Bare metal vs cloud: cost and speed

Bare metal is a physical server that's all yours: no hypervisor, no other tenants. For a machine that runs flat out all month, it's usually the cheapest compute you can rent month to month, and local NVMe beats network storage on latency.

It won't run CPU work noticeably faster than a VM on a modern hypervisor, though: AWS's own benchmark puts that overhead at about 1%. On the cloud side, the same machine costs $420 to $1,033 a month on AWS depending on how long you commit, and sending 5 TB out adds $441 more.

One steady machine, priced eight ways

The machine: 32 vCPU, 64 GB of RAM, a 500 GB disk, running all 730 hours of an average month, at list prices as of October 10, 2026. On AWS that's a c6i.8xlarge (Intel, like our Standard VMs), and the plans are EC2 Instance Savings Plans with no upfront payment unless marked.

The two bare metal rows are Latitude.sh's starting list price for its closest server, not ours: we quote bare metal per build. OVHcloud lists a similar server at $467 a month plus setup.⁴ Both have more RAM than the other machines in the table.

On this machine, AWS only beats our VMs if you sign for three years and send little out. On a 3-year plan, AWS undercuts our monthly-term VM by about $190 to $250 a month on compute and disk. That lead is gone once roughly 2.3 to 2.9 TB a month leaves your servers, because AWS charges $0.09 per GB after the first 100 GB⁵ and we don't charge for egress at all.

Without a long contract, our VMs cost less. With no commitment, our VM costs about a third less than AWS on-demand before any traffic, and less than half once 5 TB goes out. Against AWS's 1-year plan it's a tie on compute, except AWS wants a year signed and we don't. Any real traffic puts us ahead.

Bare metal is the most machine for the money. The listed server has 16 cores that run 32 threads, the same as the AWS machine's 32 vCPUs (AWS counts each thread as a vCPU), plus twice the RAM, about 4.8 TB of local NVMe and 20 TB of egress in the price. Only AWS's 3-year all-upfront plan lists lower ($420.43), and only until about half a terabyte a month goes out.

Where bare metal is faster, and where it isn't

CPU: a rounding error on a modern hypervisor. AWS says its Nitro hypervisor "delivers performance that is indistinguishable from bare metal for most workloads."⁶ Its own HPC benchmark put a virtualized instance within 1% of the same hardware run as bare metal,⁷ and Brendan Gregg, then at Netflix, measured Nitro's overhead at "often less than 1%."⁸ If a vendor tells you bare metal runs your code 10 to 30% faster, ask for the benchmark.

Storage: this is the real gap. Cloud block storage like AWS's EBS sits across a network. AWS designs gp3 volumes for "single-digit millisecond latency" and only its top tier, io2 Block Express, for an average under 500 microseconds.⁹ A local NVMe drive on a bare metal box skips that network hop. If your database spends its time waiting on disk, you'll feel it. Our Standard VMs skip the hop too: their disks are local SSDs in the same host.

Noisy neighbors. On shared hardware, another tenant's spike can show up in your tail latency. On a box nobody else uses, it can't.

So will bare metal make your service answer faster? Only when the time goes to disk waits or contention. If your service is CPU-bound or waiting on the network, moving to metal won't change much. Profile first, then decide.

Bare metal cloud, dedicated servers and custom builds

• Bare metal cloud: a physical server you rent through an API and pay for by the hour or month. Some providers do rent bare metal by the hour, from about $0.62 an hour for the server in our table.⁴ It suits a benchmark or a short burst job where you need the whole machine.
• Dedicated server: the same idea on a longer term, usually a month or more, from a fixed menu.
• Custom build: hardware sized to your workload and quoted. This is how we sell bare metal today. We don't rent metal by the hour.

What bare metal costs beyond the invoice

The price in the table is the easy part. Bare metal makes you own three problems a cloud provider usually hides.

Failover. The second box is in the table. What isn't: copying data between the two, something to route traffic, and practicing the switch-over. Local NVMe is fast partly because nothing replicates it for you.

Hardware breaks. Backblaze's fleet of 344,196 drives failed at 1.36% a year in 2025.¹⁰ When you rent, the provider swaps the dead disk, and if the disk wasn't mirrored, your service waits until it does. When you own the servers, it's your job: Ahrefs runs about 3,300 servers with a team of eleven and logged 94 hardware issues in one month.¹¹

Lead time. A cloud VM is up in minutes. Buying your own servers or data center space can take two to nine months, Ahrefs says.¹¹ Renting is faster, but ask for the lead time before you plan a launch around it.

The big savings stories are real. 37signals cut its cloud spend, not counting S3 storage, from about $180,000 a month to under $80,000 after buying about half a million dollars of servers, with the same ops team.¹² Ahrefs puts its own servers at $122 million over six years, against about $1.1 billion for the cheapest AWS option.¹¹ Both own their hardware and run it at scale, so treat those numbers as the best case, not your forecast.

Compliance doesn't require bare metal

Neither HIPAA nor PCI DSS makes you buy dedicated hardware. HHS says the HIPAA Rules "do not endorse or require specific types of technology," and a covered entity can use a cloud provider under a business associate agreement.¹³ The PCI Security Standards Council's cloud guidance accepts shared cloud hardware when tenants are properly segmented; separate physical servers are one way to do that, not a requirement.¹⁴ Bare metal doesn't make you compliant on its own.

Which one to pick

Start with the workload, not the price list. Does it run near full load all month, at a size and in a place that won't change for a year? A production database or a steady API tier, yes. A job that runs for an hour a night, traffic that jumps tenfold on Black Friday, or a new region next quarter, no. If it's a no, keep it on cloud VMs and stop here.

If it's a yes, rented metal lists for less. What's left to decide is what happens when the hardware breaks:

• It can be down while a part is swapped. One box, $456 a month, against $671.25 for one of our VMs.
• It needs a spare running. A pair of boxes is $912, against $1,342.50 for two of our VMs. Still cheaper, but the copying, routing and switch-over from the section above are now your job.

Disk-bound workloads on network storage gain the most from the move: a lower price and a faster disk. CPU-bound ones only get the lower price, so weigh it against the work of running the hardware.

Both, from American Cloud

Our bet: most teams have a few machines that belong on metal and a lot that don't. We sell both, and we run hardware ourselves: we own and operate OpenColo, our data center in Santa Clara, with engineers on site around the clock and the same team running it for over 20 years. Our SOC 2 Type II covers both the cloud and the data center.

Our VMs bill by the hour or on a monthly term, with no egress fees and no long contract. You can price your VMs on our pricing page. When something breaks, our support team's median reply time is under 2 minutes.

GGet started

For the steady machines, we build bare metal servers to order: you tell us the workload, we size the hardware and quote the build. Once you approve the quote, the server is running within 24 hours.

BBook a call

If you're moving off AWS first, here's how to move your workloads off AWS. If you'd rather put your own hardware in a rack, read colocation vs. public cloud in 2026.

Sources and dates

Footnotes

1. AWS EC2 on-demand, US East (N. Virginia), Linux: c6i.8xlarge (32 vCPU, 64 GiB) at $1.36 an hour, times 730 hours = $992.80. EBS gp3 at $0.08 per GB-month, 3,000 IOPS and 125 MB/s included: 500 GB = $40. AWS price list files for EC2 and EBS, read October 10, 2026. ↩

2. AWS EC2 Instance Savings Plans for c6i.8xlarge, US East, Linux, shared tenancy: 1 year no upfront $0.89964 an hour; 3 years no upfront $0.59875; 3 years all upfront $0.52114 (the upfront payment spread over the term; $0.52114 x 26,280 hours is about $13,700 paid on day one). Each times 730 hours, plus $40 of gp3. AWS Savings Plans pricing, price list file read October 10, 2026. ↩ ↩² ↩³

3. American Cloud Standard (Intel) VM, 32 vCPU, 64 GB RAM, 500 GB SSD: $10 per vCPU, $5 per GB of RAM and $1 per 16 GB of disk a month = $671.25 on a monthly term. By the hour it's 2% more, $0.9509 an hour: $684.68 for our 720-hour billing month, and $694.18 over the 730 hours used here. No egress charge. American Cloud pricing, October 10, 2026. ↩ ↩²

4. Latitude.sh m4.metal.medium: AMD EPYC 9124 (16 cores, 32 threads), 128 GB RAM, 2 x 480 GB plus 2 x 1.9 TB NVMe, "Starting At $0.62/hr $456/mo", 20 TB of egress included; the page doesn't say which location the starting price is for. Latitude.sh pricing, October 10, 2026. Two boxes: $912. A second listing at the same level: OVHcloud Advance-5 (AMD EPYC 8224P, 24 cores, 48 threads, from 96 GB) from $467 a month plus a $467 installation fee, OVHcloud, October 10, 2026. ↩ ↩² ↩³ ↩⁴

5. AWS data transfer out to the internet from US East: first 100 GB a month free, then $0.09 per GB for the next 10 TB. 5 TB = 4,900 GB x $0.09 = $441. AWS EC2 pricing, price list file read October 10, 2026. ↩

6. AWS, "Instances built on the AWS Nitro System", read October 10, 2026. ↩

7. c5n.18xlarge against c5n.metal on HPC codes: "The differential in all of the evaluated cases is within 1% of the performance level." AWS HPC Blog, "Bare metal performance with the AWS Nitro System", 2021. ↩

8. Brendan Gregg, "AWS EC2 Virtualization 2017: Introducing Nitro", November 29, 2017. ↩

9. Amazon EBS features: io2 Block Express is "designed to deliver an average latency of under 500 microseconds for 16KiB I/O operations"; gp3 and gp2 are "designed to offer single-digit millisecond latency." Read October 10, 2026. ↩

10. Backblaze Drive Stats for 2025, February 12, 2026. ↩

11. Ahrefs, "How Ahrefs gets a billion-dollar worth infrastructure with a 90% discount", May 8, 2024. The cost comparison covers July 2017 to June 2023; the 94 hardware issues were in June 2023. ↩ ↩² ↩³

12. David Heinemeier Hansson, "Our cloud exit has already yielded $1m/year in savings", September 15, 2023. ↩

13. HHS, Guidance on HIPAA and cloud computing (reviewed December 23, 2022) and FAQ on cloud access to ePHI (reviewed January 9, 2023). ↩

14. PCI SSC Cloud Computing Guidelines, version 3.0, April 2018. ↩