Nscale accounts for $3.7bn of the $5.7bn raised by UK datacentre companies, Tracxn figures show
Funding concentrated in one AI infrastructure firm shows how capital, power and location decide where compute actually gets built.
A single AI infrastructure company, Nscale, has attracted $3.7 billion of the $5.7 billion in equity funding raised by the United Kingdom's datacentre infrastructure sector, according to figures from the research firm Tracxn reported by The Register on 10 September. That is roughly 65 percent of the total. Tracxn tracks 154 companies in the sector, of which 24 have raised equity funding, and Nscale is the only one valued above $1 billion.
The next largest fundraisers are some way behind: GreenScale at $1.3 billion, Kao Data at $177 million, Verne at $125 million, the liquid-cooling specialist Iceotope at $107 million and Six Degrees Group at $106 million. The money is geographically concentrated, with London-based companies accounting for 93 percent of the funding Tracxn tracked.
Exits have come mainly through acquisition rather than public listing. Tracxn records 36 acquisitions and two initial public offerings, the largest deal being Equinix's $3.8 billion purchase of TelecityGroup. The Register also cites data from Onnec and Barbour ABI showing nearly 100 UK datacentre projects worth about £36 billion ($49 billion) in the construction pipeline, and an Onnec survey of 300 decision-makers in which 43 percent of operators said they had needed infrastructure upgrades or remedial work after a facility launched. The article did not detail Nscale's individual rounds, investors, site capacities or customers.
Why it matters
The funding pattern shows that the capital flowing into datacentres is increasingly about GPU-dense AI facilities rather than conventional colocation, and that it pools around a small number of firms able to raise at scale. A pipeline of £36 billion is a statement of intent rather than delivered capacity, because each project still has to secure grid connections and planning consent. The finding that more than four in ten operators needed remedial work after launch suggests that facilities designed for one generation of hardware are being asked to host another.
- Nscale$3.7bn
- GreenScale$1.3bn
- Kao Data$177m
- Verne$125m
Figures: Tracxn figures as reported by The Register
What you can learn from this
- A neocloud is a capital business before it is a software business. Firms like Nscale buy accelerators by the tens of thousands, build or lease halls to house them, and rent the resulting compute to model developers. Because the hardware and the buildings must be paid for long before revenue arrives, these companies raise very large rounds of equity and debt, and their valuation depends on contracted demand. That is why one firm can absorb most of a sector's funding without being the largest operator by floor space, and why a single lost customer can matter more than it would for a software company.
- Location follows connectivity, then collides with power. London dominates because it has dense fibre routes, internet exchange points and customers who need low latency, and 93 percent of tracked funding going to London-based companies reflects that gravity. Yet the same concentration strains the local grid, so the physical sites for new AI capacity are often far from the head office. Investors follow the company address; electrons do not, and a funding map should never be read as a capacity map.
- A pipeline figure is not built capacity. A project counts as pipeline when it is planned, proposed or under construction, and £36 billion across nearly 100 projects covers all of those stages. Each must obtain a grid connection offer, planning consent and financing, and any one of those can slip by years. Reading pipeline as delivered megawatts is a common error when assessing how much compute a region will really have, and it flatters every national AI strategy that quotes it.
- Remedial work after launch is usually a density problem. A hall commissioned for racks drawing a few kilowatts each relies on air cooling and cabling sized for that load. AI servers can draw many times more per rack and often require liquid cooling, heavier power feeds and different network topologies. When tenants arrive with newer hardware than the design assumed, operators retrofit, which is one plausible reason 43 percent of surveyed operators reported post-launch upgrades.
- Acquisition, not IPO, is the usual exit. With 36 acquisitions against two listings, the sector behaves like other infrastructure markets: assets are valued on long-term contracted revenue, which suits strategic buyers and infrastructure funds more than public equity markets. Large operators grow by buying smaller ones, as the Equinix purchase of TelecityGroup illustrates, so consolidation is the expected end state for many of the 154 companies Tracxn tracks.
We teach this
How to use this in practice
- Model the power bill of a small GPU cluster. In a spreadsheet, take 1,000 accelerators, assume a per-device draw you state explicitly (for example 1 kW including server overhead), apply a power usage effectiveness factor you also state (say 1.3) to cover cooling and losses, and multiply by hours in a year and a price per megawatt-hour from your local market. Keep every assumption in its own labelled cell so a reader can challenge it. Done is a sheet where changing the PUE or the price instantly updates the annual cost, and a note on which input moves the total most.
- Find real grid headroom near you. Look up the published network capacity or connection-queue data for your regional electricity network operator; in Great Britain, distribution operators publish capacity maps and the system operator publishes connection queue information. Record the available headroom at the substation nearest a plausible datacentre site, and note how many megawatts the modelled cluster above would need. Done is one number, its unit, the date and the source URL written down next to your own demand figure.
- Draw the stack and place the named companies on it. Sketch layers from bottom to top: land and grid connection, building and cooling, racks and accelerators, orchestration and software, then customers. Place Iceotope at the cooling layer, Kao Data and Six Degrees at building and colocation, and Nscale spanning racks through orchestration. Done is a diagram that makes clear why funding at one layer does not translate into capacity at another. Mark which layers need planning consent and which need only a lease, since that decides how long each takes to deliver.
- Read one planning application end to end. Search a UK council planning portal for a recent datacentre application and read the design and access statement. Note the stated IT load in megawatts, the cooling method, the backup generation type and the proposed grid connection. Done is a five-bullet summary and a comparison of the stated load against the spreadsheet from the first exercise, with any gap explained. If the portal lists objections or conditions, note the most common one; it is usually about power, noise or water.
Sources
- Nscale swallows lion's share of UK datacenter investment — The Register
Our reporting is an original summary; full coverage is at the links above.
Don't just read about it — build it.
Square 1 teaches the skills behind the headlines, with every line of your work graded by AI. Find your starting point in 3 minutes.
Get your free skill report