The short answer
Every vendor on this page builds modular infrastructure for AI. They differ in where they start. Most start from the module: a large prefabricated unit that still needs a utility connection, a transformer and a site build. Others start from their own cloud and sell capacity, not infrastructure.
QIT starts from the site. We look for power that is already energized, such as a bitcoin mine with spare transformer capacity, and size the compute to that power in 1 MW blocks. That changes what sits on the critical path.
- Fastest to live compute on an existing site: 15-hour on-site installation once power and pad are ready.
- No new utility service required where existing capacity can be reused, as in QIT Pod 001.
- Liquid cooling on every GPU and CPU, designed around the rack that is shipping today, NVIDIA GB300 NVL72.
- You own the GPUs. No cloud lock-in and no resale of your capacity.
- 1 MW blocks that fit stranded and partial capacity that multi-megawatt units cannot use.
Side-by-side comparison
| Company | What you get | Cooling | Unit size | Who owns the GPUs |
|---|---|---|---|---|
| QIT | Liquid-cooled GPU container on a powered site, installed in 15 hours | Direct liquid on every GPU and CPU | 1 MW IT per container | You |
| Crusoe Spark | Prefabricated AI module, primarily to deliver Crusoe Cloud capacity (Edge Zones) | Closed-loop direct liquid (Spark Liquid) | Up to 3.5 MW per unit | Mostly Crusoe; you rent compute |
| Armada | Ruggedized Galleon modules plus edge software | Direct-to-chip, ~80/20 liquid to air (Leviathan) | 1.77 MW (Leviathan); 10 MW (Orion) | Customer or operator |
| Giga Energy | Power equipment, GigaPod builds and colocation on its own land | Air-to-liquid with hot-aisle containment | Not published per GigaPod | Customer, or Giga colocation |
| Flexnode | NX units under one design-to-operation contract | Closed-loop, water-free | 1–18 MW per unit | Owned or leased |
| Vertiv | SmartMod, MegaMod CoolChip and OneCore prefabricated systems | Air or direct-to-chip | 200 kW to 5–50 MW blocks | Customer (equipment sale) |
| Dell | MDC Flex, Click and Micro modular data centers | Air; liquid available (Flex) | 5–100 racks at 5–115 kW | Customer (equipment sale) |
| Duos Edge AI | Edge pods it builds, owns and operates | Not published | 4.8 MW configuration | Duos; you buy colocation |
Sources: vendor pages and filings cited in our modular data center companies guide, including Crusoe, Armada, Giga Energy, Flexnode and Duos Technologies (SEC). Figures use each vendor’s own units. Reviewed November 3, 2026.
1. We start from power that already exists
For most AI projects, the module is not the bottleneck. The utility connection and the transformer are. A new-build pad-mount transformer can take 40–65 weeks, and a new utility service follows the utility’s own process. A module delivered in weeks still waits for both.
QIT designs around energized sites first. In QIT Pod 001, the preferred option reuses an existing 2,500 kVA transformer on an operating bitcoin mine; the alternative taps the host’s private 34.5 kV collector with a stock 2,000 kVA unit. Neither waits for a new utility service.
Module-first vendors can be installed on an existing site too, but their unit sizes and site packages are built for new capacity. Ours are built for the capacity you already have.
2. 15 hours on site, not a construction project
QIT containers are integrated and tested in the factory. On a prepared site, installation takes 15 hours. In the Pod 001 hosting design, the container is commissioned before the customer’s racks arrive: power, coolant, fire protection and network wait at every rack position, so the racks roll in, connect and go live.
Crusoe has quoted delivery “as fast as three months” for Spark; other vendors quote factory lead times and leave the site schedule to the project. Ask every supplier the same question: from arrival on site, how many hours until the first GPU runs?
3. Liquid cooling designed for the rack that ships today
An NVIDIA GB300 NVL72 rack dissipates about 135 kW, of which roughly 120 kW goes to liquid. That is far beyond what air or air-to-liquid retrofits handle well. Every QIT container uses direct liquid cooling with a cold plate on every GPU and CPU, and the coolant temperature, flow, voltage window, floor loading and service clearance follow the rack OEM’s envelope.
Several competitors offer liquid cooling as one option among air and hybrid layouts, or do not publish their cooling method. For a GB300-class deployment, liquid is not an option. It is the design.
4. You own the GPUs. No cloud lock-in.
Some of the best-known modular platforms are routes to a cloud. Crusoe Spark primarily delivers Crusoe Cloud capacity; Duos Edge AI builds, owns and operates its pods and sells colocation. Both can be good choices if you want to rent compute.
QIT provides the powered, cooled and connected space. Your racks, power shelves, cold plates and scale-out network remain yours. You choose the GPU platform, the software stack and the customers, and you keep the margin.
5. 1 MW blocks that fit real sites
A 1.77 MW, 3.5 MW or 5–50 MW unit needs that much usable power on day one. Many energized sites have less: a retired mining container frees one transformer, not a substation. QIT’s 1 MW containers turn that partial capacity into compute now and add the next block when more power is freed or connected, phasing toward 5–10 MW.
Smaller blocks also mean smaller commitments. You do not buy cooling, switchgear and space for capacity you cannot power yet.
6. Open design, no single-vendor dependency
QIT publishes its reference design. The Pod 001 drawings name a basis of design for every system and accept equals, subject to the rack OEM’s requirements. If one CDU, switchboard or cooler has a long lead time, another qualified product takes its place without redesigning the pod.
Vertically integrated vendors control their own supply chain, which helps when their factories have capacity and hurts when they do not.
Where competitors are ahead
We would rather you hear this from us:
- Installed base. Crusoe reported more than 400 modular units deployed at the Spark launch; Giga Energy states more than 175 MW of data centers built. QIT is earlier-stage; Pod 001 is our first reference project.
- Single-unit scale. If you need 10–50 MW in one block on a new campus, Armada Orion, Flexnode or Vertiv OneCore are built for that.
- Own manufacturing of electrical gear. Giga Energy builds transformers and switchgear in its own factories.
- Turnkey cloud. If you do not want to own GPUs at all, a cloud offer such as Crusoe Edge Zones is simpler.
When QIT is the right choice
Choose QIT when you have, or can reach, energized power today and want to own the compute running on it:
- A bitcoin miner with transformers that can be freed for AI. See bitcoin miners moving to AI.
- A neocloud or GPU operator that needs a liquid-cooled home for GB300 NVL72 racks now. See QIT for neoclouds.
- A site owner who wants to grow in 1 MW steps toward 5–10 MW.
Read the vendor-by-vendor comparisons for Crusoe Spark, Armada, Giga Energy, Flexnode, Vertiv, Dell, Mod42 and Duos Edge AI. Then discuss your site with QIT.
QIT is independent of the companies named on this page. Competitor information comes from their public pages and filings and may have changed since review. Reviewed November 3, 2026.
Prepared by QIT · Updated November 2, 2026
Explore all field notes