Armada.

A factory the size of a Walmart, full of GPUs

In a warehouse outside Phoenix this summer, a 400,000-square-foot floor is being fitted out with cranes, paint booths, and conveyor lines. The product rolling off them is not a sedan or a refrigerator. It is a steel box, roughly the dimensions of a shipping container, packed with NVIDIA accelerators, a liquid-cooling loop, a Starlink terminal, and enough on-board software to behave like a regional cloud region by itself.

The company building these boxes is called Armada. As of May 2026, it is worth $2 billion, employs about 250 people across San Francisco and Seattle, and is shipping its products to oil rigs, U.S. Navy vessels, Saudi gas fields and at least one mining operation that politely declined to be named.

The cloud has a geography problem

For two decades, the tech industry insisted that compute had no location - that it lived "in the cloud," a place vaguely north of your head. The reality is more boring: compute lives in roughly thirty hyperscale regions, mostly clustered around fiber backbones in Northern Virginia, Dublin, and Singapore. If your data is also there, fine. If your data is on a drilling platform in the Permian Basin, you have a problem.

The problem is not theoretical. A modern offshore rig generates terabytes of sensor data per day. A defense forward operating base needs computer-vision inference in milliseconds, not minutes. A mine in the Atacama would happily pay for predictive-maintenance models if only the satellite link could squeeze them through. None of this works if every byte has to make a round-trip to Virginia.

That insight is not unique to Armada. Edge computing has been declared the next big thing approximately every eighteen months since 2014. What is unusual is Armada's answer: don't fight physics, ship the data center.

A CEO with a strange resume

Dan Wright did not arrive here by accident. Before Armada, he was CEO of DataRobot during its run from a few hundred customers to several thousand. Before that, he was COO of AppDynamics during the last-second $3.7 billion Cisco acquisition in 2017 - a deal that closed the night before the IPO was supposed to price. He has done the enterprise-software thing twice, and was apparently uninterested in doing it a third time.

Wright co-founded Armada in 2022 with Pradeep Nair and Jon Runyan. The pitch to their first investors was less "we will build better software" and more "we will build hardware, software, and a satellite plan, and we will charge one price." It is a strange bet for three software executives to make.

The bet had immediate validation. Founders Fund and Lux Capital led a $55 million stealth round in late 2023. Microsoft's M12 fund added $40 million the next summer and put every Armada SKU on the Azure Marketplace. By mid-2025, a strategic $131 million round closed alongside the launch of Leviathan, the company's megawatt-scale unit. Then BlackRock came in for the Series B.

A small navy of boxes

Armada's product names are unsubtle. The smallest unit is the Beacon, a suitcase-sized inference box. Above it sits the Cruiser, a 20-foot containerized cluster. Above the Cruiser, the Titan - a 40-foot version with substantially more GPUs and a beefier cooling loop. Sitting above all of them is the Leviathan, a megawatt-scale modular data center designed to be co-located with stranded natural gas, a solar field, or, eventually, small modular nuclear.

What sits inside any of these boxes is essentially the same recipe. NVIDIA accelerators for compute. Starlink (or, where regulators prefer, terrestrial microwave) for connectivity. A liquid-cooling system tuned to operate at 50 degrees Celsius without complaint. And Armada's own software platform - confusingly, also given a maritime name - called Atlas, which manages workloads, security, and fleet operations across however many boxes a customer has bought.

The numbers behind the bet

The company's traction is the part that has investors using the word "oversubscribed" out loud. Customer bookings grew 540% from fiscal 2025 to fiscal 2026. In the first quarter of fiscal 2027, bookings were up roughly 2,000% year-over-year. The customer list - the parts of it that are public - includes the U.S. Navy, Saudi Aramco, Halliburton, and at least one major hyperscaler buying Armada units as a sovereign-AI fallback.

Partnerships are arranged in the same logic as the product: cover every layer of the stack with a name customers already trust. Microsoft handles the cloud connection. SpaceX handles the satellites. NVIDIA handles the silicon. Johnson Controls now handles the factory. Nscale handles the sovereign-AI play. It is an unusual coalition - a constellation of partners that, on paper, should not all agree on lunch.

Why this company exists

Armada uses the phrase "the U.S. AI stack" a lot, especially after the Series B. The framing is partly marketing and partly geopolitics. If sovereign AI is the question - the idea that nations, militaries, and companies want compute that is provably under their own control - then a portable, on-premise data center is, conveniently, an answer.

The wider mission is less patriotic. Armada wants the next trillion dollars of compute to happen outside hyperscale regions. That requires three things to be true at once: that AI inference moves close to data, that connectivity stops being a constraint, and that someone is willing to manufacture rugged infrastructure at scale. Armada has decided to be the third thing.

It is a strange thing to root for - a company whose worldview only makes sense if the cloud, as we currently understand it, is incomplete. But the world the company is betting on is recognizable. Most data is generated outside hyperscale regions. Most AI workloads benefit from running near the data. Most enterprises would rather not pay egress fees forever. Armada is wrong only if those three things stop being true at once. They probably will not.

Back to the factory floor

Return, then, to Arizona. Galleon Forge One is not yet at full capacity. The Leviathan units leaving the line this summer are largely pre-sold. The Cruisers and Titans behind them are mostly spoken for. Customers signing purchase orders today are quoted multi-month lead times - the polite way of saying the demand is already ahead of the manufacturing.

This is not the standard arc of a software startup. Software startups scale by spinning up another region. Armada scales by hiring welders. There is a romance to that, but also a discipline: you cannot fake the unit economics of a steel box. The box either ships, or it doesn't. The GPUs either boot, or they don't. There is no clever software workaround for a missing power supply.

Which is, in the end, the most interesting thing about Armada. It is an AI infrastructure company that has chosen the least scalable, most physical, most logistically painful version of the AI infrastructure problem. And it is winning anyway, because the customers who need this product cannot wait for the cloud to grow another region near them. The cloud, it turns out, has weight. Armada is the company that put it on a truck.

Where to go from here