What the Runware Sonic Pod is
Runware is an AI inference platform. On August 4, 2026 it introduced Sonic Pods, also called Sonic Inference Pods: 20 ft shipping containers that Runware describes as complete AI data centers with 1 MW of IT compute and about 1,200 GPUs.
Runware’s published details include:
- Compute: nodes of 2 to 8 GPUs, with RTX PRO 6000 as the “workhorse” and B200 and B300 also supported.
- Cooling: direct liquid cooling with a water block on each CPU and GPU, a sealed loop of about 1.5 m³ and air-cooled radiators, with a chiller roughly the size of the container placed on top.
- Sites: wind farms, solar parks and hydro, buying power at generation price, with “no grid interconnect wait.”
- Network: every pod is one node in a single distributed inference network; if a pod fails, traffic reroutes to other pods instead of relying on local redundancy.
- Build and online time: a 3-week average factory build today, with a stated goal of bringing pods online within 48 hours of arrival.
Runware reports customers including Higgsfield AI and Wix, a Romanian factory targeting 30 MW of pods per day by 2028, and more than $1 billion of committed interest representing over 20 MW. These are company statements reported by TechCrunch; we have not verified them.
Sources: Runware Sonic Pods, Runware blog, August 4, 2026, TechCrunch, August 4, 2026, TechCrunch, October 30, 2026. Reviewed November 10, 2026.
The real difference: an inference cloud or your own GPUs
Runware is clear that the pods exist to run its own service: “We do inference. We don’t sell data center space,” and “a pod site runs our workloads only.” Its products are sold as compute, not as containers:
- Serverless: your models or containers on Runware GPUs, billed by the second.
- Compute: committed GPUs priced per GPU-hour, bare metal from 72 GPUs.
- Model serving: priced per asset, output or token.
TechCrunch also reports that customers can supply their own GPUs and pay a monthly fee for pods and operations, but that option does not appear on Runware’s own pages.
QIT works the other way round. We deliver the powered, cooled and connected container on your site; the GPUs, the software stack and the customers stay yours. You keep the margin that a per-GPU-hour or per-token price includes.
Side-by-side comparison
| Dimension | QIT | Runware Sonic Pod |
|---|---|---|
| What you buy | GPU container infrastructure on your site, scoped per project. | Inference capacity: serverless, per GPU-hour or per token. |
| Who owns the GPUs | You. | Runware; customer-supplied GPUs reported by TechCrunch only. |
| Who operates | Agreed per project; you can run your own stack. | Runware schedules, scales and operates the fleet. |
| Form factor | 20 ft shipping container (5.8 × 6.8 × 2.44 m). | 20 ft container with chiller mounted on top. |
| Capacity | 1 MW IT and up to 1,200+ GPUs per container; phased into 5–10 MW sites. | 1 MW IT, about 1,200 GPUs; 2 MW pods mentioned. |
| GPU platforms | Any architecture, including HGX B300, GB300 NVL72 and RTX PRO 6000. | RTX PRO 6000 primary; B200 and B300 supported. |
| Workloads | AI training and inference. | Inference first (“Our servers are built for inference and nothing else”). |
| Cooling | Direct liquid cooling on every GPU and CPU; sealed closed loop with dry coolers, no water consumption. | Direct liquid on each CPU and GPU; sealed loop with air-cooled radiators and chiller. |
| Redundancy model | Site resilience designed into the project. | Network-level: traffic reroutes to other pods. |
| Sites | Your energized site; Texas focus, existing transformers such as bitcoin mines. | Runware-selected renewable sites, or a customer site; Europe live, US underway. |
| On site | Installed in 15 hours at a prepared site. | Goal of online within 48 hours of arrival. |
Runware figures are from its published pages and TechCrunch reporting cited above. QIT and Runware publish a near-identical container envelope; check the configuration actually offered for your site. Reviewed November 10, 2026.
Your site, or theirs
Runware places pods where it can buy cheap power, mostly renewable generation sites, and reports 160 locations available to it. That suits a network that can move inference traffic between pods.
QIT starts from a site you control. In QIT Pod 001 the container reuses an existing 2,500 kVA transformer at an operating bitcoin mine, so no new utility service is needed. If you own an energized site, QIT turns its spare capacity into compute that you own, instead of leasing the site to an inference operator.
Network redundancy versus site redundancy
Runware says a pod site can trade redundant power for very cheap power, because a failed pod’s traffic moves elsewhere in its network. That works for stateless inference requests across a large fleet.
A single owner running training jobs, dedicated customers or a small number of pods cannot reroute in the same way. QIT designs power, cooling, fire protection and network for the uptime your workload needs, at the site level.
When each option fits
Runware may fit when you want inference capacity without owning hardware: serverless endpoints, GPU-hours on a term or per-token model serving, and you do not need the compute on a particular site.
QIT may fit when you have energized power, want to own the GPUs and the margin, need training as well as inference, or need the compute on your own property for control, data residency or customer contracts.
Questions to ask any 1 MW container vendor
- Will I own the GPUs, rent them, or pay per token?
- Can the container run my own software stack and customers, or only the vendor’s workloads?
- Is 1 MW the IT load, and what does the chiller or dry cooler add at my site’s design ambient temperature?
- What happens to my workload if the container loses power: site redundancy or rerouting to someone else’s network?
- Which GPU platforms are actually available for my delivery date?
- What event starts the “online in hours” clock, and what must the site already provide?
Related comparisons and next steps
See how QIT compares with Crusoe Spark, Armada and Giga Energy’s GigaPod, or read the full side-by-side comparison.
QIT is independent of Runware. This page does not claim equivalent or superior measured performance. Discuss your site with QIT.
Prepared by QIT · Updated November 2, 2026
Explore all field notes