Armada’s 500 MW AI Data Center Goal Is Now a Factory Throughput Question

Leviathan modular AI data center unit on the Galleon Forge Two factory floor at Johnson Controls in Dublin

TL;DR · 30-second read

The Short Version

A factory in Dublin, Ireland, has started building artificial intelligence data centers the way other factories build trucks. Whole computer rooms, with their cooling systems inside, are assembled and tested indoors, then shipped out.

The first one is going to Norway. The companies behind it, the startup Armada and the building-equipment maker Johnson Controls, want to turn out roughly 25 times more next year than their named customers have announced so far.

Why it matters: if computing power is built like a product instead of a building, how fast it arrives depends on how quickly these factories can work.

Armada, a San Francisco-based maker of modular AI data centers, and Johnson Controls (NYSE: JCI) announced on November 11, 2026, that Galleon Forge Two is in continuous production. The operation sits inside a 90,000-square-foot Johnson Controls plant in Dublin, Ireland, and is supported by about 700 Johnson Controls engineers, technicians and skilled workers. The first unit built there is a Leviathan, a factory-built modular data center. It is headed to Fossefall in Norway, which has ordered five Leviathans totaling more than 9 megawatts (MW) for delivery by the first quarter of 2027.

The companies say they are targeting at least 500 MW of capacity in 2027. Dublin is the second Galleon Forge. The first, at Johnson Controls’ site in Gilbert, Arizona, is expected to begin continuous production in early 2027, and the companies plan further expansion at a Johnson Controls site in Edmonton, Canada. Australian developer WinDC also plans to deploy 11 MW of Leviathan capacity at renewable-energy sites.

Executive Summary

The announcement moves Armada’s European production from plan to practice. Leviathan units combine Armada’s compute hardware and management software with Johnson Controls’ thermal equipment: coolant distribution units, which circulate liquid to carry heat away from chips, plus chillers and building controls. The units are now being engineered, integrated and tested on a factory floor in Dublin rather than assembled on a construction site. The companies frame this as “sovereign” AI infrastructure, meaning computing that customers own and run inside their own borders without relying on an outside cloud provider.

The significant number is the 500 MW target for 2027. Under a factory model, capacity arrives as fast as plants finish units. Today only one plant is in continuous production, and the larger site in Arizona is not expected to reach that stage until early 2027. Named customer demand totals about 20 MW, roughly 4% of the goal. So the questions that will decide whether the model works are how many units the factories can turn out, and who is lined up to take them.

Why 500 MW Comes Down to Units Per Month

In a conventional data center project, the schedule is set on site. Crews pour foundations, install electrical gear and cooling plants, and then commission everything in sequence. Leviathan moves most of that work indoors. Johnson Controls’ Austin Domenici described the goal as completing “more of the complex integration work in the factory” so that compute comes online “faster and more predictably.” The first Dublin unit was engineered, integrated and tested entirely at Galleon Forge Two. The consequence is that delivered capacity now tracks factory throughput, the number of finished units leaving the plant each month, rather than the pace of a construction site.

Simple arithmetic shows why that number matters. Armada’s company description refers to 10 MW Leviathans, so 500 MW would be about 50 units at that size. Fossefall’s order, however, is five units totaling more than 9 MW, an average of a little under 2 MW each. At that size, 500 MW would take something closer to 250 units. The companies have not disclosed units per month for either plant. Dublin is the only line in continuous production today. Gilbert, at up to 400,000 square feet the larger footprint, is expected to start in early 2027, and Edmonton has no date. Reaching 500 MW within calendar 2027 therefore depends on both plants ramping quickly and on the unit mix leaning toward larger modules.

The people affected are concrete. Buyers who want 2027 capacity are effectively competing for production slots. Developers have to have land, permits and a power connection ready when units arrive, because a finished module waiting on an unenergized site produces nothing. Johnson Controls’ thermal supply chain has to keep pace with the assembly line. Factory production can make delivery more predictable, but only up to the plants’ output rate.

A Target Larger Than Disclosed Demand

Two customers are named. Fossefall has placed a firm initial order for more than 9 MW. WinDC plans to deploy 11 MW in Australia, described as a plan rather than an order. Together that is roughly 20 MW against a 500 MW goal. Undisclosed customers may account for the rest, and a capacity target is not the same as a sales forecast. Still, the announcement establishes production capability more firmly than it establishes demand.

The Fossefall order does carry real information. Five units scheduled within roughly a quarter and a half of the first unit’s completion shows the plant expects to deliver repeatedly, not one-off. It is also a modest volume. Readers should treat the 500 MW figure as an ambition the companies have stated, and the Fossefall schedule as the first measurable test of it.

What “Sovereign” Does and Does Not Cover Here

Several parts of the sovereignty claim are supported by specifics. Manufacturing happens in Ireland with a European workforce. Customers own the hardware. The Armada Platform manages power, cooling and compute without dependence on an external cloud. For a European buyer worried about where infrastructure is built and who controls it, those are meaningful points. Fossefall said that finding partners in Europe was particularly important to its vision of sovereignty.

Other layers are not addressed. The release does not say where the processors and other compute components come from. The operating software is Armada’s, from a company headquartered in San Francisco and backed by US investors. Sovereignty in practice is a matter of degree across hardware supply, software control and physical location. This announcement clearly settles the location question. Buyers will need to evaluate the others.

Site Work Shrinks, Power Constraints Do Not

Factory building reduces construction time, but it does not make a power connection appear. Both customers emphasize energy. Fossefall’s CEO, Øyvind L. Vesterdal, pointed to clean energy, political stability and climate as Nordic advantages. WinDC’s Andrew Sjoquist argued for placing compute at wind, solar and battery sites “rather than waiting for a grid that can’t scale fast enough.” Self-contained modules suit that strategy because they can be sited where power already exists. Their delivery speed is only useful if the site is ready.

Johnson Controls’ zero-water cooling claim is relevant to where these units can go. Cooling that does not consume water widens the range of sites a buyer can consider, particularly renewable-energy locations far from water infrastructure. The release says these technologies are designed to reduce energy use and noise but does not quantify either.

Background

Armada, founded in 2022 and headquartered in San Francisco, sells what it calls a Sovereign SI Grid: a distributed network of factory-built modular data centers that customers own and run as a single system. Its product range runs from suitcase-sized Galleon units for running AI models close to where work happens (known as inference) up to Leviathans for fine-tuning and inference at larger scale. Its investors include Founders Fund, Lux Capital, Shield Capital, 8090 Industries and M12. Armada uses “SI” in its branding, linking its products to superintelligence.

Johnson Controls is a building-systems and thermal-management company with a history of more than 140 years. It serves data centers, healthcare, pharmaceuticals and advanced manufacturing, and says it has built modular infrastructure for nearly two decades. Modular, or prefabricated, data centers are not new. What has changed is the power density of AI hardware. Dense AI chips generate enough heat to require liquid cooling and tightly integrated thermal systems, and that integration is the work these factories aim to standardize.

Sources

Source: Armada and Johnson Controls Launch Galleon Forge Two in Dublin, Bringing Sovereign AI Infrastructure Production to Europe, a joint announcement of continuous Leviathan production in Dublin, the first unit for Fossefall, and a 500 MW capacity target for 2027.