Hetzner GPU Cloud: European Cloud Provider GPU Pricing and Availability in 2026

Hetzner, the largest German cloud provider, entered the GPU cloud market in 2025-2026 with a focused portfolio optimized for European AI workloads. The offering centers on NVIDIA H100 (80 GB SXM), A100 (80 GB SXM), L40S (48 GB), and A16 (16 GB) instances across Hetzner's data centers in Nuremberg (Germany), Helsinki (Finland), and Falkenstein (Germany). The pricing strategy follows Hetzner's traditional approach: minimal overhead, no egress fees within Europe, and straightforward hourly billing with no upfront commitments.

Hetzer's GPU pricing undercuts the US hyperscalers in European data centers by 35-55 percent. An H100 80 GB instance on Hetzner runs 3.19/hr (8 GPU node at 25.52/hr) versus AWS P5 in eu-west-1 at 5.43/GPU/hr and Azure ND H100 v5 in West Europe at 5.75/GPU/hr. The A100 80 GB costs 1.92/hr per GPU versus AWS P4d in Frankfurt at 2.83/hr. These prices include VAT for EU customers, whereas hyperscaler pricing is typically quoted exclusive of VAT, adding 19-25 percent for European buyers.

Hetzner's H100 and A100 nodes use standard NVIDIA HGX baseboard configurations: 8 GPUs per node with NVLink 3.0 (900 GB/s GPU-to-GPU) for H100, NVLink 2.0 (600 GB/s) for A100. Inter-node networking uses InfiniBand NDR200 (200 Gb/s) on H100 clusters, lower than Lambda's NDR400 and AWS's 3200 Gbps EFA but adequate for training clusters up to 64 GPUs where scaling efficiency reaches 82-87 percent. The InfiniBand backbone covers nodes within the same data center hall; cross-hall connectivity uses 100 Gbps Ethernet.

The networking is Hetzner's primary limitation versus hyperscalers. For single-node training and fine-tuning (the most common European AI workload), the difference is negligible. For 64+ GPU distributed training, Hetzner's 200 Gbps InfiniBand versus AWS's 3200 Gbps EFA means 8-15 percent lower scaling efficiency. Hetzner's positioning is deliberate: the European AI market skews toward fine-tuning (70 percent of GPU workloads) and small-to-medium model training, not the 1000+ GPU clusters common in US-based foundation model training.

Hetzner's European data center locations provide a structural advantage for EU-based AI workloads under GDPR and the post-Schrems III regulatory regime. All Hetzner GPU data centers are in GDPR-compliant jurisdictions (Germany, Finland), with data never touching US-based cloud infrastructure. For European companies processing personal data through AI models (medical imaging, HR screening, customer profiling), Hetzner eliminates the standard clause that US hyperscalers are subject to US surveillance laws (FISA 702, EO 12333).

This regulatory compliance is not just legal comfort: it affects costs. European enterprises using AWS or GCP for GDPR-sensitive GPU workloads require additional contractual measures (Data Processing Addendums, SCC-plus provisions) and may need data residency configurations that restrict GPU instance availability. GCP A3 is available only in 3 regions globally, of which only one (europe-west4, Netherlands) is in Europe. On Hetzner, all three GPU data centers are EU-based, with the Nuremberg data center offering the lowest latency for Central European inference workloads.

Hetzner's GPU capacity is limited relative to hyperscalers. As of mid-2026, Hetzner operates approximately 3,000 H100 GPUs and 8,000 A100 GPUs across three data centers. This is roughly 1-2 percent of AWS's GPU capacity. Large GPU allocations (128+ GPUs) require pre-booking with 4-8 weeks lead time. Hetzner does not offer spot GPU pricing, and reservations are limited to 1-month and 12-month commitments with 15 percent and 25 percent discounts respectively.

The geographic limitation is also significant. Hetzner's GPU instances are only available in European data centers. For teams requiring inference deployment in North America, Asia, or Australia, Hetzner requires integration with another provider for non-EU regions. Latency from Hetzner's EU data centers to US East Coast users is 80-100 ms, to Asia Pacific 200-350 ms, making it unsuitable for global inference serving.

A total cost comparison for a typical European AI startup running a 4-GPU H100 node for inference fine-tuning across 30 days: Hetzner at $14.28/hr (4x H100) = $10,282/month; AWS eu-west-1 P5 spot = $8.12/hr (4x H100 spot) = $5,846/month plus $1,200 storage and egress = $7,046/month; AWS P5 on-demand = $16.39/hr (4x H100) = $11,801/month. Hetzner lands between AWS on-demand ($11,801) and spot ($7,046) for committed 30-day usage.

For teams with steady 24/7 GPU usage, Hetzner's 12-month commitment ($10.71/hr for 4x H100) brings the monthly cost to $7,711, competitive with AWS spot but without termination risk. The zero egress cost within European data centers saves $200-800/month for teams moving training data and model checkpoints. Factor in the 19 percent VAT included in Hetzner pricing that is added to AWS/US pricing, and the effective cost for EU-based companies narrows further.

Hetzner GPU is the right choice for EU-based organizations with data sovereignty requirements, particularly in healthcare, legal tech, and financial services where GDPR compliance and US data access avoidance are legal requirements. The cost advantage is strongest for teams running steady-state GPU workloads (fine-tuning, model training, continuous inference) in single-digit GPU configurations (1-32 GPUs) where the InfiniBand NDR200 networking is sufficient.

Hetzner is not suitable for teams needing 100+ GPU clusters, global inference deployment, or access to managed ML services (no SageMaker/Vertex equivalent). For the large and growing European AI market, Hetzner fills a niche between hyperscaler pricing and the need for data sovereignty. Given the regulatory trajectory in Europe (EU AI Act enforcement, continued Schrems challenges), the demand for sovereign GPU infrastructure is likely to grow significantly through 2026-2027.