Hetzner's GEX131 GPU Server: What €889/Month for a Blackwell RTX PRO 6000 Buys Against AWS's g5.xlarge

Hetzner's newest GPU box costs €889 a month (about $989) and ships with 96GB of VRAM on a single card. AWS's most-reached-for GPU instance, g5.xlarge, costs $734.38 a month on-demand and gives you 24GB. On the sticker price alone, AWS looks cheaper. Divide by the gigabyte of VRAM you actually get, and the ranking flips: Hetzner's card costs a third as much per GB as AWS's — and that's before you hit the wall where AWS simply doesn't sell a single GPU with 96GB of memory in an instance size a team can reach for without a six-figure annual commitment.

Here's the worked math, the trade-offs Hetzner's flat-rate model carries that the sticker price hides, and where each one actually wins.

The Headline Numbers, Side by Side​

Hetzner's GEX131 pairs an NVIDIA RTX PRO 6000 Blackwell Max-Q — 96GB of GDDR7 ECC memory, 5th-generation Tensor Cores — with a 24-core Intel Xeon Gold 5412U, 256GB of DDR5 ECC RAM, and 2×960GB NVMe SSDs. It bills at a flat €889.00/month (€1.4247/hour if you only need it for a few hours or days — Hetzner charges no setup fee specifically because it expects short-lived AI jobs). Traffic is unlimited and free.

AWS's g5.xlarge pairs a single NVIDIA A10G — 24GB of GDDR6 — with 4 vCPUs and 16GB of RAM, at $1.006/hour on-demand in us-east-1. That's $734.38 for a 730-hour month, before EBS storage (billed separately) and data-transfer egress ($0.09/GB after the first 100GB free each month).

Sticker price says AWS is $255 cheaper a month. Cost per gigabyte of VRAM — the number that actually determines what model you can load — says Hetzner is roughly 3x more efficient.

The Real Yardstick: Cost Per Gigabyte of VRAM, Across the Pricing Options​

A single data point invites the objection that on-demand is the wrong AWS price to compare against. It's fair — AWS gives you two big levers to bring g5.xlarge's cost down, and neither one closes the gap:

AWS Spot for g5.xlarge runs around $0.4419/hour (~$322.59/month), a genuine 56% discount off on-demand — but it's interruptible by design, so it's a fit for retryable batch jobs, not a service that needs to stay resident. At spot pricing, cost per GB drops to about $13.44/GB — closer to Hetzner, but the workload has to tolerate reclamation to get there, which most inference-serving deployments can't.

AWS 3-year Reserved Instances offer up to 72% off on-demand for g5.xlarge, which works out to roughly $205–$260/month depending on payment structure — landing at $8.50–$10.80/GB, genuinely competitive with Hetzner's per-GB rate. The catch is the commitment: three years, paid mostly or entirely upfront, locked to one instance type in one region, and you're still capped at 24GB no matter how long you commit for.

So across every AWS pricing lever — on-demand, spot, and 3-year reserved — the honest range is $8.50 to $30.60 per GB of VRAM, and only the least-flexible option (a 3-year prepay) reaches parity with Hetzner's flat month-to-month rate. Hetzner doesn't discount for commitment because it doesn't need to; $10.30/GB is the day-one, cancel-anytime price.

What the Other 72GB Actually Buys​

The gap matters because VRAM is the hard ceiling on what model fits on one card, and 24GB and 96GB aren't the same category of workload.

A10G's 24GB comfortably fits 7B–13B parameter models at FP16, or larger models pushed down to 4-bit quantization to survive the memory budget — usable, but you're trading answer quality for the smaller card. The RTX PRO 6000 Blackwell's 96GB fits 70B-parameter models at Q6 or Q8 quantization on a single card — the sweet spot Ollama defaults to on this hardware — without the accuracy loss of Q4, and without needing to split the model across multiple GPUs with the added latency and orchestration complexity that comes with tensor or pipeline parallelism. It also has enough headroom to run mid-size models at their native precision with room left over — some 96GB-class benchmarks report context windows scaling to 262K tokens for models that fit entirely in VRAM.

That headroom exposes an AWS gap the on-demand price alone doesn't show: AWS doesn't sell a single 96GB GPU in an accessible instance size at all. The next step up from g5.xlarge's 24GB is g6e.xlarge with a single L40S at roughly 45GB — still under half of Hetzner's card — for $1,358.53/month on-demand ($30.18/GB, no better than g5.xlarge's ratio). To actually match or exceed 96GB on AWS, you leave single-GPU instances behind entirely and jump to 8-GPU multi-card configurations like p5.48xlarge (8×H100, 640GB combined) at roughly $98/hour — north of $71,000 a month. There's no AWS instance that gives a team a single 96GB card at a price between "slightly more than g5.xlarge" and "eight-GPU cluster money." Hetzner's GEX131 sits in a gap AWS's instance lineup doesn't cover.

The Egress Bill AWS's Sticker Price Doesn't Show​

Compute isn't the only line item. AWS bills data-transfer egress at $0.09/GB after the first 100GB free each month; Hetzner's traffic is unlimited and free regardless of volume. That gap scales with how much traffic the inference service actually serves, and it's worth pricing at a realistic volume rather than assuming it's negligible.

A moderately trafficked inference API — one serving completions, embeddings, or RAG responses across a handful of tenants rather than a single low-traffic side project — can plausibly push 2TB (2,000GB) of response egress a month. On AWS, that's (2,000GB − 100GB free) × $0.09/GB = $171/month, stacked on top of the $734.38 compute bill, for an effective $905.38/month on a g5.xlarge — closing most of the gap to Hetzner's flat $989 while still capping out at 24GB of VRAM. A lighter-traffic service serving a few hundred gigabytes a month narrows that egress line to single digits or low tens of dollars, so the effect scales down for smaller deployments — but it never goes negative, and it's a cost Hetzner's flat rate simply doesn't carry at any traffic volume. For a service that's actually busy enough to justify a dedicated GPU in the first place, egress is a real, traffic-dependent tax on the "AWS is cheaper" read of the sticker price, not a rounding error.

What the Flat Rate Doesn't Include​

None of this makes Hetzner strictly better — it makes the trade-offs concrete instead of hidden inside a sticker price.

No spot equivalent. Hetzner's €889/month is the only price on offer — there's no interruptible tier for batch or dev workloads that could run for 56% less on AWS spot. If your GPU sits idle nights and weekends, you're paying Hetzner's flat rate for that idle time regardless.

No managed ML services. SageMaker-style managed training, autoscaling inference endpoints, and one-click model deployment don't exist on a Hetzner dedicated server. You provision the box, install CUDA and drivers yourself, and manage the stack — the same operational overhead as running any other bare-metal server.

EU-only data centers. Hetzner's GPU fleet is European (Falkenstein, and comparable German/Finnish sites). For teams with EU data-residency requirements this is a plus; for teams needing US-region compute or global multi-region placement, it's a hard constraint AWS's 30+ regions don't share.

Manual provisioning, not instant. Spinning up a g5.xlarge takes the same seconds as any EC2 instance. A Hetzner dedicated GPU server is provisioned as physical hardware — real, but not the sub-minute elasticity of a cloud API call, and availability of a specific SKU in a specific location isn't guaranteed the way EC2 capacity effectively is.

Where This Lands for a Self-Hosted Fleet​

For a team running AI-agent sandboxes or inference workloads that need to stay resident — not bursty, not tolerant of interruption — the per-GB math favors owning the card outright: $989/month buys 96GB with no interruption risk and no three-year prepay. For workloads that are genuinely bursty or dev/test-shaped, AWS spot at $322.59/month remains the cheaper floor, provided the workload can actually tolerate reclamation.

If you're already running a Cluster-API-managed fleet — CAPH provisioning Hetzner machines the same way it provisions any other node — a GEX131 slots in as one more managed machine rather than a separate GPU-cloud account and billing relationship to reconcile. It's still a bare-metal box you set up CUDA and drivers on yourself; nothing here replaces that operational work, and a platform sitting on top of it doesn't add a managed-ML layer AWS doesn't already lack either. What it does remove is the second invoice, the second region model, and the second control plane for a resource that, at steady state, is cheaper per gigabyte than anything AWS's GPU instance lineup offers short of a three-year prepay you'll never make on a single 24GB card.

Bex.co is the open-source, AI-native Render alternative — push a git repo, get a running HTTPS service on machines you own, GPU boxes included, provisioned through the same Cluster-API-managed fleet as everything else. Star the repo on GitHub or deploy your first app today.

Sources​

• Hetzner presents GPU server GEX131 with NVIDIA RTX PRO 6000 Blackwell Max-Q
• High performance dedicated GPU server - Hetzner GEX131
• g5.xlarge Specs & Pricing - Spot, On-Demand & Savings Plans | DoiT Compute
• Amazon EC2 G5 Instances
• g6e.xlarge pricing - AWS EC2
• RTX PRO 6000 Blackwell For AI Inference: When It Makes Sense
• NVIDIA RTX Pro 6000 (Blackwell) — 96GB GPU for LLM Inference)
• AWS Data Transfer Pricing 2026: Stop Overpaying for Cloud
• AWS Savings Plans vs Reserved Instances 2026: The Decision Guide.