AI Factory Bet: SWI Compute 3.6 GW and a $1B NVIDIA Order
The newly launched company has 3.6 gigawatts of contracted power, a $1 billion NVIDIA order and a UK campus that will not switch on until the end of 2027. The sequence is the story.
SWI Compute launched out of stealth with a proposition that inverts the usual order of an AI cloud company. It has 3.6 gigawatts of power capacity, six years of land and grid work behind it, a preferred position in NVIDIA’s cloud partner programme, about a billion dollars of NVIDIA systems on order and its first contracted customers. What it does not have is a running AI factory. The first one, a 30-megawatt hall in the UK with close to 7,000 Rubin GPUs and 3,500 Vera CPUs in NVL72 racks, comes online at the end of 2027. That is the AI factory, and it does not exist yet.
That sequence is the point. Most of the companies that built AI clouds over the last three years started with chips and then hunted for space and power, which is why so many of them ended up paying hyperscalers rent or waiting in grid queues. SWI Group did the opposite. It began buying land, power and shell in 2020, and only now, with the scarce input under control, is it buying GPUs and selling compute.
What SWI Compute actually is
SWI Compute is the compute arm of SWI Group, an investment group listed on Euronext Amsterdam under the ticker SWICH and formed from the merger of Icona and Stoneweg. The group allocates more than 80 per cent of its capital to digital infrastructure, with the stated intention of pushing that above 90 per cent, and it holds capacity on both sides of the Atlantic. In Europe the vehicle is AiOnX, a portfolio of about 2.3 gigawatts across Ireland, the UK, Denmark, Spain and Italy, one campus of which is already leased to a hyperscaler.
In the United States it is SWI Digital, formerly Genesis Digital Assets, a company SWI took a controlling stake in this year, bringing roughly 1.2 gigawatts of secured grid connections. SWI Compute sells access to NVIDIA systems built on those sites under an AI factory model, in which compute, networking, power and cooling are designed as one system rather than assembled separately.
The structure matters because it explains the funding. SWI Group deploys its own balance sheet rather than a venture fund’s, which means it can carry land and grid work for years before revenue arrives. It also means the group’s listed status gives the compute business a public valuation reference and a route to capital that most AI cloud startups do not have.
Power first, chips second
The company’s chief executive, Halil Mirakhmedov, states the thesis plainly: energy availability is the primary bottleneck facing European AI infrastructure. That is not a controversial claim in 2026, and the evidence sits in grid connection queues across Europe and in the price of firm capacity. What is unusual is the length of the bet. Six years of site acquisition and grid work before the first GPU arrives is a real-estate timeline, not a technology one, and it means SWI Compute enters the market holding something most neoclouds lack: the right to draw power nobody else can.
The commercial signal that matters most is the order book. SWI Group became a preferred member of NVIDIA’s cloud partner programme at the end of August, and within roughly two months had committed to about a billion dollars of NVIDIA systems while signing its first customer contracts. For a company that had not yet switched on a facility, that is a fast conversion, and it tells a reader two things. NVIDIA allocates supply to partners who can demonstrate power and shell rather than intent. And buyers of AI capacity will contract against a build schedule, provided the site and the grid are real.
The group’s own framing, from Max-Hervé George, is that the company went from land, power and shell to chips, tokens and applications. That is a vertically integrated infrastructure play, and it explains the shape of the business. A company that owns the shell and the grid connection decides what goes inside it; a company that rents both does not.
The bitcoin-to-AI conversion
One part of the story deserves more attention than the release gives it. SWI Digital’s grid connections come largely from former bitcoin mining sites, now being converted to high-performance computing. That conversion has become a pattern across North America, where miners who spent years securing cheap power and interconnection rights discovered that those rights were worth more than the machines they were running. The economics favour the conversion because a mining site already has the substation, the interconnection agreement and often the cooling and shell. The open question is whether it holds up at AI density, which demands far more power per square metre and far tighter latency than hashing ever did.
Who it sells to
The customer thesis is narrower than a hyperscaler’s and more interesting for it. Europe has the researchers, the founders and the enterprises, the release argues, but the capacity that exists has gone first to global hyperscalers and frontier labs, the buyers who can sign multi-year contracts at scale. The companies with the most urgent need, AI natives with frontier ambitions and startup balance sheets, have the hardest time securing any. SWI Compute targets exactly that gap, offering dedicated and on-demand capacity under European jurisdiction, close to users and data. Its first customer is a US-based AI inference platform, which is a useful signal: the demand is not only European.
The economics it is betting on
The release contains one genuinely useful piece of operating logic. An AI factory earns on three things: how many tokens it produces per megawatt, how many years it keeps producing them, and how many kinds of work it can take on. That is the correct framing for the asset class, and it explains two decisions. The first is the choice to support pretraining, fine-tuning, agentic inference, simulation and scientific computing on the same footprint, because a facility that can only do one of those is worth less when that workload’s pricing falls.
The second is the adoption of NVIDIA‘s DSX reference design for every site from 2027, which the company says can enable up to 40 per cent more GPU capacity within the same power budget on Vera Rubin NVL72. If that figure holds in practice, it separates a marginal site from a good one, because power is the constraint and every extra GPU per megawatt is free revenue.
The caution
Two things in the record argue for patience. The first is timing. The UK campus’s first 30 megawatts arrive at the end of 2027, the 530-megawatt build-out follows, and the 2.3-gigawatt target is dated 2029. Nothing runs today, and the AI infrastructure market has a long history of plans that slipped. The second is SWI Group’s own recent record. In February 2026 it agreed to acquire a majority stake in Polarise, a German NVIDIA cloud partner, with a one billion euro commitment attached. By August both parties had mutually agreed not to complete that transaction, and SWI became a financier to Polarise instead.
A group that reverses a deal of that size within six months is a group whose plans deserve a second date on them. Neither point is disqualifying. A patient infrastructure investor can absorb both, and nobody questions the market’s appetite for capacity. What remains in question is whether the schedule holds.
What to watch
Three markers will show whether this is a platform or a plan. The first is commissioning: whether the UK hall delivers 30 megawatts on schedule at the end of 2027, because grid connections slip and construction slips more. The second is the customer mix: whether contracted capacity fills with AI natives and inference platforms, as claimed, or with the same hyperscalers everyone else sells to. The third is utilisation, and it is the marker nobody reports. A power-first strategy only pays if the GPUs on those sites stay busy, yet the company has disclosed neither the terms of its first contracts nor how much of the 3.6 gigawatts already belongs to the hyperscaler tenant on its European campus.
What it signals
The launch marks a shift in who builds AI infrastructure and what they bring to it. The first wave of neoclouds came from crypto and from cloud resellers, and their advantage was speed and access to GPUs. This wave comes from real-estate and infrastructure investors, and its advantage is land, grid and patience. That fits the capital intensity of the next few years, when a single campus costs more than most startups raise in total. The comparison worth holding in mind is with the hyperscalers’ own European build-outs, which are larger in aggregate but slower to serve the long tail. SWI Compute’s bet is that the long tail pays a premium for proximity, jurisdiction and speed of contracting, and that a 30-megawatt hall in the UK can fill before a 300-megawatt campus breaks ground.
It also raises a question Europe will have to answer. If the continent’s AI capacity increasingly sits inside a handful of listed investment vehicles that also lease to hyperscalers, then sovereignty becomes a matter of jurisdiction and contract rather than of ownership. SWI Compute sells European jurisdiction as a feature. Whether European buyers get the resilience they pay for depends on terms nobody has published.