Hetzner has introduced the GEX430, its first dedicated multi-GPU server, aimed at artificial intelligence workloads that demand large amounts of video memory. The machine combines four NVIDIA RTX PRO 6000 Blackwell Max-Q GPUs, each with 96 GB of GDDR7 ECC memory, for a total of 384 GB of VRAM. The original announcement dates from October 21, 2026, with an update on November 2, and the topic gained fresh coverage on November 4 — Geekiki had not previously covered the GEX430.
According to Hetzner, each GPU connects directly to the processor via 16 PCIe 4.0 lanes, without sharing bandwidth with another card. The cards use fifth-generation Tensor Cores, fourth-generation RT Cores and FP4 support, and the company lists 3,511 AI TOPS per GPU. The rest of the configuration includes an Intel Xeon Gold 5412U processor with 24 cores and 48 threads, 128 GB of registered DDR5 ECC RAM and two 1.92 TB Gen4 NVMe SSDs in a data-centre edition.
What 384 GB of VRAM is for
The GEX430's central point is inference: the more simultaneous users a model serves, the more memory it tends to require. Hetzner cites models of 70 to 100 billion parameters at 16-bit BF16 precision, whose weights alone would require about 140 to 200 GB of VRAM. As examples, the company mentions Kolibri, from Aleph Alpha, with 78 billion parameters and about 156 GB at BF16, and Qwen3-Next-80B-A3B-Instruct, with a theoretical requirement of about 160 GB — neither would fit on a single GPU in the server.
One important detail: the 384 GB is the sum of the four cards, not a single shared memory pool. Splitting a large model across the GPUs requires suitable frameworks, as Hetzner itself notes. The other strategy is to distribute tasks: one model for customer support, another to assist developers and a third to triage requests, each on its own GPU, without reloading the model for every request.
Dedicated hardware in the European Union and price
Hetzner positions the GEX430 as an alternative for companies that want to run open-weight models such as GLM, Mistral and Qwen on dedicated hardware inside the European Union, rather than relying entirely on external platforms. According to the company, this helps keep source code, customer data and internal documents on their own infrastructure — provided the environment is configured appropriately — which can ease data-protection and compliance requirements.
“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,” said Stefan Weiberg, product manager at Hetzner, as quoted by Data Centre News. Pricing starts at €3,599 per month, with a one-off setup fee of the same amount, or $4,199 in the United States, excluding taxes and including one IPv4 address.
Image credit: illustration of a dedicated server. Credit: Hetzner.
Sources: Hetzner — GEX430: Our first multi-GPU server with 384 GB VRAM; Hetzner — High performance dedicated GPU server - Hetzner GEX430; Data Centre News — Hetzner launches first four-GPU server for AI work.
Transparency: this article was produced with the assistance of artificial intelligence, based on cross-checking public posts on X and sources on the internet, and should be verified against the original sources cited in this article.
Muse
