Bare Metal Servers: What They Cost, and When They Beat Cloud
Bare metal is a dedicated physical server rented from a provider, with no hypervisor and no other tenants on the machine. You get the whole box: every core, all the memory, the full network interface. It sits between public cloud, where you rent a slice of a machine, and colocation, where you own the hardware and rent the room. It is billed monthly rather than hourly, provisioned in days rather than seconds, and it is usually the cheapest way to run a steady workload without spending capital. Consider this your independent bare metal review, written by an advisory practice that places requirements with providers and operators across North America and internationally, is paid by the provider you choose, and does not sell servers.
Summary: For a workload that runs at consistent utilization, bare metal is typically 25 to 40 percent cheaper than equivalent cloud compute, and the saving arrives within a quarter rather than a year. It requires no capital, no procurement cycle and no data center contract. What it does not do is scale down when you are not using it, replace itself in seconds when hardware fails, or come with the managed services a cloud bill includes. Whether that trade works depends on how steady the workload is and how much of your stack depends on services you would have to replace. Almost every page written about bare metal is written by a company selling it. This one is not.
If you are working out whether bare metal fits, tell us what you are running. You do not need a defined requirement. Workload, rough scale, and what you are on today is enough. It comes to the principal directly and you will have a first read within 24 hours on whether the shape is right and what it should cost. If it is not the answer, we will say that too.
What is bare metal, and how is it different from cloud?
A bare metal server is a single physical machine allocated to one customer. Nothing is virtualized, nothing is shared, and no hypervisor sits between your software and the hardware. The provider owns the machine and the facility, handles replacement when something fails, and bills you monthly.
Public cloud is the same hardware split among many customers by a hypervisor. That layer is what makes elasticity possible, and it is also what you pay a premium for. It costs a few percent of raw performance and introduces variability from other tenants on the same host, which is the noisy neighbor problem. For most applications that variability is irrelevant. For a database under sustained load, or any system where tail latency matters more than average latency, it is not.
The bare metal vs cloud comparison, and the bare metal vs VM comparison underneath it, comes down to three things. Performance is better and more consistent on bare metal because nothing is between you and the silicon. Cost is lower for steady utilization and higher for bursty. And flexibility is worse: a virtual machine can be resized in minutes, where a bare metal server is the machine you ordered until you order a different one.
| Public cloud | Bare metal | Colocation | |
|---|---|---|---|
| What you rent | A virtual slice of a machine | The whole physical machine | Space, power and cooling |
| Who owns the hardware | The provider | The provider | You |
| Billing | Hourly or per second | Monthly or annual | Monthly per kW, plus hardware capital |
| Time to production | Seconds | Days to weeks | Nine to fifteen months |
| Capital required | None | None | Hardware, amortized three to five years |
| Scales down | Yes | No | No |
| Hardware failure | Rescheduled in seconds | Provider replaces, hours | Your spares, your problem |
| Managed services | Extensive | Minimal | None, you or an MSP |
| Relative cost, steady workload | Highest | 25 to 40 percent lower | Lowest over three years |
The practical rule: cloud rents you a slice, bare metal rents you the machine, colocation rents you the room. Moving down that list costs less per month and asks more of you in commitment and operations.
How we get you a better deal
Four things, and they apply whether the answer is bare metal, colocation or both.
1. We will tell you when bare metal is the wrong answer.
A provider selling bare metal will tell you it fits. Sometimes it does not: the workload is too variable, the stack depends on managed services, or the footprint is large enough that owning hardware in wholesale colocation beats renting it. We say so, because we are not selling either one.
2. We put providers in competition.
A requirement that goes to one provider gets one price. The same requirement in front of three or four who know they are competing gets a different price and different terms. Making the responses comparable is most of what a placement is.
3. We know what they actually contract at.
Published bare metal pricing is a rate card. What a provider will sign at volume, on term, with a defined refresh cycle, is a different number, and the gap is what you negotiate against. Data Center Knowledge used the practice as its source in August 2026 on the same point: what operators say and what they will sign are two different lists.
4. It costs you nothing.
The provider you choose pays the placement fee as a standard part of its channel program. Going direct does not save it; it leaves it with the provider.
That is the placement, and it is free. Above roughly $2 million a year of cloud or dedicated spend there is usually a question underneath it: which workloads should move at all, what the P&L looks like under each path, and whether the destination is bare metal, colocation or both. That is a fixed-fee engagement paid by you, four to six weeks, and it produces the specification the placement then runs. The two are separate and the consulting fee does not depend on the placement.
Either way, tell us what you are running. We work across North America and internationally, and the first conversation is what fits and what it should cost.
When does bare metal beat cloud?
Four conditions, and the more of them a workload meets the stronger the case.
1. Consistent utilization. Anything running above roughly 60 percent around the clock. Databases, data warehouses, core application tiers, scheduled batch processing. Cloud is priced for the ability to scale down, and a workload that never scales down pays for something it never uses.
2. Latency or performance sensitivity. No hypervisor means lower and more consistent per-request latency, and no noisy neighbors means the tail latency improves more than the average. For real-time bidding, trading systems, and high-throughput APIs, the consistency is often worth more than the cost saving.
3. High outbound data transfer. Cloud egress runs roughly 5 to 9 cents per gigabyte at volume. Bare metal providers either include transfer or price it at a fraction of that. For an egress-heavy workload this line alone can justify the move, which is covered in more depth in the cloud repatriation guide.
4. Predictable growth. If you know roughly what you will need in twelve months, you can commit to a term and price accordingly. If the shape of the workload is still unknown, cloud’s flexibility is worth the premium.
Where bare metal is the wrong answer: workloads that spike ten times against baseline, anything deeply built on proprietary managed services, applications needing presence in twenty regions simultaneously, and small footprints where the operational overhead exceeds the saving.
A worked example at enterprise scale
The numbers below are illustrative, built for a company running roughly $4 million a year of steady production workload with meaningful outbound traffic. They are the shape of the comparison rather than any client’s figures.
| Line | Cloud, already optimized | Bare metal |
|---|---|---|
| Compute, 200 machines at 32 cores and 128 GB, running 24/7 | $164,000 | $130,000 |
| Outbound data transfer, 2 PB a month | $118,000 | $18,000 |
| Storage | $26,000 | $9,000 |
| Managed services: load balancing, monitoring, managed databases | $34,000 | 0 |
| Operations and platform, replacing what was managed | 0 | $38,000 |
| Migration, amortized over year one | 0 | $14,000 |
| Monthly total, year one | $342,000 | $209,000 |
| Annual | $4.1 million | $2.5 million |
Roughly $1.6 million a year, or 39 percent, and about $1.8 million once the migration cost drops out in year two. Four things are worth noticing.
- The compute saving is the smallest part of it. Around twenty percent on the servers themselves, and that is against optimized cloud rather than list. Anyone quoting a much larger compute saving is comparing against on-demand pricing you should not be paying anyway.
- Outbound transfer is the largest single line. It falls by around eighty-five percent, because cloud prices egress per gigabyte while bare metal providers price bandwidth as a port and a commit. For a workload that answers requests at volume, this line alone often decides the question, and it is covered in more depth in the cloud repatriation guide.
- The operations line is real and it is where in-house models go wrong. Thirty-eight thousand a month for the platform layer you now run yourself: load balancing, monitoring, the databases someone used to manage for you. At this scale that is a team or a managed service provider, not a rounding error. Leave it out and this comparison shows $2.1 million a year instead of $1.6 million, and the difference turns up in month four.
- At this size the colocation question is live. A 200-machine footprint is roughly 500 kW to 1 MW, which is the point where owning hardware in a wholesale colocation lease beats renting it over three years. The usual sequence is bare metal first, because it lands the saving in a quarter and proves the workload runs on dedicated infrastructure, then colocation for the second wave once the profile is known.
If your numbers look anything like that, send us the shape of the problem. Annual spend, what you are running, and roughly what your transfer looks like. That is enough for a first conversation, and if bare metal does not fit, we will say so.
What does a bare metal server cost?
Behind the example above, this is what individual machines cost. Bare metal server pricing is quoted monthly per machine and depends on the processor generation, memory, storage and network. The ranges below are observed provider rate cards for a single server in 2026, before bandwidth and add-ons. There is no published benchmark for bare metal the way CBRE and Colliers publish colocation rents, because every provider prices its own inventory:
| Configuration | Typical monthly | Suits |
|---|---|---|
| Entry, single modern CPU, 64 to 128 GB RAM, NVMe | $150 to $400 | Web tiers, small databases, CI, staging |
| Mid, dual CPU, 256 to 512 GB RAM, NVMe array | $400 to $1,200 | Production databases, application tiers, caching |
| High memory or high core count, 1 TB+ RAM | $1,200 to $3,000 | In-memory workloads, analytics, large data platforms |
| Storage-dense, high-capacity disk arrays | $500 to $2,500 | Object storage, backup targets, data lakes |
| GPU, single to eight accelerators | $1,500 to $15,000+ | Inference, training at small scale, rendering |
Three things move the number more than the specification does.
Term. Month to month is the most expensive way to buy bare metal. A one-year commitment typically takes 10 to 20 percent off, and three years more than that. Providers price this way because their own economics are capital amortization, and a longer commitment de-risks it.
Volume. A single server is bought off a rate card. Twenty servers is a negotiation, and the discount at that point is meaningful. Most of the published pricing you see assumes you are buying one.
Bandwidth. This is the line that catches people. Some providers include a generous transfer allowance; others meter it and the bill looks very different at volume. For an egress-heavy workload, the bandwidth terms matter more than the server price, and it is the first thing to compare rather than the last.
Against cloud, the comparison worth running is dedicated server cost against the equivalent instance running 24/7 at list or committed-use pricing. For a steady workload the bare metal number is usually 25 to 40 percent lower. Against a workload that idles half the day, cloud often wins.
Who are the bare metal providers?
Bare metal server providers fall into three groups, and they behave differently.
The specialists.
Companies whose core business is bare metal hosting and dedicated servers. They have the deepest inventory, the fastest provisioning, and the most competitive pricing at volume. This is where most enterprise bare metal is bought, and where a competitive process pays off most.
The colocation operators with a bare metal line.
Several of the operators covered on this site sell managed bare metal alongside space and power. The advantage is that a hybrid of rented servers now and owned hardware later can sit in the same facility, on the same cross-connects, with one contract. For a company that expects to grow into wholesale colocation, that path is worth pricing.
The bare metal cloud providers.
AWS, Azure and Google all sell bare metal instances, and several specialists market their product as bare metal cloud, meaning dedicated hardware with cloud-style provisioning. The hyperscaler version is convenient if the rest of your estate is there and rarely the cheapest; the premium buys integration with the managed services around it. The specialist version is usually the better price.
The practice holds channel relationships across the first two groups. Which fits depends on the workload, the scale and whether colocation is in the picture later, and that is the first thing a placement establishes.
Bare metal or colocation: which dedicated path?
Both give you dedicated hardware. The difference is who owns it, and that decides everything else.
- Bare metal is faster and needs no capital. The hardware already exists, so provisioning takes days. The provider handles failure and replacement. You can test the move on a real workload without a procurement cycle or a board approval. For a company that needs a result before a fiscal year end, it is usually the only path that gets there.
- Colocation is cheaper over three years and slower to reach. You buy the servers, so the monthly cost is space and power rather than rental, and the saving is larger. But procurement, delivery, racking and burn-in run nine to fifteen months, and you carry the refresh cycle and the operations.
- The crossover is roughly $50,000 a month of dedicated spend. Below that, the operational overhead of owning hardware usually eats the saving. Above it, the three-year economics start to favor ownership, and by the time a footprint is several hundred kilowatts it is a wholesale colocation conversation rather than a server rental.
Many companies do both in sequence. Bare metal for the first wave, which lands the saving in a quarter and proves the workload runs on dedicated infrastructure. Colocation for the second, once the profile is known and the capital case is easy to make. That turns one large bet into two smaller ones.
What about GPU and AI workloads?
Bare metal GPU exists and it works well for inference, small-scale training and rendering. A single server with one to eight accelerators is a standard product from most providers, and for a workload of that size it is the right answer.
Rack-scale AI is a different market. A GB300 NVL72 system draws 140 to 160 kW per rack and needs direct-to-chip liquid cooling, and there is effectively no bare metal rental product at that scale. Companies deploying rack-scale hardware buy it and place it in a liquid-cooled colocation facility, which is why the AI and GPU colocation page covers what a facility has to have.
So the honest split: single-server and multi-GPU requirements are a bare metal conversation, and anything rack-scale is a colocation one. Anyone telling you otherwise is selling something they cannot deliver on your timeline.
What should you negotiate in a bare metal contract?
Bare metal contracts are shorter and simpler than a data center lease, but four terms carry real money.
Bandwidth and transfer.
Included allowance, overage rate, and whether it is metered on the 95th percentile or on total transfer. For an egress-heavy workload this is frequently larger than the server cost, and it is the term providers are least likely to volunteer.
Hardware refresh.
A three-year commitment on a server that is already two years old is a worse deal than the price suggests. Establish the generation you are getting and what happens at refresh.
SLA and replacement time.
What the provider commits to on hardware failure, in hours, and what the remedy is when they miss. This is the operational difference between bare metal and cloud, and it should be written down.
Exit and expansion.
What it costs to add machines mid-term, and what happens if you need to leave early. A provider who wants a three-year commitment should be flexible on adding capacity inside it.
What this page cannot tell you
It cannot tell you which of your workloads are steady enough to move, because that depends on utilization data you have and we do not. It cannot tell you what a provider will actually quote for your configuration at your volume, because that is negotiated rather than published. It cannot tell you whether bare metal or colocation is right for your footprint without knowing the footprint. And it cannot tell you what your egress line is, because that is on your bill.
Those are the analysis and the placement. This page is the framework they use.
How the practice works on bare metal
Metro Colo Advisory is an independent data center advisory practice. There are two sides to it and they are paid differently.
Placement.
You give us the requirement. We take it to providers who can serve it, run the process so the responses are comparable, and negotiate the terms that matter rather than just the monthly rate. The provider you choose pays us as a standard part of its channel program. You pay nothing. If bare metal is not the right answer, we tell you that.
Consulting, when the requirement is not yet defined.
Sometimes the question is earlier: which workloads should move at all, what the P&L looks like under each scenario, and whether the right destination is bare metal, colocation or a hybrid. That is the consulting side, paid by you at a fixed fee, and it produces the specification the placement then runs.
The two are separate. The consulting fee does not depend on the placement, and the placement does not depend on having done the consulting.
Frequently Asked Questions
What is a bare metal server?
A single physical server rented from a provider and allocated to one customer, with no hypervisor and no other tenants on the machine. You get every core, all the memory and the full network interface. The provider owns the hardware and the facility and handles replacement when something fails. It is billed monthly rather than hourly and provisioned in days rather than seconds.
What is the difference between bare metal and a virtual machine?
A virtual machine is a slice of a physical server, created and managed by a hypervisor that allows many customers to share one machine. Bare metal is the whole machine with no hypervisor. The practical differences are performance, which is better and more consistent without the virtualization layer, cost, which is lower for steady workloads, and flexibility, which is better on a VM because it can be resized in minutes.
How much does a bare metal server cost?
Typically $150 to $400 a month for an entry configuration, $400 to $1,200 for a mid-range dual-processor machine, and $1,200 to $3,000 for high-memory or high-core-count servers. GPU configurations run from $1,500 to $15,000 or more. Term length and volume move the number more than the specification does, and bandwidth is priced separately by most providers.
Is bare metal cheaper than cloud?
For a workload running consistently above roughly 60 percent utilization, usually yes, by 25 to 40 percent on equivalent compute. For bursty workloads that scale down for part of the day, often no, because cloud only charges for what runs. The comparison has to be against optimized cloud pricing rather than list, and it has to include bandwidth on both sides.
What is bare metal as a service?
Bare metal as a service is dedicated physical hardware delivered with cloud-like provisioning: an API, automated deployment, and hourly or monthly billing without a long procurement cycle. It is the same underlying product as a traditional dedicated server, packaged so it can be spun up and torn down programmatically rather than ordered by ticket.
Can you run Kubernetes on bare metal?
Yes, and it is common. Bare metal Kubernetes removes the virtualization layer between the container runtime and the hardware, which improves performance and removes a licensing and management layer. What you take on is the networking and load balancing that a cloud provider’s managed Kubernetes handles for you. For a steady, large cluster the economics are usually favorable; for a small or highly variable one they are usually not.
What is the difference between bare metal and colocation?
Bare metal is a server you rent, where the provider owns the hardware and the building. Colocation is space and power you rent for hardware you own. Bare metal is faster to deploy and needs no capital; colocation is cheaper over three years and gives you control of the asset. The crossover is usually around $50,000 a month of dedicated spend.
How long does it take to deploy bare metal?
Days to a few weeks for standard configurations, because the hardware already exists in the provider’s inventory. Custom specifications or large volumes take longer, since the provider may need to procure. Migrating a live workload onto it is the longer part of the timeline: typically three to six months for a production system running in parallel during cutover.
Do I need a broker to buy bare metal?
You do not need one to sign a contract. You need one if you want competitive pricing and to know which terms move. A provider quoting into no competition quotes their rate card. The same requirement in front of three providers produces a different number, and the provider pays the placement fee either way, so the cost to you is nothing.
Related reading
The decision before the server: the cloud repatriation guide for whether to move off cloud at all, and the cloud versus colocation calculator for a first read on your numbers.
The alternative path: wholesale colocation for footprints above a few hundred kilowatts, what colocation is for the basics, and the colocation pricing guide for what facilities charge.
Specific workloads: the AI and GPU colocation page for rack-scale requirements, the hybrid cloud guide for the architecture most companies land on, and the disaster recovery guide.
Contracting and moving: the contract terms guide and the migration guide. For how the placement side works, see how it works and why companies use an independent advisor. Or compare all providers.
Send us the requirement
The workload, roughly what scale, what you are running on today, and whether you expect to grow into owned hardware later. That is enough for a first read on whether bare metal fits and what it should cost, and you will have it within 24 hours.
North American and international coverage. Paid by the provider you choose. If nothing fits, we say so.
