Latest in AI
SWI Compute's UK AI Factory: Rubin Capacity, Timing and Access
SWI's new compute business has a power portfolio and hardware orders behind it, but the useful capacity numbers depend on which stage of the plan you mean.
SWI Group announced SWI Compute on October 26, setting out plans for European NVIDIA AI factories. Its first large UK installation targets 30 MW and roughly 7,000 Rubin GPUs by the end of 2027. This is an infrastructure rollout announcement, not a Rubin cloud developers can immediately start using.
The distinction matters if you are comparing places to train or serve models. A provider's power portfolio, ordered hardware and customer-ready cluster describe different stages of delivery. Treating them as interchangeable makes a future supply announcement look like available capacity.
What SWI Compute announced
The issuer's announcement says SWI Compute has ordered approximately $1 billion of NVIDIA systems and contracted its first customers. Its parent reports a 3.6 GW power platform across Europe and the US.
The UK plan uses Vera Rubin NVL72 racks, with around 3,500 Vera CPUs alongside the GPUs. The company targets 2.3 GW of AI factory capacity by 2029. Those are company plans, rather than independently demonstrated operating capacity.
SWI was already named in NVIDIA's October 21 European cloud-partner announcement. The October 26 release supplies the company launch and deployment detail; it is not the first public mention of the business.
SWI Compute capacity and availability
The current product website identifies the first large site as Cambridge and describes dedicated and on-demand access for AI companies, labs and enterprises.
| Item | What the current sources establish |
|---|---|
| Initial Cambridge phase | 30 MW, with roughly 7,000 Rubin GPUs planned |
| Target commissioning | By the end of 2027 |
| Full UK campus | Press release says 530 MW; product page says 500 MW |
| Customer access | Request capacity and receive an offer and timing from the team |
| Public GPU-hour pricing | Not listed on the product page |
The 3.6 GW power platform is not a claim that 3.6 GW of GPU compute is running today. It spans the parent's infrastructure portfolio. Buyers need the capacity and delivery date attached to their particular offer, rather than the largest number in the announcement.
The campus-size discrepancy also deserves a qualification. Both official sources agree on the initial 30 MW phase, but they give different figures for the eventual campus. Neither figure should be presented as a settled delivered total.
Why the architecture matters to a buyer
NVIDIA describes its European cloud expansion around the DSX platform, which brings compute, networking, power and cooling into a coordinated infrastructure design. SWI's site says new locations from 2027 will follow the Vera Rubin DSX reference design.
For a buyer, the useful question is how a cluster performs on the intended workload. A training run needs reliable access to connected GPUs for its duration. An inference service needs enough capacity at busy times, alongside a workable response-time target. The same GPU count does not answer both questions.
A capacity discussion should therefore cover the actual accelerator configuration, network topology, storage throughput, reservation conditions and support responsibilities. Ask which parts of the proposed service are already operating, which are on order and what happens if the delivery schedule moves.
What to watch before committing a workload
The next meaningful evidence is commissioned infrastructure with a concrete service offer: where it runs, when access begins, what it costs and what performance customers can reproduce. A reference design and a hardware order help describe the plan, but they do not replace those details.
Regional AI capacity becomes useful to a developer when a model can run there on acceptable commercial and operational terms. The gap between a secured power connection and that usable service is where this rollout will need to prove itself.
I'm Paras. I work at the intersection of AI, engineering and distribution.
I help technical teams turn complex infrastructure and product updates into clear stories that developers and decision-makers can use.