated floor or desk-side position with clear intake and exhaust paths A suitable UK circuit and an agreed UPS decision after measured-load review An acoustic and heat check under the accepted workload A named owner for accounts, updates, backups and incident response	Workstation A ventilated floor or desk-side position with clear intake and exhaust paths A suitable UK circuit and an agreed UPS decision after measured-load review An acoustic and heat check under the accepted workload A named owner for accounts, updates, backups and incident response	Rack server A suitable 19-inch rack, rail depth, handling route and secure operating location A qualified electrical design based on the final PSU population and measured load Cooling and heat-rejection capacity for sustained accelerator operation Appropriate switching, cabling, remote management and network segmentation A named operational owner or contracted support route	Rack server A suitable 19-inch rack, rail depth, handling route and secure operating location A qualified electrical design based on the final PSU population and measured load Cooling and heat-rejection capacity for sustained accelerator operation Appropriate switching, cabling, remote management and network segmentation A named operational owner or contracted support route	Rack server A suitable 19-inch rack, rail depth, handling route and secure operating location A qualified electrical design based on the final PSU population and measured load Cooling and heat-rejection capacity for sustained accelerator operation Appropriate switching, cabling, remote management and network segmentation A named operational owner or contracted support route	Rack server A suitable 19-inch rack, rail depth, handling route and secure operating location A qualified electrical design based on the final PSU population and measured load Cooling and heat-rejection capacity for sustained accelerator operation Appropriate switching, cabling, remote management and network segmentation A named operational owner or contracted support route	Rack server A suitable 19-inch rack, rail depth, handling route and secure operating location A qualified electrical design based on the final PSU population and measured load Cooling and heat-rejection capacity for sustained accelerator operation Appropriate switching, cabling, remote management and network segmentation A named operational owner or contracted support route	Rack server A suitable 19-inch rack, rail depth, handling route and secure operating location A qualified electrical design based on the final PSU population and measured load Cooling and heat-rejection capacity for sustained accelerator operation Appropriate switching, cabling, remote management and network segmentation A named operational owner or contracted support route	Rack server A suitable 19-inch rack, rail depth, handling route and secure operating location A qualified electrical design based on the final PSU population and measured load Cooling and heat-rejection capacity for sustained accelerator operation Appropriate switching, cabling, remote management and network segmentation A named operational owner or contracted support route
Platform	Single-GPU mid-tower AI workstation	Dual-GPU full-tower AI workstation	Single-GPU professional AI workstation	Dual-GPU professional AI workstation	GENOAX2 PCIe Gen5 4U rack platform	GENOAX2 PCIe Gen5 4U rack platform	GENOAX2 PCIe Gen5 4U enterprise rack platform	GENOAX2 PCIe Gen5 4U enterprise rack platform	Eight-GPU H200 NVL PCIe rack platform	SYS-822GS-NB3RT 8U HGX B300 platform	GB300 NVL72 full-rack architecture
System RAM	64GB	128GB	128GB	128GB	Specified to order	Specified to order	Specified to order	Specified to order	1024GB	Specified to order	Specified to order
Storage	2TB NVMe	2TB NVMe	2TB NVMe	2TB NVMe	Specified to order	Specified to order	Specified to order	Specified to order	11.68TB NVMe	Specified to order	Specified to order
Network	Specified to order	10GbE and 2.5GbE Ethernet	Specified to order	10GbE and 2.5GbE Ethernet	Specified to order	Specified to order	Specified to order	Specified to order	200Gb ConnectX-6 plus management/base networking	Specified to order	Specified to order
Power	Specified for the ordered configuration and intended facility.	Specified for the ordered configuration and intended facility.	Specified for the ordered configuration and intended facility.	Specified for the ordered configuration and intended facility.	Specified for the ordered configuration and intended facility.	Specified for the ordered configuration and intended facility.	Specified for the ordered configuration and intended facility.	Specified for the ordered configuration and intended facility.	Specified for the ordered configuration and intended facility.	Specified for the ordered configuration and intended facility.	Up to 142kW full-rack power
Model compatibility GPU-memory fit only. Final speed, context and concurrent users depend on the selected model and software.	18 models fit without splitting the model across GPUs. DeepSeek-OCR 2 12GB GPU planning floor Qwen3-Embedding 8B 20GB GPU planning floor BGE-M3 4GB GPU planning floor, or CPU for light use Qwen3-ASR 1.7B 8GB GPU planning floor View 14 more models Whisper Large V3 6GB GPU planning floor for an optimised inference runtime Voxtral Mini 4B Realtime 24GB GPU planning floor FLUX.2 Klein 4B About 13GB VRAM Wan2.2 TI2V 5B 24GB VRAM with documented CPU offload settings HunyuanVideo 1.5 14GB VRAM with model offloading Hunyuan3D 2.1 24GB single-GPU planning floor Llama Guard 4 28GB GPU planning floor Stable Diffusion 3.5 Large 24GB planning floor with an optimised or offload workflow GPT-OSS 20B 16GB on one GPU Qwen3.5 4B 16GB on one GPU Gemma 4 12B 32GB on one GPU Qwen3-TTS 1.7B 8GB on one GPU Qwen3-Reranker 8B 24GB on one GPU Step-Audio 2 Mini 24GB on one GPU Open more model options 15 additional models Recommended memory route Studio 96 7 models on this route Qwen3.6 35B-A3B 80GB-class GPU for the named BF16 weights at reduced context GPT-OSS 120B One 80GB GPU Gemma 3 27B 64GB GPU for BF16 at a bounded context Qwen-Image 2512 64GB GPU as a tight full-weight floor Qwen3.5 27B 64GB on one GPU Gemma 4 31B 64GB on one GPU FLUX.2 Klein 9B 64GB on one GPU Explore system Compare pair Recommended memory route Frontier Native 2.3TB 5 models on this route DeepSeek V4 Flash At least 176GB aggregate GPU memory with a supported sharding route Mistral Small 4 256GB aggregate with supported model parallelism Llama 4 Scout At least 240GB aggregate for BF16 weights and minimal overhead Qwen3-Coder-Next 160GB aggregate across at least 2 GPUs MiniMax M2.5 256GB aggregate across at least 2 GPUs Explore system Compare pair Multi-GPU opportunity Enterprise 768 1 model on this route Step 3.7 Flash 416GB aggregate across at least 4 GPUs Software support and GPU layout are confirmed during sizing. Explore system Compare pair Multi-GPU opportunity Frontier Native 2.3TB 2 models on this route Kimi K3 At least 1.561TB aggregate GPU memory for the released weight files GLM-5 1536GB aggregate across at least 8 GPUs Software support and GPU layout are confirmed during sizing. Explore system Compare pair See full compatibility for Team 32	18 models fit without splitting the model across GPUs. DeepSeek-OCR 2 12GB GPU planning floor Qwen3-Embedding 8B 20GB GPU planning floor BGE-M3 4GB GPU planning floor, or CPU for light use Qwen3-ASR 1.7B 8GB GPU planning floor View 14 more models Whisper Large V3 6GB GPU planning floor for an optimised inference runtime Voxtral Mini 4B Realtime 24GB GPU planning floor FLUX.2 Klein 4B About 13GB VRAM Wan2.2 TI2V 5B 24GB VRAM with documented CPU offload settings HunyuanVideo 1.5 14GB VRAM with model offloading Hunyuan3D 2.1 24GB single-GPU planning floor Llama Guard 4 28GB GPU planning floor Stable Diffusion 3.5 Large 24GB planning floor with an optimised or offload workflow GPT-OSS 20B 16GB on one GPU Qwen3.5 4B 16GB on one GPU Gemma 4 12B 32GB on one GPU Qwen3-TTS 1.7B 8GB on one GPU Qwen3-Reranker 8B 24GB on one GPU Step-Audio 2 Mini 24GB on one GPU Open more model options 15 additional models Recommended memory route Studio 96 7 models on this route Qwen3.6 35B-A3B 80GB-class GPU for the named BF16 weights at reduced context GPT-OSS 120B One 80GB GPU Gemma 3 27B 64GB GPU for BF16 at a bounded context Qwen-Image 2512 64GB GPU as a tight full-weight floor Qwen3.5 27B 64GB on one GPU Gemma 4 31B 64GB on one GPU FLUX.2 Klein 9B 64GB on one GPU Explore system Compare pair Recommended memory route Frontier Native 2.3TB 5 models on this route DeepSeek V4 Flash At least 176GB aggregate GPU memory with a supported sharding route Mistral Small 4 256GB aggregate with supported model parallelism Llama 4 Scout At least 240GB aggregate for BF16 weights and minimal overhead Qwen3-Coder-Next 160GB aggregate across at least 2 GPUs MiniMax M2.5 256GB aggregate across at least 2 GPUs Explore system Compare pair Multi-GPU opportunity Enterprise 768 1 model on this route Step 3.7 Flash 416GB aggregate across at least 4 GPUs Software support and GPU layout are confirmed during sizing. Explore system Compare pair Multi-GPU opportunity Frontier Native 2.3TB 2 models on this route Kimi K3 At least 1.561TB aggregate GPU memory for the released weight files GLM-5 1536GB aggregate across at least 8 GPUs Software support and GPU layout are confirmed during sizing. Explore system Compare pair See full compatibility for Company 64	25 models fit without splitting the model across GPUs. Qwen3.6 35B-A3B 80GB-class GPU for the named BF16 weights at reduced context GPT-OSS 120B One 80GB GPU Gemma 3 27B 64GB GPU for BF16 at a bounded context DeepSeek-OCR 2 12GB GPU planning floor View 21 more models Qwen3-Embedding 8B 20GB GPU planning floor BGE-M3 4GB GPU planning floor, or CPU for light use Qwen3-ASR 1.7B 8GB GPU planning floor Whisper Large V3 6GB GPU planning floor for an optimised inference runtime Voxtral Mini 4B Realtime 24GB GPU planning floor FLUX.2 Klein 4B About 13GB VRAM Qwen-Image 2512 64GB GPU as a tight full-weight floor Wan2.2 TI2V 5B 24GB VRAM with documented CPU offload settings HunyuanVideo 1.5 14GB VRAM with model offloading Hunyuan3D 2.1 24GB single-GPU planning floor Llama Guard 4 28GB GPU planning floor Stable Diffusion 3.5 Large 24GB planning floor with an optimised or offload workflow GPT-OSS 20B 16GB on one GPU Qwen3.5 4B 16GB on one GPU Qwen3.5 27B 64GB on one GPU Gemma 4 12B 32GB on one GPU Gemma 4 31B 64GB on one GPU Qwen3-TTS 1.7B 8GB on one GPU Qwen3-Reranker 8B 24GB on one GPU FLUX.2 Klein 9B 64GB on one GPU Step-Audio 2 Mini 24GB on one GPU Open more model options 8 additional models Recommended memory route Frontier Native 2.3TB 5 models on this route DeepSeek V4 Flash At least 176GB aggregate GPU memory with a supported sharding route Mistral Small 4 256GB aggregate with supported model parallelism Llama 4 Scout At least 240GB aggregate for BF16 weights and minimal overhead Qwen3-Coder-Next 160GB aggregate across at least 2 GPUs MiniMax M2.5 256GB aggregate across at least 2 GPUs Explore system Compare pair Multi-GPU opportunity Enterprise 768 1 model on this route Step 3.7 Flash 416GB aggregate across at least 4 GPUs Software support and GPU layout are confirmed during sizing. Explore system Compare pair Multi-GPU opportunity Frontier Native 2.3TB 2 models on this route Kimi K3 At least 1.561TB aggregate GPU memory for the released weight files GLM-5 1536GB aggregate across at least 8 GPUs Software support and GPU layout are confirmed during sizing. Explore system Compare pair See full compatibility for Studio 96	25 models fit without model splitting; 2 more may fit with supported multi-GPU loading. Qwen3.6 35B-A3B 80GB-class GPU for the named BF16 weights at reduced context GPT-OSS 120B One 80GB GPU Gemma 3 27B 64GB GPU for BF16 at a bounded context DeepSeek-OCR 2 12GB GPU planning floor View 21 more models Qwen3-Embedding 8B 20GB GPU planning floor BGE-M3 4GB GPU planning floor, or CPU for light use Qwen3-ASR 1.7B 8GB GPU planning floor Whisper Large V3 6GB GPU planning floor for an optimised inference runtime Voxtral Mini 4B Realtime 24GB GPU planning floor FLUX.2 Klein 4B About 13GB VRAM Qwen-Image 2512 64GB GPU as a tight full-weight floor Wan2.2 TI2V 5B 24GB VRAM with documented CPU offload settings HunyuanVideo 1.5 14GB VRAM with model offloading Hunyuan3D 2.1 24GB single-GPU planning floor Llama Guard 4 28GB GPU planning floor Stable Diffusion 3.5 Large 24GB planning floor with an optimised or offload workflow GPT-OSS 20B 16GB on one GPU Qwen3.5 4B 16GB on one GPU Qwen3.5 27B 64GB on one GPU Gemma 4 12B 32GB on one GPU Gemma 4 31B 64GB on one GPU Qwen3-TTS 1.7B 8GB on one GPU Qwen3-Reranker 8B 24GB on one GPU FLUX.2 Klein 9B 64GB on one GPU Step-Audio 2 Mini 24GB on one GPU 2 models may fit with multi-GPU loading DeepSeek V4 Flash At least 176GB aggregate GPU memory with a supported sharding route Qwen3-Coder-Next 160GB aggregate across at least 2 GPUs Open more model options 6 additional models Recommended memory route Frontier Native 2.3TB 3 models on this route Mistral Small 4 256GB aggregate with supported model parallelism Llama 4 Scout At least 240GB aggregate for BF16 weights and minimal overhead MiniMax M2.5 256GB aggregate across at least 2 GPUs Explore system Compare pair Multi-GPU opportunity Enterprise 768 1 model on this route Step 3.7 Flash 416GB aggregate across at least 4 GPUs Software support and GPU layout are confirmed during sizing. Explore system Compare pair Multi-GPU opportunity Frontier Native 2.3TB 2 models on this route Kimi K3 At least 1.561TB aggregate GPU memory for the released weight files GLM-5 1536GB aggregate across at least 8 GPUs Software support and GPU layout are confirmed during sizing. Explore system Compare pair See full compatibility for Studio 192	18 models fit without splitting the model across GPUs. DeepSeek-OCR 2 12GB GPU planning floor Qwen3-Embedding 8B 20GB GPU planning floor BGE-M3 4GB GPU planning floor, or CPU for light use Qwen3-ASR 1.7B 8GB GPU planning floor View 14 more models Whisper Large V3 6GB GPU planning floor for an optimised inference runtime Voxtral Mini 4B Realtime 24GB GPU planning floor FLUX.2 Klein 4B About 13GB VRAM Wan2.2 TI2V 5B 24GB VRAM with documented CPU offload settings HunyuanVideo 1.5 14GB VRAM with model offloading Hunyuan3D 2.1 24GB single-GPU planning floor Llama Guard 4 28GB GPU planning floor Stable Diffusion 3.5 Large 24GB planning floor with an optimised or offload workflow GPT-OSS 20B 16GB on one GPU Qwen3.5 4B 16GB on one GPU Gemma 4 12B 32GB on one GPU Qwen3-TTS 1.7B 8GB on one GPU Qwen3-Reranker 8B 24GB on one GPU Step-Audio 2 Mini 24GB on one GPU Open more model options 15 additional models Recommended memory route Enterprise 384 7 models on this route Qwen3.6 35B-A3B 80GB-class GPU for the named BF16 weights at reduced context GPT-OSS 120B One 80GB GPU Gemma 3 27B 64GB GPU for BF16 at a bounded context Qwen-Image 2512 64GB GPU as a tight full-weight floor Qwen3.5 27B 64GB on one GPU Gemma 4 31B 64GB on one GPU FLUX.2 Klein 9B 64GB on one GPU Explore system Compare pair Recommended memory route Frontier Native 2.3TB 5 models on this route DeepSeek V4 Flash At least 176GB aggregate GPU memory with a supported sharding route Mistral Small 4 256GB aggregate with supported model parallelism Llama 4 Scout At least 240GB aggregate for BF16 weights and minimal overhead Qwen3-Coder-Next 160GB aggregate across at least 2 GPUs MiniMax M2.5 256GB aggregate across at least 2 GPUs Explore system Compare pair Multi-GPU opportunity Enterprise 768 1 model on this route Step 3.7 Flash 416GB aggregate across at least 4 GPUs Software support and GPU layout are confirmed during sizing. Explore system Compare pair Multi-GPU opportunity Frontier Native 2.3TB 2 models on this route Kimi K3 At least 1.561TB aggregate GPU memory for the released weight files GLM-5 1536GB aggregate across at least 8 GPUs Software support and GPU layout are confirmed during sizing. Explore system Compare pair See full compatibility for Value Rack 128	18 models fit without model splitting; 5 more may fit with supported multi-GPU loading. DeepSeek-OCR 2 12GB GPU planning floor Qwen3-Embedding 8B 20GB GPU planning floor BGE-M3 4GB GPU planning floor, or CPU for light use Qwen3-ASR 1.7B 8GB GPU planning floor View 14 more models Whisper Large V3 6GB GPU planning floor for an optimised inference runtime Voxtral Mini 4B Realtime 24GB GPU planning floor FLUX.2 Klein 4B About 13GB VRAM Wan2.2 TI2V 5B 24GB VRAM with documented CPU offload settings HunyuanVideo 1.5 14GB VRAM with model offloading Hunyuan3D 2.1 24GB single-GPU planning floor Llama Guard 4 28GB GPU planning floor Stable Diffusion 3.5 Large 24GB planning floor with an optimised or offload workflow GPT-OSS 20B 16GB on one GPU Qwen3.5 4B 16GB on one GPU Gemma 4 12B 32GB on one GPU Qwen3-TTS 1.7B 8GB on one GPU Qwen3-Reranker 8B 24GB on one GPU Step-Audio 2 Mini 24GB on one GPU 5 models may fit with multi-GPU loading DeepSeek V4 Flash At least 176GB aggregate GPU memory with a supported sharding route Mistral Small 4 256GB aggregate with supported model parallelism Llama 4 Scout At least 240GB aggregate for BF16 weights and minimal overhead Qwen3-Coder-Next 160GB aggregate across at least 2 GPUs MiniMax M2.5 256GB aggregate across at least 2 GPUs Open more model options 10 additional models Recommended memory route Enterprise 384 7 models on this route Qwen3.6 35B-A3B 80GB-class GPU for the named BF16 weights at reduced context GPT-OSS 120B One 80GB GPU Gemma 3 27B 64GB GPU for BF16 at a bounded context Qwen-Image 2512 64GB GPU as a tight full-weight floor Qwen3.5 27B 64GB on one GPU Gemma 4 31B 64GB on one GPU FLUX.2 Klein 9B 64GB on one GPU Explore system Compare pair Multi-GPU opportunity Enterprise 768 1 model on this route Step 3.7 Flash 416GB aggregate across at least 4 GPUs Software support and GPU layout are confirmed during sizing. Explore system Compare pair Multi-GPU opportunity Frontier Native 2.3TB 2 models on this route Kimi K3 At least 1.561TB aggregate GPU memory for the released weight files GLM-5 1536GB aggregate across at least 8 GPUs Software support and GPU layout are confirmed during sizing. Explore system Compare pair See full compatibility for Value Rack 256	25 models fit without model splitting; 5 more may fit with supported multi-GPU loading. Qwen3.6 35B-A3B 80GB-class GPU for the named BF16 weights at reduced context GPT-OSS 120B One 80GB GPU Gemma 3 27B 64GB GPU for BF16 at a bounded context DeepSeek-OCR 2 12GB GPU planning floor View 21 more models Qwen3-Embedding 8B 20GB GPU planning floor BGE-M3 4GB GPU planning floor, or CPU for light use Qwen3-ASR 1.7B 8GB GPU planning floor Whisper Large V3 6GB GPU planning floor for an optimised inference runtime Voxtral Mini 4B Realtime 24GB GPU planning floor FLUX.2 Klein 4B About 13GB VRAM Qwen-Image 2512 64GB GPU as a tight full-weight floor Wan2.2 TI2V 5B 24GB VRAM with documented CPU offload settings HunyuanVideo 1.5 14GB VRAM with model offloading Hunyuan3D 2.1 24GB single-GPU planning floor Llama Guard 4 28GB GPU planning floor Stable Diffusion 3.5 Large 24GB planning floor with an optimised or offload workflow GPT-OSS 20B 16GB on one GPU Qwen3.5 4B 16GB on one GPU Qwen3.5 27B 64GB on one GPU Gemma 4 12B 32GB on one GPU Gemma 4 31B 64GB on one GPU Qwen3-TTS 1.7B 8GB on one GPU Qwen3-Reranker 8B 24GB on one GPU FLUX.2 Klein 9B 64GB on one GPU Step-Audio 2 Mini 24GB on one GPU 5 models may fit with multi-GPU loading DeepSeek V4 Flash At least 176GB aggregate GPU memory with a supported sharding route Mistral Small 4 256GB aggregate with supported model parallelism Llama 4 Scout At least 240GB aggregate for BF16 weights and minimal overhead Qwen3-Coder-Next 160GB aggregate across at least 2 GPUs MiniMax M2.5 256GB aggregate across at least 2 GPUs Open more model options 3 additional models Multi-GPU opportunity Enterprise 768 1 model on this route Step 3.7 Flash 416GB aggregate across at least 4 GPUs Software support and GPU layout are confirmed during sizing. Explore system Compare pair Multi-GPU opportunity Frontier Native 2.3TB 2 models on this route Kimi K3 At least 1.561TB aggregate GPU memory for the released weight files GLM-5 1536GB aggregate across at least 8 GPUs Software support and GPU layout are confirmed during sizing. Explore system Compare pair See full compatibility for Enterprise 384	25 models fit without model splitting; 6 more may fit with supported multi-GPU loading. Qwen3.6 35B-A3B 80GB-class GPU for the named BF16 weights at reduced context GPT-OSS 120B One 80GB GPU Gemma 3 27B 64GB GPU for BF16 at a bounded context DeepSeek-OCR 2 12GB GPU planning floor View 21 more models Qwen3-Embedding 8B 20GB GPU planning floor BGE-M3 4GB GPU planning floor, or CPU for light use Qwen3-ASR 1.7B 8GB GPU planning floor Whisper Large V3 6GB GPU planning floor for an optimised inference runtime Voxtral Mini 4B Realtime 24GB GPU planning floor FLUX.2 Klein 4B About 13GB VRAM Qwen-Image 2512 64GB GPU as a tight full-weight floor Wan2.2 TI2V 5B 24GB VRAM with documented CPU offload settings HunyuanVideo 1.5 14GB VRAM with model offloading Hunyuan3D 2.1 24GB single-GPU planning floor Llama Guard 4 28GB GPU planning floor Stable Diffusion 3.5 Large 24GB planning floor with an optimised or offload workflow GPT-OSS 20B 16GB on one GPU Qwen3.5 4B 16GB on one GPU Qwen3.5 27B 64GB on one GPU Gemma 4 12B 32GB on one GPU Gemma 4 31B 64GB on one GPU Qwen3-TTS 1.7B 8GB on one GPU Qwen3-Reranker 8B 24GB on one GPU FLUX.2 Klein 9B 64GB on one GPU Step-Audio 2 Mini 24GB on one GPU 6 models may fit with multi-GPU loading DeepSeek V4 Flash At least 176GB aggregate GPU memory with a supported sharding route Mistral Small 4 256GB aggregate with supported model parallelism Llama 4 Scout At least 240GB aggregate for BF16 weights and minimal overhead Qwen3-Coder-Next 160GB aggregate across at least 2 GPUs Step 3.7 Flash 416GB aggregate across at least 4 GPUs MiniMax M2.5 256GB aggregate across at least 2 GPUs Open more model options 2 additional models Multi-GPU opportunity Frontier Native 2.3TB 2 models on this route Kimi K3 At least 1.561TB aggregate GPU memory for the released weight files GLM-5 1536GB aggregate across at least 8 GPUs Software support and GPU layout are confirmed during sizing. Explore system Compare pair See full compatibility for Enterprise 768	25 models fit without model splitting; 6 more may fit with supported multi-GPU loading. Qwen3.6 35B-A3B 80GB-class GPU for the named BF16 weights at reduced context GPT-OSS 120B One 80GB GPU Gemma 3 27B 64GB GPU for BF16 at a bounded context DeepSeek-OCR 2 12GB GPU planning floor View 21 more models Qwen3-Embedding 8B 20GB GPU planning floor BGE-M3 4GB GPU planning floor, or CPU for light use Qwen3-ASR 1.7B 8GB GPU planning floor Whisper Large V3 6GB GPU planning floor for an optimised inference runtime Voxtral Mini 4B Realtime 24GB GPU planning floor FLUX.2 Klein 4B About 13GB VRAM Qwen-Image 2512 64GB GPU as a tight full-weight floor Wan2.2 TI2V 5B 24GB VRAM with documented CPU offload settings HunyuanVideo 1.5 14GB VRAM with model offloading Hunyuan3D 2.1 24GB single-GPU planning floor Llama Guard 4 28GB GPU planning floor Stable Diffusion 3.5 Large 24GB planning floor with an optimised or offload workflow GPT-OSS 20B 16GB on one GPU Qwen3.5 4B 16GB on one GPU Qwen3.5 27B 64GB on one GPU Gemma 4 12B 32GB on one GPU Gemma 4 31B 64GB on one GPU Qwen3-TTS 1.7B 8GB on one GPU Qwen3-Reranker 8B 24GB on one GPU FLUX.2 Klein 9B 64GB on one GPU Step-Audio 2 Mini 24GB on one GPU 6 models may fit with multi-GPU loading DeepSeek V4 Flash At least 176GB aggregate GPU memory with a supported sharding route Mistral Small 4 256GB aggregate with supported model parallelism Llama 4 Scout At least 240GB aggregate for BF16 weights and minimal overhead Qwen3-Coder-Next 160GB aggregate across at least 2 GPUs Step 3.7 Flash 416GB aggregate across at least 4 GPUs MiniMax M2.5 256GB aggregate across at least 2 GPUs Open more model options 2 additional models Multi-GPU opportunity Frontier Native 2.3TB 2 models on this route Kimi K3 At least 1.561TB aggregate GPU memory for the released weight files GLM-5 1536GB aggregate across at least 8 GPUs Software support and GPU layout are confirmed during sizing. Explore system Compare pair See full compatibility for H200 1.1TB	30 models fit without model splitting; 3 more may fit with supported multi-GPU loading. Qwen3.6 35B-A3B 80GB-class GPU for the named BF16 weights at reduced context DeepSeek V4 Flash At least 176GB aggregate GPU memory with a supported sharding route Mistral Small 4 256GB aggregate with supported model parallelism Llama 4 Scout At least 240GB aggregate for BF16 weights and minimal overhead View 26 more models GPT-OSS 120B One 80GB GPU Gemma 3 27B 64GB GPU for BF16 at a bounded context DeepSeek-OCR 2 12GB GPU planning floor Qwen3-Embedding 8B 20GB GPU planning floor BGE-M3 4GB GPU planning floor, or CPU for light use Qwen3-ASR 1.7B 8GB GPU planning floor Whisper Large V3 6GB GPU planning floor for an optimised inference runtime Voxtral Mini 4B Realtime 24GB GPU planning floor FLUX.2 Klein 4B About 13GB VRAM Qwen-Image 2512 64GB GPU as a tight full-weight floor Wan2.2 TI2V 5B 24GB VRAM with documented CPU offload settings HunyuanVideo 1.5 14GB VRAM with model offloading Hunyuan3D 2.1 24GB single-GPU planning floor Llama Guard 4 28GB GPU planning floor Stable Diffusion 3.5 Large 24GB planning floor with an optimised or offload workflow GPT-OSS 20B 16GB on one GPU Qwen3-Coder-Next 160GB aggregate across at least 2 GPUs Qwen3.5 4B 16GB on one GPU Qwen3.5 27B 64GB on one GPU Gemma 4 12B 32GB on one GPU Gemma 4 31B 64GB on one GPU MiniMax M2.5 256GB aggregate across at least 2 GPUs Qwen3-TTS 1.7B 8GB on one GPU Qwen3-Reranker 8B 24GB on one GPU FLUX.2 Klein 9B 64GB on one GPU Step-Audio 2 Mini 24GB on one GPU 3 models may fit with multi-GPU loading Kimi K3 At least 1.561TB aggregate GPU memory for the released weight files Step 3.7 Flash 416GB aggregate across at least 4 GPUs GLM-5 1536GB aggregate across at least 8 GPUs See full compatibility for Frontier Native 2.3TB	30 models fit without model splitting; 3 more may fit with supported multi-GPU loading. Qwen3.6 35B-A3B 80GB-class GPU for the named BF16 weights at reduced context DeepSeek V4 Flash At least 176GB aggregate GPU memory with a supported sharding route Mistral Small 4 256GB aggregate with supported model parallelism Llama 4 Scout At least 240GB aggregate for BF16 weights and minimal overhead View 26 more models GPT-OSS 120B One 80GB GPU Gemma 3 27B 64GB GPU for BF16 at a bounded context DeepSeek-OCR 2 12GB GPU planning floor Qwen3-Embedding 8B 20GB GPU planning floor BGE-M3 4GB GPU planning floor, or CPU for light use Qwen3-ASR 1.7B 8GB GPU planning floor Whisper Large V3 6GB GPU planning floor for an optimised inference runtime Voxtral Mini 4B Realtime 24GB GPU planning floor FLUX.2 Klein 4B About 13GB VRAM Qwen-Image 2512 64GB GPU as a tight full-weight floor Wan2.2 TI2V 5B 24GB VRAM with documented CPU offload settings HunyuanVideo 1.5 14GB VRAM with model offloading Hunyuan3D 2.1 24GB single-GPU planning floor Llama Guard 4 28GB GPU planning floor Stable Diffusion 3.5 Large 24GB planning floor with an optimised or offload workflow GPT-OSS 20B 16GB on one GPU Qwen3-Coder-Next 160GB aggregate across at least 2 GPUs Qwen3.5 4B 16GB on one GPU Qwen3.5 27B 64GB on one GPU Gemma 4 12B 32GB on one GPU Gemma 4 31B 64GB on one GPU MiniMax M2.5 256GB aggregate across at least 2 GPUs Qwen3-TTS 1.7B 8GB on one GPU Qwen3-Reranker 8B 24GB on one GPU FLUX.2 Klein 9B 64GB on one GPU Step-Audio 2 Mini 24GB on one GPU 3 models may fit with multi-GPU loading Kimi K3 At least 1.561TB aggregate GPU memory for the released weight files Step 3.7 Flash 416GB aggregate across at least 4 GPUs GLM-5 1536GB aggregate across at least 8 GPUs See full compatibility for Frontier Rack 20TB
Known limitations	8 things to consider The accepted model or context needs more than 32GB Several heavy services need guaranteed simultaneous capacity The buyer needs redundant power or enterprise remote management 32GB is one GPU's physical memory; context, cache, batching and concurrency reduce the usable model envelope. No tokens-per-second, user-count, latency, image or video performance has been measured for this revision. A workstation is not presented as a redundant datacentre appliance. Confirm model speed, usable context, concurrency and output quality with the exact software and workload. The selected components and delivery route are matched to the configured system.	9 things to consider One model needs a guaranteed contiguous 64GB space The site cannot accept the measured dual-GPU load Rack serviceability or redundancy is required 64GB is aggregate capacity across two 32GB GPUs, not one universal memory pool. No multi-GPU model fit, thermals, noise or wall power has been measured. 64GB is aggregate capacity across two independent 32GB GPUs; the selected supplier lists no NVLink. A workstation is not presented as a redundant datacentre appliance. Confirm model speed, usable context, concurrency and output quality with the exact software and workload. The selected components and delivery route are matched to the configured system.	9 things to consider A 32GB route passes the acceptance test Several independent GPU services are required The buyer needs rack resilience or service density The exact RTX PRO 6000 Workstation or Max-Q selection remains a pre-quote decision. 96GB capacity does not prove a named model, context, speed or concurrency. The exact RTX PRO Workstation or Max-Q edition must be confirmed in the quote. A workstation is not presented as a redundant datacentre appliance. Confirm model speed, usable context, concurrency and output quality with the exact software and workload. The selected components and delivery route are matched to the configured system.	9 things to consider One 96GB GPU passes the workload acceptance test Enterprise rack serviceability or density is required The office cannot accept the final electrical, heat or noise envelope 192GB is aggregate capacity across two 96GB GPUs, not one universal memory pool. No dual-GPU topology, model fit, cooling, acoustic or performance result has been measured. 192GB is aggregate capacity across two 96GB GPUs; no NVLink or universal pooling claim is made. A workstation is not presented as a redundant datacentre appliance. Confirm model speed, usable context, concurrency and output quality with the exact software and workload. The selected components and delivery route are matched to the configured system.	8 things to consider One model needs more than 32GB per GPU without proven sharding A quieter tower meets demand The site cannot support 4U forced-air operation 128GB is aggregate capacity across four 32GB GPUs. The custom passive GPU, exact lane map, thermals and warranty remain supplier gates. Aggregate GPU memory is not one universal memory pool. Confirm model speed, usable context, concurrency and multi-GPU efficiency with the exact software and workload. Rack, electrical work, UPS, cooling, cabling, colocation and on-site installation require a separate scope.	9 things to consider Four workers meet demand The model requires NVLink or NVSwitch The site cannot support the final electrical and cooling design 256GB is aggregate capacity across eight 32GB GPUs. Power, thermal behaviour, PCIe topology and GPU warranty remain unverified. No single-model speed, concurrency or multi-GPU efficiency has been measured. Aggregate GPU memory is not one universal memory pool. Confirm model speed, usable context, concurrency and multi-GPU efficiency with the exact software and workload. Rack, electrical work, UPS, cooling, cabling, colocation and on-site installation require a separate scope.	9 things to consider The workload requires HGX NVSwitch Two or fewer workers meet demand The buyer lacks an enterprise rack operating route 384GB is aggregate capacity across four 96GB GPUs. This is a PCIe system; no NVLink, NVSwitch or universal pooled-memory claim is made. The old Max-Q rack wording is invalid and no exact Server Edition BOM has yet been quoted. Aggregate GPU memory is not one universal memory pool. Confirm model speed, usable context, concurrency and multi-GPU efficiency with the exact software and workload. Rack, electrical work, UPS, cooling, cabling, colocation and on-site installation require a separate scope.	9 things to consider Four workers meet demand The workload requires an HGX fabric Power, cooling or support is outstanding 768GB is aggregate capacity across eight 96GB GPUs. This is a PCIe architecture, not an HGX system and not one pooled 768GB memory space. No Kimi K3 fit, context, throughput, concurrency or fine-tuning result has been measured. Aggregate GPU memory is not one universal memory pool. Confirm model speed, usable context, concurrency and multi-GPU efficiency with the exact software and workload. Rack, electrical work, UPS, cooling, cabling, colocation and on-site installation require a separate scope.	11 things to consider A PCIe RTX PRO worker platform meets demand The purchase depends on unverified speed or concurrency The exact supplier and support route is outstanding 1,128GB aggregate HBM is below the current 1.5609TB official Kimi K3 weight-file total. Kimi K3 is only an aggressive-quantisation candidate on this route until an exact converted checkpoint is tested. No native load, full context, tokens-per-second, concurrency or quality result is claimed. The £349,000 guide price covers the complete H200 deployment described on this page. The released Kimi K3 weights exceed this system's aggregate GPU memory. A smaller conversion would be a different model version and would need separate testing. Aggregate GPU memory is not one universal memory pool. Confirm model speed, usable context, concurrency and multi-GPU efficiency with the exact software and workload. Rack, electrical work, UPS, cooling, cabling, colocation and on-site installation require a separate scope.	10 things to consider Independent PCIe workers meet the workload The decision depends only on aggregate VRAM Facility or software compatibility is not yet accepted 2.304TB HBM is a calculated capacity envelope, not proof that the official Kimi K3 checkpoint loads or serves one-million-token context. No Kimi K3 throughput, TTFT, concurrency, wall power or quality result has been measured. Pricing is quote-only. Model-fit entries remain candidates until reproducible tests exist. Aggregate GPU memory is not one universal memory pool. Confirm model speed, usable context, concurrency and multi-GPU efficiency with the exact software and workload. Rack, electrical work, UPS, cooling, cabling, colocation and on-site installation require a separate scope.	10 things to consider A single HGX system meets demand The programme lacks a named technical operator Budget or value depends on unverified public performance claims This is a complete 72-GPU liquid-cooled rack, not an the selected OEM 4U product or ordinary office delivery. 20TB is the rounded marketing class; 72 × 288GB equals 20,736GB aggregate GPU memory. Pricing is quote-only. 20TB is aggregate rack accelerator memory and not a generic one-process promise. Aggregate GPU memory is not one universal memory pool. Confirm model speed, usable context, concurrency and multi-GPU efficiency with the exact software and workload. Rack, electrical work, UPS, cooling, cabling, colocation and on-site installation require a separate scope.
Product page	View Team 32	View Company 64	View Studio 96	View Studio 192	View Value Rack 128	View Value Rack 256	View Enterprise 384	View Enterprise 768	View H200 1.1TB	View Frontier Native 2.3TB	View Frontier Rack 20TB

Every system is configured around the selected workload, software, facility, delivery and support needs.

Open 4U OEM GPU server chassis showing passive GPUs, cooling fans, processors and memory slots — OEM supplier render showing one possible internal layout. Components vary with the ordered build. OEM supplier reference image.

Diagram combining model weights, context, cache and active requests into a memory headroom check — Memory fit depends on the workload, context, cache and simultaneous demand, then needs a representative test. GPU Servers technical illustration.

Exploded supplier render of passive GPUs arranged above an open 4U rack chassis — Supplier layout render used to explain GPU density and airflow. It does not represent a confirmed package configuration. OEM supplier reference image.

After the shortlist

Turn the shortlist into the right configuration.

Share the model, workload, users, site conditions and budget. We will help you choose the most suitable system and software route.

Talk through your shortlist Discuss the shortlist

Decision check

GPU Servers Vs AI Workstations: Fit, Evidence & Next Steps

When assessing GPU servers vs AI workstations, start with the real workload, operating boundary, available evidence and credible alternatives.

Relevant supporting considerations include AI workstation vs server, GPU workstation and rackmount GPU server. A sound GPU servers vs AI workstations decision should make inputs, limitations, responsibilities and the next practical check clear.

Compare GPU Servers vs AI Workstations for Your Workload

Build a two-to-four system shortlist.

Review the selected systems.

Compare the machine and its operating boundary

Turn the shortlist into the right configuration.

GPU Servers Vs AI Workstations: Fit, Evidence & Next Steps