AI infrastructure startup Runware wants to manufacture data centers the way automakers build cars, with finished units rolling off a production line. Chief Executive Flaviu Radulescu has reportedly said the company plans to produce 30 megawatts of data center capacity per day by 2028, and to bring every pod online within 48 hours at any location already equipped with power supply and connectivity. Those achievements, however, have yet to be proven on a larger scale.

The appeal of the approach lies in the speed of production and commissioning. Runware intends to build and test its pods inside a factory before delivering them for installation, a method designed to reduce the amount of work performed on site. Whether the company can achieve these objectives on time and at lower cost remains another problem to be tackled.

Pods coming off a line, like cars

Runware's modular AI data center, the Sonic Pod, is 20 feet long and houses about 1,200 graphics processing units. Each unit delivers 1 megawatt of computing power, and the closed-loop liquid cooling system requires no permanent water connection. The company already assembles these modular data centers in Romania and expects a second factory to open by early 2027.

Building a single Sonic Pod currently takes about three weeks. According to Radulescu, that could drop to one day with an assembly line. The 48-hour target refers to something else entirely: it excludes fabrication and grid access, and instead covers the installation and setup of a finished pod at an already prepared site.

Runware expects a surge in customer demand to propel its growth strategy. The company claims to have secured $1 billion in committed interest, representing more than 20 megawatts of capacity, though this interest has yet to be converted into completed sales. In an October 30 announcement, the company said mass production is projected to begin in late 2027.

Nvidia's standard, Crusoe's factory, Runware's inference bet

Runware is not the only company seeking to accelerate data-center construction. On October 21, Nvidia unveiled the DSX POD, a reference design for ready-made systems built on Nvidia's Vera Rubin computing platform. The units are assembled and tested before being shipped to buyers, and Nvidia claims installation at prepared sites takes two weeks instead of the six months required with conventional methods.

Crusoe is one of those partners. Its Spark platform connects a 352,000-square-foot Colorado factory with Crusoe Cloud, and the company has operated modular facilities since 2020. It says it can deliver productive AI capacity in as little as three months.

The timelines, however, measure different stages: Runware's 48 hours covers setup, Nvidia's 14 days covers deployment, and Crusoe's three months covers working capacity.

Runware's business model differs as well. Rather than selling data-center space, it sells inference services — the work of running trained AI models to generate outputs. Customers can supply their own GPUs, although Runware still relies on Nvidia for networking.

A $75.8 billion market with hard limits

The pressure to speed up modular data-center manufacturing follows a rise in demand for the facilities. According to Grand View Research, the market was worth around $29 billion in 2024 and could reach $75.8 billion by 2030, an annual growth rate of 17.4%.

According to JLL, total global data-center capacity could reach 200 gigawatts by 2030, requiring almost $3 trillion in investment. The same firm predicts that inference will become more popular than training starting from 2027. In Japan, JERA, Dell, and RHAELM are planning a repeatable design for a $15 billion, 400-megawatt project.

Faster construction will not solve the industry's power problems, however. JLL says grid-connection waits in major markets exceed four years, while the IEA expects data centers to consume 950 terawatt-hours of electricity by 2030.

Public opposition adds another challenge. Pew found that 54% of Americans consider data centers bad for the environment, up from 39% in January. Texas paused grid approvals for projects under review in August and halted environmental permits in September.

Manufacturing carries its own set of risks. As TechCrunch has reported, electric vehicle startup Arrival struggled to make mass production profitable. Runware now faces a similar test: proving whether its assembly line can perform at the speed and scale it claims.

This article is based on reporting by Cryptopolitan.