SeaPRwire – The number that travels with this story is 48 hours. It is worth pinning down what it covers before the rest of the plan gets evaluated. Runware’s 48 hours is the time from a pod arriving at a customer site to that pod running, covering installation and set-up. It is not the build time. Building a pod used to take three weeks and is meant to fall to one day as the line ramps.
The factory is the actual claim. Runware wants to assemble Sonic Inference Pods, the modular data centers it started shipping ready to install in August, on a production line the way cars are assembled. The target is 30 megawatts of capacity per day at its Romania plant by 2028, which works out to just under a gigawatt a month. By comparison, the United States was expected to add about 10 gigawatts of data center power this year. A year of Runware output at that rate would come close to the entire national addition.
That comparison is the company’s framing, not an independent forecast, and it describes a capacity that does not exist yet. What exists is two factories: one that has been building pods for a few years, and a second that broke ground this year, is due to finish within months, and is scheduled to come online early next year. Production rises from there.
Prefabrication is not Runware’s invention. Crusoe and Schneider Electric both sell prefabricated pods, and Nvidia spent this year pushing its Vera Rubin AI factories and DSX POD as ways for companies to build capacity faster. Runware’s distinction is narrower and easier to check: it says it has built a factory, rather than designing a reference design that someone else operates.
The commercial terms are where modularity turns into a product decision. Customers can supply a site or have Runware find one. They can rent compute from Runware, or bring their own GPUs and pay a monthly fee for the pod and its operations, with power and internet billed separately. Pods follow a standard design and can be configured to customer requirements and to current GPU generations.
The GPU part of that is the one with a genuine mechanical advantage. Chips take a long time to arrive after ordering, and a pod that ships faster than the chip does leaves a window in which the infrastructure can be adapted to the hardware that is actually coming. Radulescu puts it plainly: pod delivery times are shorter than current GPU lead times, and the in-house hardware team can design, build and test components for new GPU generations. Whether that window is a month or a quarter decides how much of it is useful.
The dependencies inside the pod are less modular than the pod itself. Runware designs and builds its cooling and other systems from standard components, but pumps and fans come from Germany and networking still comes from Nvidia. A production line that runs at 30 megawatts a day consumes those parts at a rate no single plant controls, and the schedule it promises is only as firm as its slowest supplier.
Cooling is the design choice with the clearest operational consequence. The pods use a closed-loop liquid system with air-cooled radiators, so they need power and an internet connection and not a water connection. That removes the permitting fight that has slowed water-cooled projects, and it also caps how much heat a given footprint can reject, which is why the target sites look nothing like hyperscale campuses. Runware says it is working with power providers on a site pipeline covering 397 megawatts, with potential to expand to 1.2 gigawatts, on sites that typically already have power.
The silence claim sits in the same category. Pods run completely silent up to 86 degrees Fahrenheit, according to the company, and it says it uses renewable energy wherever possible, with an aim of making renewables the main source. Those are property claims for a product sold to neighbours as much as to customers, and they are the kind of claim that gets tested by the first site that sits near housing.
Commitments are the number to watch. Runware says it has received $1 billion in committed interest for Sonic pods since it began offering them to outside customers a month ago, representing more than 20 megawatts now being worked toward. Committed interest and signed capacity are different instruments, and the gap between the two is usually where prefab hardware businesses discover their real demand.
Money is the constraint on the plan as stated. Runware has raised more than $60 million, including a $50 million Series A last December, and says it plans a Series B. It is finishing and launching the factory on the Series A. A plant that needs people, inventory, supply-chain logistics and working capital is a different financing problem from a company that designs pods and contracts out assembly, and the company is describing the change before it has funded it.
The comparison Radulescu gets asked about is Arrival, the electric vehicle startup once valued above $13 billion that could not mass-produce cost-effectively against conventional plants. His answer is that the cost structure differs: factory labour is cheap relative to the value of deployed AI compute, and buying widely available parts keeps costs down. He also points to a plant manager hired from the automotive industry. That is the right kind of hire for the problem, and it is not evidence that the problem is solved.
The comparison that decides the argument is not against other pod vendors but against building conventionally. A hyperscale build takes years and is financed as infrastructure. A pod site takes months and is financed closer to equipment. Runware is betting that the second structure wins on speed even at a higher cost per megawatt, and that bet only pays if the pods arrive configured and the site is ready to receive them. Neither of those is a manufacturing problem, and both are conditions the customer controls rather than the factory.
What would count as evidence is concrete: the second plant opening on schedule early next year, a sustained daily output figure that is audited rather than projected, and pods delivered faster than the GPUs inside them. Until then, the 48-hour number describes a well-defined installation window, the 30-megawatt figure describes a 2028 target, and the case for factory-built data centers rests on whether a line that has never run at that rate can be made to.