**tl;dr** - if you know what you are doing, and you understand the disadvantages, then it makes sense to move to bare metal once you are spending more than $50 a month on VPS hosting regardless if it is 1 VPS or many. 

## What you must know or want to learn

* how to run a hypervisor
* how to run your own offsite backups
* how to do a full restore of a VM on a different piece of hardware
* how to do your own monitoring of the server and the backups

## Disadvantages that you must be comfortable with

* you are responsible for all management of the server and VMs (no training wheels!)
* recovery time from hardware failure could be up to a day to setup and migrate a new dedicated server
* potentially lower quality support options
* potentially using older generation processors to save money
* lower cost dedicated server providers may take shortcuts to infrastructure like cooling and fire prevention

## Benchmark VPS vs Bare Metal

| Provider     | ? Big Three ? | Digital Ocean | Linode | Vultr    | OVH Eco   |
|:-------------|--------------:|--------------:|-------:|---------:|----------:|
| Name         |  2nd Cheapest | Premium Intel | Shared |  Regular | Dedicated
| Price        |    **$10-12** |           $28 |    $40 |      $80 |       $27
| Cores        |             1 |             2 |      4 |        6 |     **8**
| RAM (GB)     |             2 |             4 |      8 |       16 |    **32**
| Storage (GB) |         50-55 |            80 |    160 |      320 |   **480**
| Memory Speed |        58-78% |          100% |   100% |     265% |  **424%**
| IO Speed     |       52-144% |          100% |    90% | **260%** |  **258%**
| CPU Speed    |       43-173% |          100% |   240% | **328%** |  **325%**
| Web Runs     |        17-63% |          100% |   150% | **333%** |  **300%**
| Tasks        |        28-60% |          100% |   127% | **381%** |      195%

## Analysis

For $27 a month you can get 2x the performance and 4x the amount of memory and 3x the amount storage as you would get with a $40 VPS. However the other advantages of using a VPS provider outweigh the savings. If you are spending more than $50 a month on VPSs, then I think the savings of moving to a dedicated server is worth the tradeoffs. 


VPS providers don't discount the price per vCPU or price per GB of RAM as you increase the specs of your server. I think that makes sense when you are selling 4 or fewer vCPUs in a package, but they really should offer discounts when you get to 6+ vCPUs or 12+ GB of RAM. Without offering discounts at those levels, it just doesn't make sense to pay that much when you need that amount of processing power or memory. 

## Providers

For VPS providers I sampled Digital Ocean, Linode and Vultr. You won't be disappointed with any of these three providers as your VPS provider. They are all really good at pricing their options where you can expect the performance to be in line with the price charged. 

For dedicated servers I only tested OVH's Eco server in the Kimsufi line. Those are servers that are at least 6 years old, but still very capable machines. It is hit or miss what stock is available for the Kimsufi line, but when they have a server in the $20-30 a month range it is usually an excellent value. 

If you are looking for something more powerful or more modern, you will likely find better deals for $35+ a month at Hetzner, but only in their European data centers. 

## Benchmark Explanation

All benchmark data came from [VPSBenchmarks](https://www.vpsbenchmarks.com/). I paid for the test of the dedicated server, the data for the VPSs are from tests that the site makes public. I chose to normalize the performance metrics to be in comparison with the 2 vCPU Digital Ocean VPS to make it easier to see the difference in performance in % terms. The raw data is posted in the [comments](https://www.reddit.com/r/webdev/comments/xyuq5s/comment/irirckd/). 

As you are most likely leasing a server to host a website, the two most important metrics is the Web Runs and the Tasks. 

Web Runs is how many simultaneous requests the machine can handle at a time to host a Rails app that VPS Benchmark uses. The Digital Ocean server can handle 30 simultaneous requests at a time without error, but anything beyond that and some requests will fail. The dedicated server can handle 3x that, or 90 simultaneous requests. This is a good real world test that yields results based on a combination of the CPU speed and the RAM (about 60/40 respectively).

Tasks is how many high CPU utilization tasks per hour can be processed by the machine in a particular VPS Benchmark test. The number of cores has the highest correlation to the result. Note that the OVH dedicated server only gets about half as many tasks done per hour as the number of cores would suggest is possible, which I attribute to the age of the processor. 


? Big Three ? - this is a range of performance from low to high out of Digital Ocean, Linode, and Vultr. I provide this as a range since the point of this post is not about low-end VPSs, but many people may be curious about the results of cheaper options. If you are interested in the cheapest options ($5-6) then you can find that information at VPS Benchmark.

## Tutorial

I'm in the process of writing a step by step tutorial on how to setup and run your dedicated server with multiple virtual machines with a single IP address (a limitation of Kimsufi servers). I'll post that on this subreddit soon!
I can think of three scenarios where it makes sense to use bare metal

* You've been hired by a company to build an on-prem cloud

* You've been hired by a cloud provider to architect their cloud

* You want to learn to how get hired for the above

For web developers, bare metal makes no sense, ever. I could practically fill a book with the reasons, but a primary one was already provided:

> recovery time from hardware failure could be up to a day to setup and migrate a new dedicated server

From a competitive standpoint, that's fantastic news for me, running on VPS/Cloud

Because on a properly setup VPS, recovery is _automatic_ when underlying hardware fails. I might get an alert as I go about my day instead of stopping everything I'm doing, because the entire company is bleeding money, while I rebuild another single point of failure in a panic

I'll gladly tell the client who was just down for the whole day, that I build with redundancy, automatic failover, automatic recovery, with backup trimming and restoration so simple they can do it themselves with a few clicks

You go right ahead and espouse the benchmarks which will have no meaningful performance gain on the end user

I'm not usually this harsh, but this hasn't been a reasonable proposition for at least a decade, and reeks of an agenda. I'm gonna guess your second post will more strongly push the brand you mentioned; and when you do, it will get reported as spam.