NVIDIA DSX AI Factory: SWI Compute's UK Plan and the 40% Power Claim

What SWI Compute announced

SWI Compute is a new company inside SWI Group, the investment group listed on Euronext Amsterdam that was formed by the merger of Icona and Stoneweg. According to its launch release of October 26, 2026, it became a Preferred NVIDIA Cloud Partner in August, has ordered about $1 billion of NVIDIA systems and has signed its first customers, the first being a US based AI inference platform.

The release puts the first large AI factory in the UK. It starts at 30 MW with close to 7,000 Rubin GPUs and 3,500 Vera CPUs in Vera Rubin NVL72 racks, and is expected online by the end of 2027. Those 30 MW are the first phase of a UK campus planned to grow to 530 MW, and a first step towards 2.3 GW of NVIDIA AI factory capacity by 2029. The parent group says it holds 3.6 GW of power capacity across Europe and the US.

The current deployment is different hardware. It runs NVIDIA Blackwell Ultra systems, with Vera CPU racks to follow, and Delta Computers is the systems integrator for more than 8,000 GPUs. SWI Compute's CEO, Halil Mirakhmedov, called energy availability the primary bottleneck facing European AI infrastructure, as reported by Unite.AI.

What NVIDIA DSX is

The release describes the NVIDIA DSX AI Factory Platform as a full stack that codesigns compute, networking, power and cooling as one integrated system. Every new SWI Compute site from 2027 will follow the NVIDIA Vera Rubin DSX AI Factory reference design. The release also claims that on Vera Rubin NVL72, DSX can enable up to 40% more GPU capacity within the same power budget.

Five days earlier, at GTC Berlin on October 21, NVIDIA introduced DSX POD, a reference design for complete Vera Rubin AI factories that partners build inside modular containers and validate before delivery. NVIDIA says these factory built units cut deployment at a prepared site from six months to as little as 14 days, with rack level power and liquid cooling inside the integration scope. Seventeen partners, among them Dell Technologies, HPE, Lenovo, Schneider Electric and Vertiv, are due to deliver units in early 2027.

Why it matters for an AI data center operator

A fixed grid connection is the hard limit on most European sites, and SWI Compute's own CEO says so. If a reference design lets more GPUs run inside the same megawatts, the design changes how much revenue a site can earn. It also moves the risk. A site that plans for more GPUs per megawatt runs with less spare power and less spare cooling, so a wrong estimate shows up quickly in rack power alarms and cooling loop temperatures.

The 40% figure is an upper bound from the vendor side. What an operator can bank is what the meters show after commissioning: real rack power against rated power, cooling loop temperatures and flow under full training load, and the share of installed GPUs that are actually busy. The release itself says an AI factory earns on how many tokens it produces per megawatt, which is a measured number.

Prebuilt modules add a second operations question. A unit that arrives tested from the factory still has to be accepted on site, connected to the power and cooling plant, and entered into the asset and monitoring systems before its first job. If acceptance and inventory lag behind the 14 day installation, the gain in deployment time is lost in paperwork.

What the operator sees in Sensaka

Sensaka for AI Data Centers keeps power, cooling, accelerators and jobs in one inventory, which is the record needed to test a power budget claim. In the AI Cockpit the operator asks plain questions. Which racks or zones are near their rated power? Is the liquid-cooling loop healthy: temperature difference, flow and leaks? Which GPUs are idle and could be reclaimed? Each question is matched locally to a preset operational question and gets its own board, and one click renders the board as a 30-second data video for the capacity review.

GPU pool operations shows pools, idle GPUs and quotas by project and tenant, so the operator sees how much of the installed capacity is doing work. Liquid cooling monitoring reads CDU status, supply and return temperatures, pressure, flow and leak events, and places them beside rack power, PUE and WUE. This is monitoring only; control of the cooling plant stays with the site's own systems and procedures.

For new modules, the CMDB relationships from rack to server, accelerator, job and tenant give each delivered unit a place in the inventory on the day it is connected. Jobs from Slurm and Kubernetes appear with their queue reasons and failure reasons, and acceptance or fault tickets sync with ServiceNow or Jira where those run. Sensaka works with NVIDIA, AMD and other accelerators.

Takeaway

SWI Compute is betting that NVIDIA DSX turns a fixed grid connection into more GPUs. Every new AI factory built on that design will need a measured answer in its first months to how many GPUs are running inside its megawatts and how much cooling margin is left. That answer comes faster when rack power, cooling readings and GPU use already sit in one record.

Frequently asked questions