Runware, an AI inference company, says it wants to manufacture AI data centers the way carmakers build cars. In an October 30, 2026 interview with TechCrunch, CEO Flaviu Radulescu said the company plans to scale its Romanian factory to 30 megawatts of capacity per day by 2028, nearly a gigawatt a month, with pods brought online within 48 hours of arrival. Everything here is a plan or a company claim; the reporter herself writes that "execution is the test."
This post lays out what Runware has actually documented, what is new, and how to read the numbers.
TL;DR: claims, sources and status
| Claim | Where it comes from | Status |
|---|---|---|
| 30 MW per day by 2028 (about 1 GW a month) | CEO to TechCrunch, Oct 30 | A plan, not production |
| Two factories; second online early 2027 | Same interview | Company statement |
| Pods online within 48 hours | Same interview | Covers install and setup, not build |
| $1B in committed interest, over 20 MW, in a month | Same interview | Company claim; "interest" is not signed orders |
| 397 MW site pipeline, up to 1.2 GW | Same interview | Company claim |
| 1 MW and about 1,200 GPUs per 20-foot container | Runware launch post, Aug 4 | Company specification |
| 30 to 80 percent lower cost per GPU-hour | Runware launch post | Company claim, no public benchmark |
| $150M per GW facilities vs $15B conventional | Runware launch post | Company claim, GPUs excluded |
| Raised more than $60M, $50M Series A in Dec 2025 | TechCrunch, Runware | Reported |
What a Sonic Pod is
Runware's launch post describes the pod as 1 MW of IT compute in a 20-foot shipping container, with a chiller about the size of the container sitting on top. Inside are around 1,200 GPUs across custom caseless servers, custom racks, Runware's own PCIe switching, high-frequency CPUs and local NVMe so models stay warm. Cooling is a single closed loop that recirculates about 1.5 cubic meters of liquid and, the company says, holds GPU temperatures within 2 degrees C of target in ambient temperatures up to 50 degrees C.
The workhorse node uses RTX PRO 6000 GPUs, with support for B200 and B300 and an expectation of Vera Rubin. The pod needs ground, power and a network connection, and runs as one node in a distributed inference network that reroutes traffic if a pod goes down. A pod needs a crane to place and the chiller to be lifted on and connected, which the post admits is "more complicated than it sounds."
What is new this week
The August post was a product launch with a target of up to 10,000 nodes online in the second half of 2026 and more than 1 GW of inference compute in 2027. The October interview adds the factory story:
- A production line rather than one-off builds, with a plant manager hired from the automotive industry.
- A 30 MW per day target by 2028. TechCrunch calculates that this would equal roughly the 10 GW the whole US was expected to add this year in about a year of output.
- Build time per pod falling from three weeks to about a day.
- Two options for customers: rent compute from Runware, or bring their own GPUs and pay a monthly fee for the pod and operations, with power and internet billed separately.
The arithmetic is internally consistent: 30 MW a day is about 11 GW a year. But that is a design target for a year not yet reached, and the company has not said how many pods have shipped to date.
Why this matters for AI infrastructure
Data center capacity is the constraint behind much of this year's AI news. Transformers now carry lead times of around 160 weeks, optical parts are sold out into 2029, and communities are pushing back: San Francisco adopted a moratorium, New York faces a proposed statewide pause, and protests in India hit Amazon's Thane project over water and power. Microsoft even pulled out of an Essex AI datacentre project on grid power grounds.
Runware's pitch is aimed at exactly those frictions. Small sites that already have power, no water connection, no evaporative cooling and a standardized container could avoid some permitting fights. Radulescu told TechCrunch the pods run silently up to 86 degrees F and that the site pipeline is built on locations that typically already have power. Whether the communities concerned would accept dozens of containers next to a substation is untested, and a closed loop still rejects heat, which air-cooled radiators dump into the local environment.
The competition
TechCrunch notes the idea is not new: Crusoe and Schneider Electric sell prefabricated pods, and Nvidia announced Vera Rubin AI factories and the DSX POD reference designs earlier this year. Runware's distinction is that it says it has built a factory rather than only a design. Scale is the open question. For comparison, our post on China's 24 GW of compute versus the US 56 GW shows how large the national numbers are, which makes 1 GW a month an extraordinary figure for a company with about $60 million of funding.
The risks, in the reporter's words and ours
TechCrunch compares the plan with electric vehicle startup Arrival, once valued above $13 billion, which could not mass-produce vehicles cost-effectively. A factory needs people, inventory, logistics and capital. Runware says it designs most systems itself and buys widely available parts, but some pumps and fans come from Germany and it depends on Nvidia for networking. GPU supply is a bigger dependency: the pods are only useful with GPUs, and the company says its delivery times are shorter than current GPU lead times.
Other things to weigh:
- "Committed interest" is not revenue. The $1B figure is interest in more than 20 MW. Runware's own facilities figure of about $150M per GW works out to roughly $0.15M per MW, so 20 MW of pod shells would be a few million dollars. A $1B headline must therefore include GPUs and operations, and the company did not break it down.
- Cost claims lack a public benchmark. The launch post promises to benchmark your model on the hardware that would serve it, and says it prefers results on your model to published numbers.
- Facilities are only part of the bill. The $150M against $15B comparison excludes GPUs on both sides, and cites OpenAI's CFO for the conventional figure. GPUs dominate total cost.
- Distributed reliability. Running on many small sites trades concentrated risk for coordination risk. Runware says the blast radius is one pod and traffic reroutes.
- Environmental claims. "No water" refers to normal operation of the closed loop; renewable power is an aim, not a guarantee.
How Runware got here
The launch post is unusually candid about the company's path, and it explains why the factory idea is not a random pivot. Runware began in 2022 around a fast image-generation product, PicFinder, then discovered that rented GPUs of the same model performed very differently depending on surrounding hardware, CPU clock speed in particular. In summer 2023 the founders assembled their own liquid-cooled servers, a job that took three days for the first machine, and measured roughly 50 percent more performance from the same GPUs after tuning components and software.
By 2024 the company was colocating servers and found that few facilities would accept liquid-cooled racks at all, so it designed the server and the building together, starting at the end of 2023. The first prototype pod entered production in early 2025 with smaller GPUs; it was later upgraded for video and language models. The company says capacity shortages, cancelled or deferred projects and local moratoria in early 2026 convinced it to ramp production and sell pods to others, and it claims 160 locations available today.
That history matters for how you weigh the claims. The pod is not vaporware: Runware says it has run its own inference on it for months. What is unproven is the step from a handful of pods built by a hardware team to a continuous production line. Hardware iteration cycles are slow, as the post says itself: mistakes take months rather than minutes, and its first cooling supplier never delivered. A factory multiplies both the output and the cost of an error, which is exactly the scenario the TechCrunch comparison with Arrival is meant to evoke.
It is also worth noting what Runware does not claim. It says it sells inference, not data center space, and that it is not trying to out-build hyperscalers. The new element in October is that it is now also offering pods to customers who bring their own GPUs, which turns a cost-saving internal project into a product with its own customers, support obligations and spare-parts logistics.
What this means for builders
If you buy inference rather than build it, the practical point is to ask providers where the capacity physically sits, how much is owned versus rented, and how quickly it can grow. Prefabricated capacity could shorten the time between ordering GPUs and serving traffic, which is where the industry is currently stuck. If you serve media models, Runware's own business, benchmark your workload before assuming the headline discount applies, since the company itself says the saving depends on model size and parallelism.
For readers tracking infrastructure, the thing to watch is not the 2028 target but the next two checkpoints Runware has set itself: the second factory coming online early next year and the first deliveries from the $1B pipeline. Those will show whether a car-style line can work for racks of GPUs.
Related reading
- Transformer lead times of 160 weeks and AI data centers
- Lumentum optical parts sold out into 2029
- San Francisco's data center moratorium
- New York data center moratorium explained
- Microsoft pulls out of an Essex AI datacentre
- China 24 GW compute vs US 56 GW
Details are as reported on October 30, 2026 and reflect Runware's statements and plans.