What is modular compute?
Modular compute is computing infrastructure delivered in self-contained, repeatable units. Each unit packages the hardware that runs AI workloads—GPUs, CPUs, networking and storage—together with the power distribution, cooling and monitoring it needs. Units are integrated and tested in a factory, shipped to a site and connected to power and fiber. Capacity grows by adding units, not by building a new hall.
In practice, modular compute is the compute side of a modular data center. A modular data center describes the facility; modular compute describes the outcome a buyer pays for: a defined amount of GPU capacity, ready for service on a known date, in a block that can be repeated.
The idea is not new. High performance computing centers and cloud providers have used standard pods for years. What changed is the AI era: rack power for AI infrastructure rose from a few kilowatts to 40–150 kW and beyond, and grid constraints made the energized connection, not the building, the scarcest input. Modular compute is a response to both.
The building block: what a modular compute unit contains
A modular compute unit for AI workloads typically contains:
- Compute hardware: GPU servers or rack-scale systems such as NVIDIA B200, B300 and Vera Rubin, and AMD Helios, with CPUs and local storage.
- Networking: the high-bandwidth fabric between GPUs and the uplinks to the site network and carriers.
- Power distribution: switchboards, busway or panelboards and rack power, sized for the IT load and the cooling load.
- Cooling: direct liquid cooling to the chip, coolant distribution and heat rejection, usually dry coolers or chillers outside the module.
- Controls and monitoring: sensors, leak detection, fire protection and the management software that reports the unit’s state.
QIT plans around 1 MW of IT capacity per 20 ft container, phased into 5–10 MW sites. QIT M1 is chip-agnostic, so each block can rack NVIDIA B200, B300 or Vera Rubin, AMD Helios or other liquid-cooled accelerators.
Modular compute vs a traditional data center build
| Question | Modular compute | Traditional data center |
|---|---|---|
| Unit of growth | One block at a time, often 1 MW | A building or data hall, often tens of megawatts |
| Where integration happens | In the factory, tested before shipping | On site, after construction |
| What sets ready for service | Energized power, fiber and the pad | Construction, then power, then fit-out |
| Capital committed up front | Matches the first blocks of demand | Shell and shared plant for the final size |
| Best fit | Existing powered sites, phased AI capacity, edge and interim capacity | Very large campuses with long planning horizons |
Neither model removes the utility from the schedule. Where the transformer, switchgear or interconnection agreement is still open, those items take longer than any module. The modular advantage is largest on a site whose connection is already energized.
How modular compute scales
Scaling modular compute means adding blocks against usable power. Define whether a “1 MW” block is IT load or facility load: at a PUE of 1.2, each 1 MW IT block needs about 1.2 MW at the site boundary, so 5 MW of usable supply supports four whole blocks and 10 MW supports eight, before design margin.
Plan the shared infrastructure for the final size from day one: the utility service, main switchgear, fiber routes and space for heat rejection. Each later block then becomes a delivery and connection exercise, not a construction project. This is how 5–10 MW sites grow in phases while earlier blocks are already earning.
See the site selection checklist for how to qualify a site’s power before ordering blocks.
Where modular compute fits
- GPU clouds and neoclouds adding capacity for customer clusters while a larger campus is built. See capacity for GPU clouds.
- AI inference close to users, where an edge site needs dense compute without a full data center build.
- Owners of powered land, such as former bitcoin mining or industrial sites, converting existing service to AI. See bitcoin miners moving to AI and HPC.
- Enterprises and research teams that need dedicated GPU capacity for model training and deployment on their own sites.
- Sovereign and regulated workloads, where data must stay in a specific location.
Planning modular compute: power, cooling and economics
Power. Start with evidence of usable power: the service agreement, single-line diagram and a written utility response to the proposed load. Ready for service is measured from the energized connection.
Cooling. AI racks at these densities need liquid cooling. Check the coolant temperatures the GPU platform requires against summer design conditions at the site. QIT uses direct liquid cooling on every GPU and CPU with a sealed loop and dry coolers, so no water is consumed in operation. See the liquid cooling guide.
Economics. Compare options in dollars per kW-month over the period you will use the capacity, usually three to five years, with the same scope on each side: hardware, power, cooling, operations and energy. Capital for GPU infrastructure is recovered over years, so contract term, prepayment and the ability to redeploy a block matter as much as the build cost. The modular data center cost guide breaks down the main lines.
Modular compute vs Modular, the AI software company
Searches for “modular compute” also return Modular, the company behind the MAX platform and the Mojo programming language. That is a different subject. Modular builds AI software: an inference stack that runs AI models across NVIDIA, AMD and Apple silicon hardware, with tools for writing custom GPU kernels. Its products help developers and platform teams get performance from existing hardware.
Modular compute in the infrastructure sense is about the physical layer underneath: the GPUs, power, cooling and site that any AI software stack runs on. The two are complementary. QIT is not affiliated with Modular.
See modular.com for Modular’s software platform.
Common questions about modular compute
Is modular compute the same as GPU as a service?
No. GPU as a service sells access to compute by the hour or month. Modular compute is physical infrastructure that a GPU cloud, an enterprise or a site owner deploys and operates, or contracts someone to operate.
How fast can modular compute be ready for service?
On a site whose utility connection is already energized, factory-built units can be ready for service in weeks. Setting and connecting a QIT M1 on a prepared site takes about 15 hours; commissioning and network turn-up follow. Where power is not yet energized, the utility work sets the date.
Which GPUs does modular compute support?
That depends on the unit’s power and cooling design. QIT M1 is chip-agnostic: it supports NVIDIA B200, B300 and Vera Rubin, and AMD Helios and other liquid-cooled accelerators. These rack-scale platforms need direct liquid cooling and high rack power.
Can modular compute be moved?
Containerized units can be relocated, but disconnection, transport and recommissioning take planning. Treat relocation as an option, not a routine operation.
Where QIT fits
QIT builds liquid-cooled GPU containers: 1 MW IT blocks, integrated and tested in the factory, for phased 5–10 MW sites with existing power in Texas. We start from your site, its energized capacity and your workload, and plan to a ready-for-service date.
Read the GPU container specification and the deployment plan, or discuss a project.
Prepared by QIT · Updated November 2, 2026
Explore all field notes