Bare-Metal vs Cloud GPU Instances: Which Is More Cost-Effective on Nebula Block?

The lowest hourly rate isn't the whole story

When renting GPUs to run AI models, many people start with a simple question: which configuration has the lowest hourly rate?

But for inference workloads that run continuously, the hourly rental price does not tell the whole story. A more expensive configuration can still be more cost-effective if it processes more tokens in the same amount of time or meets stricter latency requirements. That is why Bare Metal and virtual machines (VMs) should be compared on both cost and real-world performance.

On Nebula Block, both options are billed hourly, based on the configuration and how long the instance exists. Before deciding which one offers better value, it helps to understand how billing works, look at the prices that are published, and know how to check the actual rate for each configuration.

The figures in this article are based on information available as of October 28, 2026.

1. How do Bare Metal, VMs and Containers differ?

According to the GPU Cloud docs, Nebula Block offers three GPU deployment options: virtual machines, containers and Bare Metal.

GPU configurations are available in Canada, the United States, Finland, France and Norway. Depending on the region, you can choose anything from a single GPU to multi-GPU nodes, typically with eight GPUs and up to ten in some locations.

The hardware lineup includes B300, B200, H200, H100, A100, L40/L40S and RTX GPUs.

2. Billing doesn't stop when you stop using an instance

On Nebula Block, charges are calculated hourly and deducted from your account's credit balance for as long as the instance exists. That is easy to overlook when you only use GPUs for testing, benchmarking or model development during working hours.

A few billing rules are worth keeping in mind:

• Account requirements: GPU instances require Tier 3, which comes with a $10 deposit.
• Spending limits: accounts have an hourly spending limit. If you use multi-GPU nodes, you may need to raise it above the default for new accounts.
• When credits run out: the instance is automatically deleted when your credit balance is exhausted. For long-running workloads, consider enabling auto-pay to avoid interruptions.
• Spot instances: some configurations are available at lower, market-based Spot rates but can be reclaimed. When that happens, compute billing pauses, while your data is retained at a lower storage rate until you restart or delete the instance.

The most important point is that stopping an instance does not stop billing. If you only need GPUs on weekdays, leaving an instance around for the whole weekend creates costs you didn't intend.

3. How much do GPUs cost on Nebula Block?

The List Products API documentation includes sample prices for VMs. The table below uses only those samples. It gives a general idea of pricing and should not be treated as an exact quote for any configuration.

The two single-GPU H100 examples show that the same GPU can be priced differently: $2.632 and $2.696 per hour, a gap of $0.064. The documentation doesn't say disk capacity is the only reason, so when you compare configurations, keep the GPU, CPU, RAM and storage as similar as you can. The API also returns one price for the whole configuration, not per GPU, so divide by the GPU count to compare fairly.

How to check Bare Metal and VM pricing

The List Products API returns the hourly price and form factor for each configuration. The product_type field tells you whether a product is a VM, Container or Bare Metal instance, and price_per_hour is the price for that configuration. When comparing, you should:

1. Select the same region and GPU family.
2. Compare CPU, RAM and storage as well.
3. Retrieve the actual prices for both VM and Bare Metal through the API or the console.
4. Divide price_per_hour by gpu_count when comparing cost per GPU.
5. Check is_available to confirm the configuration can be deployed.

Because the API needs an account token, you'll see the real Bare Metal rate inside your own account.

4. How quickly can GPU costs add up?

To illustrate, let's use the sample VM rates: $2.632 per hour for one H100 and $57.20 per hour for an 8-GPU B200 node. The calculations assume 730 hours per month and are only meant to show how costs accumulate.

These calculations use sample prices from the documentation and are not official quotes.

The $126.34 weekend row deserves attention. If you only need GPUs during working hours, leaving an instance running, or simply leaving it undeleted, for another 48 hours can create a real charge even when no workload is running. The effect is bigger with multi-GPU nodes: at the sample rate of $57.20 per hour, an 8-GPU B200 VM would cost more than $41,000 if kept on for a month.

Before optimizing GPU prices, review your instance usage history and work out which resources can be deleted when they're no longer needed.

5. When does Bare Metal actually pay off?

Bare Metal can cost more per hour than a VM and still be more economical if the performance gain is large enough. For inference workloads, one useful metric is cost per token. If Bare Metal processes more tokens in the same amount of time, a higher hourly rate can be offset by better throughput.

Bare Metal has a lower cost per token than a VM when:

Bare Metal hourly price ÷ VM hourly price < Bare Metal throughput ÷ VM throughput

In other words, the price ratio between Bare Metal and a VM has to be lower than their throughput ratio.

Here's a hypothetical illustration, not a Nebula Block measurement. If Bare Metal processed 10% more tokens per second than a VM on your model, it could cost up to 10% more per hour and still break even. Against the sample VM price of $2.632, that ceiling is about $2.90 an hour. If the speed-up were only 5%, the ceiling would drop to about $2.76.

To get a real number for your system, run the same workload on both and measure throughput in tokens per second. Cost per million tokens is then:

hourly price ÷ (tokens per second × 3,600) × 1,000,000

Results depend on the model, context length, serving configuration and how the workload uses the GPU, so one test on one model isn't enough to generalize. Throughput isn't the only factor either. If your application has a strict p99 latency target, Bare Metal can still be a reasonable choice even when its cost per token is only comparable to a VM's. The most reliable way to decide is to get the actual H100 Bare Metal price from your account, then compare measured results on both options.

6. Should you choose Bare Metal or a VM?

Neither option is always better. The right choice depends on your workload, your performance requirements and how you use your resources.

• Choose a VM if you need a flexible environment and want to start testing with a configuration whose pricing is publicly listed.
• Consider Bare Metal if your inference workload is latency-sensitive or needs tighter performance control. Benchmark your own workload before paying a premium.
• Compare cost per token if your main goal is to optimize inference cost. The hourly rental rate is only one part of the equation.
• Delete instances you no longer need, so you aren't paying for idle resources.

Before deciding, make sure you're comparing equivalent configurations and measuring performance with the same workload. Looking only at listed prices, or at a single benchmark, can lead you to a choice that doesn't fit what you actually need.

Conclusion

Bare Metal isn't automatically better than a VM, and a VM isn't always cheaper once you measure the cost of each token processed.

The key is to weigh GPU rental costs alongside real-world performance. If Bare Metal raises throughput or cuts latency enough to offset its higher hourly rate, the premium may be justified. If your workload can't take advantage of those benefits, a VM may be the more economical option.

Learn more at

• Email: contact@nebulablock.com
• Website: nebulablock.com
• Docs: docs.nebulablock.com
• Book a call: nebulablock.com/contact