Modular Data Centers for Liquid-Cooled GPUs | Syaala

Modular Data Centers

Modular data centers
for liquid-cooled GPUs.

Three sizes, built and tested in the factory. Ready for service in 4 to 8 months against 18 to 24 for a conventional build, liquid cooled to 250 kW per rack, and added one block at a time as capacity is contracted.

The building block

1.6 MW Block

The step that repeats. Its own power, battery and cooling plant, sized for GB300 NVL72 class racks and added one block at a time.

IT capacity
1.6 MW
Rack positions
8 at up to 200 kW
Redundancy
N+1, or 2N
Land
0.08 to 0.15 acre

Land: modules and coolers, from the bare pad to the pad with its service perimeter. Transformer, access, setbacks and permitting vary by site.

Full specifications
Syaala 1.6 MW liquid-cooling block: IT, power, energy storage and pump modules with dry coolers
Speed and cost

Problem Power and buildings take years, and GPU capital loses value every month it waits.

Faster to site. Cheaper per MW.

Speed

75% faster

to site than a conventional build

3 to 6 months

from order to modules delivered on site

Modules are built and tested in the factory while the site is prepared, so the two run side by side instead of one after the other. Ready for service follows in 4 to 8 months from order.

Conventional build: average construction time, Turner and Townsend Data Centre Cost Index 2024, before any utility wait. Site gates such as utility energisation, carrier fibre and permits can pace the date, and are stated in the written site fit.

Cost

50% cheaper

per MW than industry-leading conventional builds

  • Factory acceptance tested
  • Tested again on site
  • NRTL evaluation per project
  • Designed to NEC and NFPA
  • Battery to UL 9540 and NFPA 855

  • Pay as you grow

    Capacity is added one block at a time as it is contracted, not a whole building ahead of demand.

  • Factory hours, not field hours

    Power, cooling and IT rooms arrive built and tested. Site work comes down to the pad, the connections and commissioning.

  • No cooling towers

    GPU heat leaves on a warm-water loop to dry or adiabatic coolers, with water used only for adiabatic assist at high ambient.

  • One design, repeated

    The same engineered block on every site: the design, the drawings and the test programme are reused, not redone.

What waiting costs

$0.8M to $2.0M

per MW, every month a GB300 cluster waits for power

Compared with conventional data center construction at $8 million to $12 million per MW of IT (Cushman and Wakefield, U.S. Data Center Development Cost Guide 2025). Syaala prices each project in writing after a site fit. Basis: GB300 NVL72 rack capital of $5.0 million to $6.5 million at 132 to 140 kW ($36 million to $49 million per MW of IT), a 10 to 15 percent cost of capital and a three to six year economic life. Indicative, not a quotation.

Density and compatibility

Problem GPU racks now draw 130 kW and more, far beyond what an air-cooled hall was built to carry.

Built around the rack, up to 250 kW.

Every size runs direct-to-chip liquid cooling on a warm-water loop, with a separate circuit for the small share of heat left in the air. Rack count and density are set within each size: fewer racks at higher density, or more at lower.

Matched to the rack

Designed for NVIDIA GB300 NVL72, HGX B300 and Vera Rubin NVL72 class platforms, including the rack-scale systems Supermicro, Dell, Lenovo and HPE build on them. On every project the design is checked against the rack maker's facility specification: power feed, liquid connections, coolant flow and supply temperature.

Configured to the project

  • Rack platform and density per position
  • Redundancy: N+1, or 2N on the 1.6 MW Block
  • Heat rejection selected for the site climate
  • Utility, site generation or bridge power at the module interface
  • Platforms mixed pod by pod in the 9 MW Zone
  • Building management and power monitoring built in

NVIDIA, GB300, HGX, Vera Rubin, Supermicro, Dell, Lenovo and HPE are trademarks of their owners. They are named to describe compatibility; no endorsement, certification or reference design programme is implied. Densities are indicative design values and vary with rack selection and GPU generation.

Servicing

Problem Dense modules are cramped: once a room is full, racks and power gear are hard to reach.

Ships narrow. Opens wide for service.

The IT and power modules of the 1.6 MW Block travel at standard shipping width on ordinary trucks, with no over-width permits. On site, a side extension opens each room to its full working width.

The side extension, in planIn transit the module is at shipping width with its rack positions empty. On site the extension opens beside it: racks stand in the fixed module with a clear front aisle, and a separate service aisle runs inside the extension.In transitRack positions,fitted on siteOn siteFront aisleRacksService aisleExtension opened on site
Zones only, not to scale. Illustrative of the 1.6 MW Block IT module.

  • Racks roll straight in

    A clear front aisle gives every rack a straight route in and out, with nothing overhead in the way.

  • Connections from behind

    Power and cooling connections sit on a separate service aisle inside the extension, away from the rack fronts.

  • Service while it runs

    Power modules, pumps and coolers are N+1, so any one can come out for service while the block carries its load. An N+1 block runs at N during maintenance; the 2N option keeps a full second path.

  • Batteries in their own module

    Energy storage sits in a separate module outside the IT room, designed to UL 9540 and NFPA 855.

The 4.8 MW B300 Block is serviced around a central spine: racks load from its two wings, cooling water runs under the floor and power overhead. The 9 MW Zone sits in a one-storey hall and is serviced like a data hall.

Build and test

Problem A stick-built site is tested only at the end, when every fault costs the most time.

Tested in the factory. Proven again on site.

The technical datasheet covers all three sizes: modules, rack densities, redundancy and site fit.

In the factory

  • A factory acceptance test for every module before it ships, on load banks and witnessed by Syaala
  • The test programme takes each power module, pump and cooler out in turn to show the load still holds
  • Third-party (NRTL) evaluation, arranged per project

On site

  • Staged energisation and commissioning, block by block
  • An integrated systems test with site generation and fibre
  • Site acceptance with your racks installed
  • A documentation package at handover

Designed to

  • NEC (National Electrical Code) and NFPA requirements
  • UL 9540 and NFPA 855 for battery storage
  • ASHRAE TC 9.9 thermal guidelines

Site gates such as utility energisation, carrier fibre and permits can pace the date, and are stated in the written site fit.

For site owners

Problem A 5 to 50 MW site is too small for a hyperscale campus, so good power goes unused.

Have a site with power? Put AI capacity on it.

The platform is sized for power-ready land from a few megawatts to about 50 MW. Time to power matters more than size, so the first question for any site is when it can be energised.

About 5 MW of IT

Three 1.6 MW Blocks, or one 4.8 MW B300 Block

10 MW and up

A 9 MW Zone, opened in thirds from 3 MW, or blocks added side by side

20 to 50 MW

Zones and blocks added in place, phase by phase, as capacity is contracted

Figures are IT capacity. Site power is sized above the IT load for cooling and losses, and the written site fit states it.

Land per unit

1.6 MW Block
0.08 to 0.15 acre Modules and coolers, from the bare pad to the pad with its service perimeter
4.8 MW B300 Block
0.17 to 0.26 acre Modules and coolers, from the bare pad to the pad with its service perimeter
9 MW Zone
1.5 to 3.7 acres Hall, substations, batteries and cooler yard, up to a reference site with roads

Land varies by site: the transformer, generation, access road, setbacks and any easements come on top of the block figures, and local permitting sets the final layout.

Site types and the sizes that suit them
Site typeWhy it fitsBest match
Energised industrial and brownfield sitesFormer mining, mill and manufacturing sites with an existing interconnection or utility contract. The fastest route to power.All three sizes
Power-ready greenfield and campus landA utility feeder and room to repeat. Blocks add in place; a zone suits a campus with a building shell.Blocks; the Zone from 10 MW
Sites where the grid date is the gatePower can start on site generation or a bridge supply and step down to the utility when it arrives.1.6 MW and 4.8 MW Blocks
Metro and building-based sitesLimited land, or an existing building. A single block needs little room; the zone fits a one-storey hall.1.6 MW Block; 9 MW Zone

The site provides

  • Utility power, with generation sized per site
  • A site transformer at 480 V for the 1.6 MW Block or 415 V for the B300 Block; medium voltage for the 9 MW Zone
  • Network and fibre into the site
  • A prepared pad and site works
  • Chilled water for residual air heat (blocks only)
  • A one-storey hall, for the 9 MW Zone only

Syaala provides

  • A written site fit before any order
  • The modules, matched to your racks and tested in the factory
  • Delivery to site
  • The commissioning and test programme through site acceptance
  • Equipment financing for qualified projects

What it is used for

  • AI training on dense GB300, B300 and Vera Rubin clusters
  • High-throughput inference at 130 kW per rack and up
  • GPU clouds adding capacity one step at a time as demand is contracted
  • Enterprise and sovereign AI that needs capacity under its own control
Questions

Modular data centers, answered.

What is a modular data center?

A data center built from factory-made modules. The power, cooling and IT parts are built and tested in a factory, then set on a prepared site and joined. Syaala modular data centers come in three sizes, 1.6 MW, 4.8 MW and 9 MW, all with direct-to-chip liquid cooling for GPU racks.

How long does a modular data center take to deploy?

Typically 4 to 8 months from order to ready for service, depending on equipment reservation and site readiness. Site gates such as utility energisation, carrier fibre and permits can pace the date, and are stated in the written site fit.

What rack densities are supported?

130 kW to 250 kW per rack on a direct-to-chip warm-water loop. GB300 NVL72 class racks run at about 150 kW, HGX B300 racks at 150 kW, and Vera Rubin class racks up to 250 kW in the 9 MW Zone.

How do the 1.6 MW Block, the 4.8 MW B300 Block and the 9 MW Zone differ?

The 1.6 MW Block is the step that repeats: eight rack positions with their own power, battery and cooling plant, N+1 or 2N. The 4.8 MW B300 Block carries 32 liquid-cooled HGX B300 racks at 150 kW, N+1 in each half. The 9 MW Zone places factory-built power and cooling kits in a hall of three pods with four power paths, any three of which carry the load.

Is a modular data center the same as a containerized data center?

A containerized data center is one kind of modular data center. Each Syaala size is engineered as a complete system for liquid-cooled GPUs, with its IT, power, battery and cooling modules sized together, and the 9 MW Zone puts factory-built kits inside a one-storey hall.

How does capacity grow?

Blocks and zones add in place on the same site, from one block to a 50 MW campus. The 9 MW Zone opens in thirds, from 3 MW.

What redundancy is standard?

N+1 on power and cooling as standard, with a 2N option with separate A and B paths on the 1.6 MW Block. The 4.8 MW B300 Block is N+1 in each half, and the 9 MW Zone has four power paths, any three of which carry it.

Which GPU platforms and servers does it support?

NVIDIA GB300 NVL72, HGX B300 and Vera Rubin NVL72 class platforms, including the rack-scale systems Supermicro, Dell, Lenovo and HPE build on them. Each project is checked against the rack maker's facility specification for power, liquid connections, coolant flow and supply temperature. The names describe compatibility; no endorsement or certification is implied.

How are the modules tested?

Twice. The test programme puts every module through a factory acceptance test on load banks before it ships, taking each power module, pump and cooler out in turn to show the load still holds, and third-party (NRTL) evaluation is arranged per project. On site it continues with staged commissioning, an integrated systems test with site generation and fibre, and site acceptance with the racks installed.

Can the modules be serviced while they run?

Yes. Power modules, pumps and coolers are N+1, so any one can be taken out while the block carries its load (an N+1 block runs at N during maintenance; the 2N option keeps a full second path). On the 1.6 MW Block a side extension opens on site to give a clear rack aisle and a separate service aisle for power and cooling connections.

What does the site need to provide?

Utility power with generation sized per site, a site transformer at 480 V for the 1.6 MW Block or 415 V for the B300 Block (medium voltage for the 9 MW Zone), network and fibre, a prepared pad and site works, and chilled water for residual air heat on the blocks. The 9 MW Zone also needs a one-storey hall built by the site contractor.

Is financing available?

Equipment financing is available for qualified projects. Ask for it with a specification request and it is scoped with the site fit.

Have a site in mind

Tell us the power, the racks and the date.

Share what you know. Every field is optional, and the reply is a written site fit for the size that suits it.

Start a project intake
Prefer not to own it

Colocation on the same platform.

Rack-ready GPU capacity on power-ready sites in Texas, the Southeast and the Mid-Atlantic, 20 to 250 kW per rack.

Explore colocation