Hetzner launches first four-GPU server for AI work

Hetzner has introduced the GEX430, its first multi-GPU server, with four Nvidia Blackwell GPUs.

The server is aimed at businesses and professional users running artificial intelligence workloads, including training, inference, fine-tuning, rendering and simulation. It is built around four Nvidia RTX PRO 6000 Blackwell Max-Q GPUs with a combined 384GB of GDDR7 ECC graphics memory.

The launch expands the German hosting provider's dedicated server range as demand grows for infrastructure that can handle larger AI models and more simultaneous users. Multi-GPU systems are becoming increasingly important for companies that want to keep AI workloads on their own hardware rather than rely entirely on external platforms.

Hetzner said the system lets customers run memory-intensive models and operate several models at the same time on a single server. Businesses can also assign different applications to individual GPUs, such as one model for customer support, another for software development assistance and a third for handling incoming requests.

The company placed particular emphasis on graphics memory for inference workloads, where demand rises with the number of concurrent requests. The GEX430's 384GB of VRAM is intended to support large language models serving many users at once, including chatbot deployments and internal knowledge assistants.

Each GPU has 96GB of GDDR7 ECC VRAM and supports fifth-generation Tensor Cores, fourth-generation RT Cores and FP4.

The wider hardware specification includes an Intel Xeon Gold 5412U processor with 24 cores and 48 threads, 128GB of DDR5 ECC registered memory and two 1.92TB datacentre-edition Gen4 NVMe SSDs. The server is offered as a dedicated machine rather than as part of a shared environment.

One of the main selling points is data location and control. Hetzner said customers can run open-weight AI models such as GLM, Mistral and Qwen on dedicated hardware within the European Union, which may appeal to organisations with stricter internal requirements for handling documents, customer records and source code.

That positioning reflects a wider market trend in Europe, where businesses are weighing the benefits of public AI services against concerns about data governance, model choice and infrastructure control. Hosting and colocation providers have increasingly sought to offer systems tailored to AI workloads as customers look for alternatives to hyperscale platforms.

Pricing for the GEX430 starts at €3,599 per month, with a one-off setup fee of €3,599. Hetzner also listed a US price of $4,199 per month and a setup charge of $4,199.

Stefan Weiberg, Product Manager at Hetzner, outlined the company's view of the product's role in inference-heavy deployments. "Especially when it comes to AI inference, graphics memory is just as crucial as computing power. The more requests a model processes simultaneously, the more memory it generally requires. With four GPUs and a total of 384 GB of VRAM, the GEX430 offers plenty of headroom for demanding AI applications. Companies can use it, for example, to run multiple language models in parallel or to provide their AI services for a large number of concurrent users," he said.

Hetzner has built its business around dedicated servers and virtualised server infrastructure, with data centre operations in Germany and Finland and additional international infrastructure in Singapore and the United States. The GEX430 adds a higher-density GPU option to that portfolio at a time when hosting providers are adapting their product lines to accommodate more specialised AI computing demand.