Deal types
Colocation vs. bare-metal vs. compute-as-a-service: what your contract must cover
How ownership shifts across the three GPU procurement models, and what your contract must cover in each.
LegalBooks Legal Team
Attorneys licensed in the US (California & Nevada) and Canada (Ontario)
Aug 5, 2026 · 10 min
Colocation, bare-metal, and compute-as-a-service differ by who owns the hardware and who controls the facility - and that determines what your contract must cover. Colo contracts must nail facility specs and the power-billing basis. Bare-metal contracts must lock hardware spec, substitution, and clean title. Compute-as-a-service contracts must secure real capacity, tenant isolation, and portability.
The three models sit on a spectrum of control. In colocation, you own the hardware and rent space, power, and cooling. In bare-metal, the provider owns the hardware and rents you dedicated, single-tenant physical servers with direct hardware access. In compute-as-a-service (also called cloud GPU or GPU-as-a-service), the provider owns and manages everything and you consume capacity on shared infrastructure. The more you offload to the provider, the more your contract - not your ops team - has to protect you.
Colocation vs. bare-metal vs. compute-as-a-service, at a glance
| Colocation | Bare-metal | Compute-as-a-service | |
|---|---|---|---|
| Who owns the hardware | You | Provider | Provider |
| Who runs the facility | Provider | Provider | Provider |
| Tenancy | Single (your gear) | Single-tenant physical | Usually multi-tenant / shared |
| Pricing model | CapEx + rack/power MRC | Fixed monthly / committed | Per-hour or per-second + egress |
| What the contract must cover | Facility specs, power-billing basis, site control (SNDA) | Hardware spec, substitution, clean title, single-tenancy | Real capacity, isolation, SLA, portability |
| Signature failure mode | Paying for overhead (gross vs. IT load); facility can't deliver density | "Equivalent hardware" swap; a financier's lien on the GPUs | "Reserved" capacity reallocated; egress lock-in |
Ownership and control split drawn from GigaGPU, Atlantic.Net, and Inflect, 2026.
Which model fits - and why the contract, not the sticker price, decides risk
The economics follow utilization, not the headline rate. Compute-as-a-service runs several times the per-hour cost of dedicated hardware - one comparison puts cloud GPU at roughly three to ten times dedicated pricing (GigaGPU) - but it wins for short, spiky, or experimental work because you pay only for what you use. Bare-metal and colocation win for sustained, high-utilization workloads: an 8x H100 cloud instance at about $80–100 USD per hour crosses over against bare-metal at roughly $12,000–20,000 USD per month once sustained utilization reaches about 40–65% (Inflect, 2026). Colocation reaches the lowest long-run per-unit cost but front-loads capital - reported at roughly $15,000–40,000 USD per GPU server - and generally only pays off at 20-plus servers with a dedicated operations team (GigaGPU).
Whichever model fits, the contract is where the real risk lives, because each model hands the provider a different piece of your stack. The rest of this article covers what your contract must cover in each.
What your colocation contract must cover
In colocation, you own the hardware, so the contract is about the facility delivering what it sells and billing you correctly for power. The two terms that move the most money are the facility specification and the power-billing basis.
Facility specification, as built. Tier and redundancy (Tier III / N+1 / 2N), per-rack power density, and cooling verified against commissioning documents - not the marketing label. This is deep enough that we cover it separately in the pre-signing checklist for a colocation MSA.
The power-billing basis. Whether you are billed on total facility load or on IT load changes the number materially: billing on gross facility load means you pay for the building's overhead, not only your servers. Define the basis, and decide who underwrites power-price increases (a cap, a named index, and transparency into the calculation).
Site control. If the provider leases the building, get a subordination, non-disturbance, and attornment (SNDA) agreement so a foreclosing lender cannot disturb your deployment, and match the ground-lease term to your term.
Access and exit. Remote-hands terms, and a clean right to remove your hardware at the end without hostage-style fees.
Signature failure mode: paying for the facility's overhead through a gross-load billing basis, or discovering the plant cannot cool your rack density after signing.
What your bare-metal contract must cover
Bare-metal is where ownership and control split most dangerously, so this is the model to read closely. The provider owns the hardware; you get dedicated, single-tenant access to it. That means two risks land on you that do not exist in colocation, where you own the gear: the provider can change what hardware you actually get, and the provider's lender may have a claim on it.
Lock the hardware specification. Tie the contract to a named GPU model, count, memory, and interconnect (for example, the specific accelerator, node configuration, and fabric) - not a category like "high-performance GPU." A category buys you whatever the provider chooses to deliver.
Define substitution - do not accept "equivalent hardware" undefined. An "equivalent hardware" clause lets the provider swap the servers you contracted for with hardware it calls equivalent. If "equivalent" is undefined, you can sign for one GPU generation and receive a slower one at the same price. Require that any substitution be equal-or-better against a stated benchmark, with named comparable models, and gated by notice, your consent, or a price adjustment.
Confirm clean title to the GPUs. Because the provider owns the hardware, it may have financed it - and a financier's security interest can put the very servers you depend on at repossession risk if the provider defaults. Run a UCC-1 (US) or PPSA (Canada) lien search, require a title-and-no-undisclosed-encumbrances representation, and negotiate lender non-disturbance and step-in rights. The full mechanics are in five clauses that quietly cost you in GPU-capacity agreements.
Guarantee single-tenancy. Bare-metal's value is dedicated hardware with no hypervisor and no "noisy neighbor" (Inflect, 2026). Put the single-tenant, no-shared-scheduling promise in writing.
Set a hardware-failure remediation time. Since the provider owns and services the iron, commit a node-replacement time and spare capacity rather than "commercially reasonable efforts" - the reasoning is in what's market in data-center uptime SLAs.
Data handling on decommission. A secure-wipe and return-or-destroy obligation when a server is swapped or the deal ends.
Signature failure mode: an "equivalent hardware" swap that quietly downgrades your cluster, or a financier repossessing the servers your workloads run on.
What your compute-as-a-service contract must cover
In compute-as-a-service you own nothing physical, so the contract has to secure everything you cannot hold in your hands: that the capacity is real, that your workload is isolated, that the service performs, and that you can leave.
A real capacity guarantee. "Reserved" capacity should be reserved - held for you, not reallocated to a higher-paying customer or oversubscribed. Instant provisioning often comes with no guaranteed availability during demand spikes (GigaGPU), so pin down whether you have committed capacity or best-effort access.
Tenant isolation and security. On shared, multi-tenant infrastructure, contract for isolation of your data and workloads, and the security and compliance posture you need (single-tenant options exist where isolation is non-negotiable).
A workload SLA with teeth. Availability measured at a layer that reflects your workload, remediation, and - where you have leverage - an exit for chronic failure. What's standard and what to push for is in what's market in data-center uptime SLAs.
Portability and exit. Cheap to enter, expensive to leave: nail data and model export in usable formats, egress costs (per-hour cloud pricing usually carries data-egress fees on top (GigaGPU)), and a transition-assistance period so a switch does not strand your workloads.
Signature failure mode: "reserved" capacity that gets reallocated when you need it, or egress and format lock-in that makes leaving cost more than staying.
Key takeaways
The model decides what the contract must cover. Ownership and control shift from you (colocation) to the provider (compute-as-a-service), and your protections have to shift with them.
Colocation: the facility and the power basis. You own the hardware, so verify the plant and define whether power is billed on gross or IT load.
Bare-metal: spec, substitution, and title. The provider owns the hardware, so lock the exact spec, define "equivalent" substitution, and confirm no financier can repossess the servers.
Compute-as-a-service: capacity, isolation, and portability. You own nothing physical, so the contract must guarantee real capacity, isolate your workload, and let you leave.
Match the model to utilization, not the sticker price. The cheaper per-hour rate is not the cheaper deal for sustained workloads - the crossover is around 40–65% utilization.
Frequently asked questions
What is the difference between colocation, bare-metal, and compute-as-a-service?
Colocation is renting space, power, and cooling for hardware you own. Bare-metal is renting dedicated, single-tenant physical servers the provider owns, with direct hardware access. Compute-as-a-service is consuming GPU capacity on the provider's shared, usually multi-tenant infrastructure, billed by usage. Ownership and control shift from you to the provider across the three.
Which is cheaper - colocation, bare-metal, or compute-as-a-service?
It depends on utilization. Compute-as-a-service is cheapest for short, unpredictable workloads but runs several times the per-hour cost of dedicated hardware; bare-metal and colocation win for sustained, high-utilization work. Reported crossover against an on-demand cloud instance sits around 40 to 65 percent sustained monthly utilization, so the right model follows how steadily you run, not the sticker price.
What must a bare-metal GPU contract cover?
Lock the hardware specification to a named GPU model, count, memory, and interconnect rather than a category. Define any substitution right - 'equivalent hardware' must mean equal-or-better against a stated benchmark, with a consent or price-adjustment mechanism. Confirm the provider owns the hardware free of liens, since a financier's security interest can put your dedicated servers at repossession risk. Add single-tenancy, a hardware-failure remediation time, and secure data wipe on decommission.
What is 'equivalent hardware' substitution and why does it matter?
An 'equivalent hardware' clause lets the provider swap the servers you contracted for with hardware it calls equivalent. If 'equivalent' is undefined, you can sign for one GPU model and receive a slower one at the same price. The fix is to define equivalence against a named benchmark (equal-or-better, with comparable models listed) and require notice, consent, or a price adjustment before any substitution.
What must a compute-as-a-service contract cover that colocation doesn't?
Because you own nothing in compute-as-a-service, the contract must secure what you can't hold physically: a real capacity guarantee (reserved capacity that can't be reallocated or oversubscribed), tenant isolation and security on shared infrastructure, a workload SLA with remediation, and portability - data and model export, egress costs, and transition assistance so you can actually leave.
Pick the model, then get the contract right
Choosing between colocation, bare-metal, and compute-as-a-service is only the first decision; the contract that matches the model is where deals are won or lost. LegalLayer is a tech-native legal team built for compute infrastructure: in just 2026 we have negotiated more than $266 million USD in client contract value; data-center work makes up the bulk, and drafts and redlines are back within six hours, guaranteed. We work every model and both sides of the table - as technical peers, licensed in the US (California and Nevada) and Canada (Ontario).
This article is general information, not legal advice. For advice on your specific situation, consult a qualified attorney. LegalBooks provides counsel for compute infrastructure deals across the US and Canada.