This guide is built from published pricing and provider documentation rather than sustained hands-on use on both platforms. GPU cloud pricing changes frequently, so verify current figures before committing.
Lambda and Hetzner get compared as if they compete for the same customer. They mostly don’t. Lambda rents you an H100 by the hour with a 99.9% uptime SLA. Hetzner rents you a dedicated server by the month, and its GPU options top out well below H100 territory.
The honest version of this comparison isn’t “which is cheaper.” It’s “which billing model and hardware tier actually fits what you’re building,” and that answer depends entirely on your workload shape.
In this guide:
- Why This Isn’t a Fair Fight
- Lambda: Hourly, Managed, SLA-Backed
- Hetzner: Monthly, Dedicated, EU-Based
- Real Pricing Side by Side
- When Hourly Billing Wins
- When Monthly Dedicated Wins
- Who Should Choose What
- Frequently Asked Questions About Lambda vs Hetzner
- Verdict
- Sources and Further Reading
Lambda vs Hetzner: Why This Isn’t a Fair Fight
The Lambda vs Hetzner comparison solves different problems for different people, and treating them as interchangeable options in the same shopping comparison sets up a bad decision either way.
Lambda is built for machine learning workloads specifically. You provision an H100 or similar, pay by the hour, and get a 99.9% uptime SLA backed by an infrastructure provider whose whole business is AI compute. Hetzner is a general-purpose dedicated server host that happens to offer GPU configurations. You commit to a server for a month, and its GPU lineup targets rendering, transcoding and lighter inference rather than frontier-model training.
The billing models reflect that difference. Lambda’s per-hour rate makes sense for bursty, project-based work where you might need eight H100s for three days and nothing for the following month. Hetzner’s monthly rate makes sense for a workload that runs continuously and doesn’t need the newest silicon.
Neither approach is wrong. They’re answers to different questions.
Lambda: Hourly, Managed, SLA-Backed
In the Lambda vs Hetzner comparison, Lambda’s positioning is squarely toward serious machine learning work. On-demand H100 pricing runs roughly $2.49 to $3.29 per hour depending on configuration, with a published 99.9% uptime SLA and InfiniBand networking available for multi-node training.
What Lambda does not offer is worth noting as much as what it does. There’s no spot or interruptible pricing tier, so you always pay the full on-demand rate. There’s no serverless or per-second billing option for short bursts. And the footprint is largely US-based, which matters if data residency or latency to European users is a factor.
What you get in exchange is a platform built specifically for training and fine-tuning at scale, with the networking and support infrastructure that implies. For teams running actual GPU clusters rather than single instances, that specialisation is the point.
Hetzner: Monthly, Dedicated, EU-Based
On the Hetzner side of the Lambda vs Hetzner comparison, its GPU servers are dedicated hardware rented monthly, an entirely different commercial model from Lambda’s on-demand hourly compute.
The current lineup, as published:
- GEX44, RTX 4000 Ada, 20GB VRAM, from roughly €184/month
- GEX45, RTX PRO 4000 Blackwell
- GEX131, RTX PRO 6000 Max-Q
None of these are H100-class cards. They’re workstation and prosumer GPUs suited to rendering, video transcoding, smaller-model inference and fine-tuning rather than large-scale frontier training. That’s a genuine ceiling on what Hetzner can do for you, not a limitation of the pricing model.
What Hetzner offers instead is EU data centre location, which matters directly for GDPR-sensitive workloads, and a monthly-dedicated cost structure that becomes very cheap per hour once you divide it out, provided you actually use the server continuously.
Lambda vs Hetzner: Real Pricing Side by Side
Lambda vs Hetzner pricing, approximate figures as published, converted to a common basis where reasonable:
| Lambda | Hetzner | |
|---|---|---|
| Billing model | Hourly, on-demand | Monthly, dedicated |
| Top-tier GPU | H100 | RTX PRO 6000 Max-Q (GEX131) |
| Entry GPU pricing | ~$2.49 to $3.29/hr (H100) | ~€184/month (GEX44, RTX 4000 Ada) |
| SLA | 99.9% published | Standard dedicated-server terms |
| Spot or interruptible pricing | None | Not applicable, dedicated model |
| Data centre region | Primarily US | EU |
| Networking | InfiniBand available | Standard dedicated networking |
Converting Hetzner’s GEX44 to an hourly-equivalent rate for this Lambda vs Hetzner comparison lands somewhere around €0.25 per hour if run continuously for a full month, dramatically below Lambda’s H100 rate. But that comparison is misleading on its face: you’re comparing a 20GB workstation card against a flagship 80GB training accelerator. It’s the wrong pair to compare directly.
Lambda vs Hetzner: When Hourly Billing Wins
In the Lambda vs Hetzner decision, Lambda’s model, and hourly cloud GPU generally, suits several specific patterns.
Bursty, project-based work. A training run that needs eight GPUs for three days and nothing after. Paying by the hour means you only pay for the burst.
Uncertain duration. Early-stage experimentation where you don’t know if a workload needs two days or two weeks. Monthly commitment is a bad bet when the timeline is unknown.
Needing the newest hardware. H100-class silicon isn’t available in Hetzner’s lineup at all. If your workload genuinely needs that tier, Lambda or a comparable hourly provider is the only path.
Multi-node training. Lambda’s InfiniBand networking supports the distributed training patterns that large model work requires, which dedicated single servers generally don’t.
For deeper coverage of hourly and marketplace GPU rental generally, our Vast.ai review covers the cheaper, less predictable end of the same hourly category.
Lambda vs Hetzner: When Monthly Dedicated Wins
In the Lambda vs Hetzner decision, Hetzner’s model suits a different, equally valid pattern.
Continuous, predictable workloads. An inference endpoint that runs 24/7 at steady load. Once utilisation is high enough, monthly dedicated pricing beats hourly cloud by a wide margin.
Workloads that fit a workstation-class GPU. If a 20GB or similar card handles your models, there’s no reason to pay for H100-class hardware you don’t need.
EU data residency requirements. GDPR and similar regulatory requirements that mandate EU-based processing rule out Lambda’s largely US footprint outright.
Budget-constrained continuous serving. A small team running a model-serving endpoint on a fixed budget, where the monthly figure needs to be predictable rather than usage-dependent.
The crossover point is usage intensity. Below roughly 50% to 60% monthly utilisation, hourly billing on comparable hardware is often cheaper. Above that, dedicated monthly pricing pulls ahead, and the gap widens the closer you get to continuous use.
Lambda vs Hetzner: Who Should Choose What
| If you are | Start with | Why | Link |
|---|---|---|---|
| Training or fine-tuning on H100-class hardware | Lambda | Only option here with that hardware tier | Check pricing |
| Running bursty, unpredictable workloads | Lambda or a marketplace like Vast.ai | Hourly billing matches uncertain duration | N/A |
| Running a steady inference endpoint on modest hardware | Hetzner | Monthly dedicated is far cheaper at high utilisation | Check pricing |
| Bound by EU data residency requirements | Hetzner | EU data centres; Lambda is largely US-based | N/A |
| Wanting the cheapest possible hourly rate | Vast.ai | Marketplace pricing undercuts managed providers | N/A |
| Needing 96GB in one card, owned outright | RTX PRO 6000 Blackwell | Ownership beats rental past a real usage threshold | N/A |
Frequently Asked Questions About Lambda vs Hetzner
Is Hetzner cheaper than Lambda?
Per hour of equivalent hardware, not meaningfully comparable, because Hetzner doesn’t offer H100-class GPUs. Per hour of workstation-class hardware run continuously, Hetzner is considerably cheaper than any hourly cloud provider. The comparison only makes sense within a hardware tier.
Can I get an H100 on Hetzner?
No. Hetzner’s GPU lineup tops out at the RTX PRO 6000 Max-Q, a workstation-class card. For H100 or similar frontier training hardware, Lambda or a comparable provider is necessary.
Does Lambda offer monthly billing?
Lambda’s core model is hourly on-demand. Reserved and contract pricing exists for larger sustained commitments, but the accessible entry point for most users is hourly.
Is Hetzner suitable for training large models?
For fine-tuning smaller models and running inference, yes. For training frontier-scale models from scratch, no, the hardware tier doesn’t support it. Match the workload to the card, not the other way round.
Why does data residency matter here?
If your data or compliance regime requires EU-only processing, Lambda’s largely US-based infrastructure may disqualify it regardless of price or performance. Hetzner’s EU presence is a structural advantage for that specific requirement, not a minor detail.
What’s the real crossover point between hourly and monthly?
Roughly 50% to 60% monthly utilisation is the common rule of thumb across cloud compute generally. Below that, hourly billing usually wins. Above it, monthly dedicated pricing pulls ahead, assuming comparable hardware tiers.
The Networking Difference That Doesn’t Show Up in Price Tables
One factor rarely gets weighed properly in a Lambda vs Hetzner comparison: what happens when a single GPU isn’t enough.
Lambda’s InfiniBand networking exists specifically to support multi-node training, where several machines need to communicate at very high bandwidth to keep a distributed training job efficient. This is infrastructure most workloads never touch, but for anyone training or fine-tuning at a scale that spans multiple GPUs across multiple machines, it’s the difference between a job that scales cleanly and one that bottlenecks on network transfer.
Hetzner’s dedicated servers use standard networking between machines, adequate for most inference and single-server fine-tuning but not designed for the tight synchronisation that distributed training demands. This isn’t a criticism of Hetzner so much as a reflection of what its GPU lineup is for: workstation-class cards for workstation-class jobs, not distributed frontier training.
The practical Lambda vs Hetzner implication: if your roadmap includes scaling from one GPU to a cluster, that trajectory favours Lambda or a similar training-focused provider from the start, even if a single Hetzner server would handle today’s workload. Migrating a distributed training pipeline between providers later is considerably more painful than picking the right starting point.
Estimating Your Actual Utilisation
The crossover point in the Lambda vs Hetzner decision, and in hourly-versus-monthly billing generally, depends on a number most teams have never actually calculated: what percentage of the month a GPU would realistically sit busy.
A rough way to estimate it for a Lambda vs Hetzner decision: track how many hours in a typical week your workload would need a GPU running, then divide by the total hours in a week. A model-serving endpoint answering requests around the clock lands close to 100%. A research workload run in bursts during working hours might land closer to 20% to 30%. A nightly batch job that runs for two hours each night is under 10%.
Below roughly 30% utilisation, hourly billing on Lambda or a marketplace provider is very likely cheaper than any monthly dedicated option, because you’re only paying for the hours actually used. Above 60%, the monthly dedicated model on Hetzner starts winning by a wide and growing margin, since the per-hour cost of a dedicated server falls the more continuously it runs.
The middle ground, 30% to 60%, is genuinely close and depends on the specific hardware tier and rates at the time you’re comparing. That’s the range worth running actual numbers for rather than defaulting to either provider on instinct.
Common Mistakes in This Comparison
Three patterns worth avoiding in any Lambda vs Hetzner decision, or any hourly-versus-dedicated choice generally.
Comparing hourly rates across hardware tiers. An hourly H100 rate and a monthly workstation-card rate aren’t comparable numbers. Match the hardware to your actual model requirements first, then compare pricing within that tier.
Committing to monthly dedicated before confirming utilisation. A server that sits mostly idle costs more per useful hour than hourly billing would have, even though the sticker price looks lower. Estimate utilisation honestly before committing.
Ignoring data residency until after choosing a provider. If EU data residency is a hard requirement, it should filter the provider list first, before price or performance enter the comparison. Discovering the constraint after committing to Lambda is an expensive mistake to unwind.
Verdict
The Lambda vs Hetzner question isn’t really about two products competing for the same purchase decision, and the honest answer to “which is better” is “it depends which question you’re actually asking.” Lambda is the right choice for H100-class training and bursty, unpredictable workloads where hourly billing and a strong SLA matter more than raw cost per hour. Hetzner is the right choice for continuous, moderate workloads on workstation-class hardware, particularly where EU data residency is a requirement rather than a preference.
In any Lambda vs Hetzner comparison, the mistake worth avoiding is picking based on the wrong axis: choosing Hetzner because it looks cheaper per hour without checking whether its hardware tier can run your models at all, or choosing Lambda for a steady-state inference endpoint that would cost a fraction as much on dedicated monthly hardware.
Concrete Lambda vs Hetzner next step: work out your actual utilisation pattern before comparing prices. If you can’t predict whether a workload runs 10% or 90% of the month, that uncertainty itself is the answer, and it points toward hourly billing regardless of which provider you choose.
Sources and Further Reading
- Lambda GPU Cloud pricing, current on-demand rates and available hardware tiers
- Hetzner GPU server matrix, current dedicated GPU server configurations and monthly pricing
- Vast.ai review, the marketplace alternative for cheaper, less predictable hourly compute
- RTX PRO 6000 Blackwell review, the ownership alternative once rental costs exceed buying
- LLM API cost optimization, reducing compute spend before deciding how to rent it
For related coverage on this site, see our Vast.ai review for the cheapest marketplace option, the RTX PRO 6000 Blackwell review for when owning beats renting, and our LLM inference server comparison for the software side of serving models on either platform.
