Hetzner GEX430

The Hetzner GEX430 marks a significant milestone for dedicated hosting, representing Hetzner's first bare-metal server configured with a multi-GPU topology. Packing four NVIDIA RTX PRO 6000 Blackwell Max-Q GPUs with 96 GB of GDDR7 ECC memory each, the system delivers a massive 384 GB of aggregate VRAM. Designed for engineers and organizations seeking sovereign, unshared compute without the hyperscaler premium, the GEX430 provides full root access to enterprise-grade workstation silicon inside a standardized European data center environment.

Positioned in the upper enterprise dedicated tier, this platform bridges the gap between single-GPU developer workstations and prohibitively expensive SXM-based cloud clusters. At 3,599 EUR per month (plus a one-time 3,599 EUR setup fee), the server gives production teams an alternative to on-demand cloud platforms like AWS or RunPod, offering steady-state economics for compute-heavy applications. Deploying the Hetzner GEX430 for AI tasks gives engineering teams dedicated, non-throttled access to hardware capable of handling high-concurrency inference, long-context deployments, and full model fine-tuning.

For developers seeking the best hardware for running AI models locally or in private dedicated infrastructure, this rig solves the primary bottleneck in modern machine learning: GPU memory capacity. Rather than dealing with shared virtualization or noisy neighbors, practitioners get four discrete GPUs, each connected directly to the host CPU over dedicated PCIe lanes. This setup delivers predictable latency and high throughput for production inference and agentic pipelines.

The computational engine of the GEX430 relies on NVIDIA's Blackwell architecture scaled down into the workstation-grade RTX PRO 6000 Max-Q form factor. The system delivers 4 x 3,511 AI TOPS, backed by 5th-generation Tensor Cores and native hardware acceleration for FP4 precision alongside FP8, FP16, and BF16 formats.

Key technical specifications include:

• GPU Configuration: 4x NVIDIA RTX PRO 6000 Blackwell Max-Q
• Total VRAM: 384 GB GDDR7 ECC (96 GB per GPU)
• AI Compute: 4 x 3,511 AI TOPS with 4th-gen RT Cores and 5th-gen Tensor Cores
• Interconnect: Direct-to-CPU PCIe 4.0 x16 lanes per card (no intermediate PCIe switch contention)
• Processor: Intel Xeon Gold 5412U (24 physical cores, 48 threads)
• System Memory: 128 GB DDR5 ECC Registered RAM
• Storage: 2x 1.92 TB Gen4 NVMe Datacenter Edition SSDs (configured in software RAID 1)
• Networking: 1 Gbit/s guaranteed unmetered uplink
• Location & OS Support: Helsinki (HEL 1 datacenter); Debian and Ubuntu (64-bit)

A critical architectural detail is the GPU interconnect. Unlike dense enterprise nodes that utilize NVLink bridges or NVSwitch fabrics, the GEX430 routes each card directly to the Intel Xeon Gold 5412U socket via 16 dedicated PCIe 4.0 lanes. While this eliminates bus-sharing bottlenecks between GPUs, inter-GPU communication must traverse the CPU root complex. For distributed tensor-parallel workloads across all cards, communication overhead is bounded by PCIe 4.0 bandwidth (roughly 31.5 GB/s bidirectional per card) rather than the terabytes-per-second speeds of NVLink.

Despite this limitation, the presence of GDDR7 ECC memory ensures high per-GPU memory bandwidth for token generation. When evaluating the Hetzner GEX430 VRAM for large language models, the 384 GB total pool offers substantial headroom for parameter caching, KV cache expansion, and large context windows across all four devices.

The primary advantage of deploying a 384GB GPU for AI workloads is raw memory capacity. The GEX430 easily serves as dedicated hardware for running 70B-100B at BF16 (across all four GPUs) parameter models without requiring lossy quantization.

• 70B to 72B Models (Llama 3.1 70B, Qwen 2.5 72B): These models fit completely unquantized at native 16-bit precision (BF16 or FP16), occupying approximately 140 GB to 150 GB across the four cards using 4-way tensor parallelism (tp=4). This leaves over 230 GB of aggregate VRAM entirely dedicated to the KV cache, allowing 128k context lengths even at elevated batch sizes.
• Mixture of Experts (Mixtral 8x22B): With 176 billion total parameters (39 billion active), Mixtral 8x22B fits comfortably in an unquantized FP8 format or an INT8/AWQ quantization across the 384 GB pool, running with low latency.
• Large Flagship Models (DeepSeek-R1, Llama 3.1 405B): Extremely large architectures like Llama 3.1 405B or DeepSeek-R1 (671B MoE) cannot run at 16-bit precision on 384 GB. However, DeepSeek-R1 fits using FP4 or aggressive INT4/K-quant schemes (e.g., GGUF or EXL2 quantizations around 3.5 to 4.0 bpw). Leveraging the Blackwell architecture's native FP4 Tensor Cores allows these massive weights to execute with hardware-accelerated matrix operations.
• Multi-Model Concurrency: Instead of pooling all four cards for a single model, the GEX430 can host four discrete 7B to 32B models simultaneously. A single 96 GB card can run a full unquantized Qwen 2.5 32B or Llama 3.1 8B instance with vast memory reserved for context.

When profiling Hetzner GEX430 AI inference performance, the sweet spot for single-model deployment sits at 70B parameter models quantized to FP8 or running in native BF16 via vLLM, SGLang, or TensorRT-LLM.

For Hetzner GEX430 tokens per second metrics, users running Llama 3.1 70B via 4-way tensor parallelism can expect generation speeds in the range of 40 to 65 tokens per second for single streams at FP8/BF16, scaling to aggregate throughputs exceeding 600 tokens per second under concurrent batched serving. Because inter-GPU transfers occur over PCIe 4.0 rather than NVLink, pipeline parallelism (pp=4) or running multiple independent models across the cards often yields better scaling efficiency than tensor-parallel weight splitting.

The GEX430 is built for engineering teams that have outgrown desktop hardware and need dedicated cloud infrastructure.

For teams seeking the best hardware for local AI agents 2026 demands, the GEX430 provides an ideal infrastructure anchor. Multi-agent systems require high-availability hosting for planning models, reasoning engines, and parallel execution agents. A team can allocate GPU 0 and 1 to a 70B reasoning model, GPU 2 to an embedding and reranking service, and GPU 3 to multimodal vision processing (such as running Qwen2-VL), all on a single tenant box.

While full pre-training of massive models is constrained by the PCIe interconnect, fine-tuning via LoRA and QLoRA on 70B models is straightforward. System memory (128 GB) combined with 384 GB of ECC VRAM allows practitioners to load base models in 16-bit, attach rank adapters, and train against custom datasets using frameworks like DeepSpeed or Axolotl.

European organizations requiring strict GDPR compliance can leverage this Hetzner GEX430 local LLM configuration to process proprietary documents and customer interactions completely within the EU (Helsinki datacenter), avoiding third-party API data leakage.

When evaluating the GEX430 as a candidate for the best AI chip for local deployment and dedicated hosting, consider how it stacks up against standard alternatives:

• Hetzner GEX430 vs. 4x NVIDIA A100 (80GB SXM) Instances: Public clouds charge between $8,000 and $14,000 per month for a reserved 4x A100 SXM node. The GEX430 costs a fraction of that (roughly €3,600/month) while offering more total VRAM (384 GB vs. 320 GB) and next-generation Blackwell compute features like FP4. However, the A100 SXM platform includes 600 GB/s NVLink, which outperforms the GEX430 in heavy all-reduce operations during large-scale pre-training.
• Hetzner GEX430 vs. On-Premises Quad RTX 4090 Workstations: A custom build with four consumer RTX 4090 GPUs offers only 96 GB of total non-ECC VRAM (24 GB each), preventing the execution of unquantized 70B models or large batch processing. The GEX430 provides four times the addressable memory, datacenter-grade ECC reliability, and eliminates the operational overhead of cooling, power management, and physical maintenance.

The GEX430 stands out as a high-capacity, cost-effective server for practitioners who prioritize massive aggregate VRAM, predictable monthly costs, and total infrastructure control over proprietary cluster interconnects.