Loading
Taking longer than expected.
Reload page

B3IQ Blog · Knowledge · 7 min read

Managed GPU Hosting vs Colocation: What Is Included?

Who owns the hardware, who runs it, and which responsibilities belong in the agreement when you choose managed hosting or colocation.

B3IQ Team

Fig. 1: Managed Hosting vs Colocation — white text over a blurred landscape.

TL;DR

  • Your contract determines who handles physical operations and who maintains the software stack. “Managed” and “colocation” do not guarantee a particular split.
  • With managed GPU hosting or colocation, you can own the machine while a provider houses it. With rented bare metal, the provider owns the machine. On-premises operation puts the machine in your facility.
  • B3IQ’s private fleet routes your keys only to your owned machines. Physical hosting does not remove your responsibility for application behavior.
  • Check the contract for hardware repairs, software support, and incident recovery. Verify any promised response time or capacity guarantee in writing.

Four ways to operate dedicated GPU hardware

The four models differ along two questions. Who owns the machine, and who handles its physical operation? Those answers tell you whether you need to buy hardware and who will deal with it when a component fails.

With customer-owned managed GPU hosting, you own the machine and pay a provider to host and operate it. Some services marketed as managed GPU hosting instead rent you provider-owned hardware, so confirm ownership separately. B3IQ places its private fleet in this category. You can review its ownership model and machine options. The private fleet guide describes how your account accesses its owned machines, but hosting and repair commitments still need to be checked in the applicable terms.

With traditional colocation, you own the machine and place it in a provider’s facility. The provider supplies facility services such as space and power, while you arrange hardware maintenance yourself or buy remote-hands support. A colocation contract may offer more help, so check its scope rather than relying on the label.

With a rented bare-metal GPU server, the provider owns the machine and operates the facility. You rent dedicated access without buying the hardware. If you run the machine on premises, you own it and handle both the facility requirements and physical maintenance, either yourself or through contractors.

These categories describe ownership and physical work. None, by itself, tells you who patches the operating system, manages models, or recovers an application after an outage.

Who does what: the responsibility breakdown

The facility operator supplies the space, power, cooling, and network connections, but each contract defines how much capacity you receive. Under colocation, you place your GPUs in a provider’s facility and arrange the services your machines need. Managed GPU hosting also puts the machines in a provider’s facility, though the provider may take on more operating work. With rented bare metal, the provider owns and houses the machine. On premises, you provide the facility services yourself.

Hardware ownership alone does not determine who pays for a replacement or performs a repair. Warranty coverage, insurance, and the service agreement affect both. With colocation, you generally arrange parts and use your staff or paid remote hands for work at the rack. A managed host may coordinate repairs on customer-owned hardware, while a bare-metal provider generally repairs its own equipment. On premises, your staff handles the physical work or calls a service contractor. For any owned machine, check who diagnoses failures, obtains warranty parts, pays for replacements, and can touch the hardware.

Software responsibilities need a separate line-by-line review. Ask who installs and patches the operating system, maintains the GPU runtime, loads model weights, and backs up data. Monitoring also needs a defined boundary. A provider may watch machine health without monitoring your inference service or restoring it after an application failure. Under colocation and on-premises operation, you normally plan those tasks yourself. Managed hosting and rented bare metal may include some of them, but the service label does not establish which ones.

B3IQ’s private-fleet guide documents a routing boundary, not a complete operating agreement. By default, your API keys route only to machines your account owns. If no owned machine can serve a request, the gateway returns an error rather than falling back to public capacity. You still need the applicable contract to establish who handles OS changes, backups, hardware incidents, and application recovery.

Where B3IQ's managed hosting draws the line

B3IQ’s model puts customer-owned machines in managed hosting, with B3IQ handling physical operations. The ownership page describes the offering, while the machines page shows the hardware options. You still decide what your application runs and what its workload requires. Owning a machine does not make B3IQ responsible for your application’s behavior.

B3IQ’s ownership page lists power and cooling, Oregon rack space, networking and bandwidth, monitoring and on-call support, physical security, machine insurance, and remote hands under Managed Hosting. Confirm the applicable contract’s limits and exclusions for each item. Provisioning a machine or providing model and runtime tools does not, by itself, establish who patches the OS, manages model versions, maintains backups, or restores service after an incident. Assign those tasks explicitly before you deploy.

B3IQ’s private-fleet documentation draws a separate line around access. In private use, your keys route only to machines your account owns. If no owned machine can serve a request, B3IQ returns an error rather than falling back to public capacity. Earn is an optional, explicit choice for a whole machine, and you can request a switch back to private use. While Earn is active, your keys cannot reach that machine and SSH is disabled. B3IQ schedules the change in either direction.

Questions to ask before signing a contract

Ask each provider to put its answers in the contract or an attached service schedule. A sales description may tell you what a service offers, but the signed terms determine what you can require.

  • Can the facility power and cool your machines? Ask the provider to confirm the power available to each machine and the cooling method it will use. If your machines need liquid cooling, ask who supplies and maintains the connections.
  • What network and storage service will you receive? Ask which connections, bandwidth charges, and storage services the price covers. Confirm who restores connectivity or access to stored data after an outage.
  • When can you reach your machines? Ask whether you can visit the site, what notice or escort the provider requires, and which tasks its remote hands staff will perform when you cannot visit.
  • What does support promise to do, and when? Ask whether a stated time measures the first response or the resolution of an incident. Confirm which incidents the provider handles, how you escalate them, and whether the commitment appears in the contract.
  • Who replaces failed hardware? Ask who diagnoses a failed GPU, obtains the replacement, and performs the swap. Check the hardware warranty and whether labor or spare parts carry separate charges.
  • Who owns and insures the machines? If you buy the hardware, ask when title passes to you and whose insurance covers damage or loss while the provider holds it. If you rent a dedicated GPU server, confirm whether you acquire any ownership rights at all.
  • What will leaving cost? Ask who pays to shut down, remove, pack, and ship your machines, and how long release can take. Check whether any advertised service level or reserved capacity guarantee appears in the signed terms, including its exclusions and remedy.

Matching the model to your deployment

If you need dedicated GPUs but do not want to buy machines, rented bare metal avoids a hardware purchase. It may suit a short project or demand that changes enough to make ownership difficult to plan. If you expect steady use and want to own the machines, compare the full cost of operating and eventually removing them before assuming ownership will cost less.

Your capacity to handle physical operations narrows the choice among ownership models. On-premises operation requires you to provide the space, power, cooling, and staff. Colocation moves the facility work to a provider, but you need to check what its staff will do when a machine fails. Managed GPU hosting may take on more of that work. The contract, rather than the service label, determines how much. You can review B3IQ’s current machine options if customer-owned managed hosting fits your needs.

Owning a machine does not remove your responsibility for the application running on it. Before choosing a model, get the physical support duties, software boundaries, and incident commitments in writing.