UK-headquartered AI cloud Nscale has raised 3.36 billion dollars ahead of its US listing. Its prospectus shows roughly 25,000 GPUs live against 461,000 active and contracted, a useful measure of how far AI capacity announcements run ahead of energised reality.
What was announced
- Nscale filed for a US IPO on 18 September 2026 under the ticker NSCL, and on 25 September announced 3.36 billion dollars of pre-IPO funding led by Third Point, with NVIDIA, Apollo, Citadel and others participating. 2.36 billion dollars arrives up front, with a further 1 billion dollars from NVIDIA due by mid-November 2026.
- As of 31 August 2026 the company had approximately 25,000 active GPUs against 461,000 active and contracted GPUs, most of them planned as Vera Rubin clusters.
- Its active and contracted total contract value was 103.4 billion dollars, of which about 2.6 billion dollars was active. The contracts are mostly long-term take-or-pay with a weighted average life of about 5.7 years.
- Nscale has five active and 12 contracted datacentre sites, representing about 1.37 GW of active and contracted power. DCD reports the five active sites offer 55 MW.
- Revenue for the six months to June 2026 was 140.6 million dollars, up from 10.4 million dollars a year earlier. The net loss for the same period was 1.02 billion dollars, against 368.9 million dollars in the first half of 2025.
- Anthropic has agreed payments of up to about 44.6 billion dollars for dedicated Vera Rubin NVL72 capacity at Nscale's Monarch campus in West Virginia, where Phase I is planned at 2 GW and expected online in the first half of 2028.
The ETON view
The headline numbers are huge, but the useful figure for anyone buying capacity is the ratio. About 25,000 GPUs are running today against 461,000 on the books, and around 2.6 billion dollars of a 103.4 billion dollar contract book is active. Most of that pipeline is tied to a handful of frontier labs and hyperscalers on multi-year take-or-pay terms, and it depends on new sites, new power and new Vera Rubin supply arriving on schedule.
For hosting providers, enterprises and smaller AI companies, that has two practical consequences. First, the newest GPU generation will be allocated to the largest contracted buyers first, so plan on current and previous-generation platforms for anything you need in the next 12 to 18 months. H100, H200, L40S and RTX PRO class servers are available now and run comfortably in air-cooled UK halls at 6-12 kW per rack, which rack-scale liquid-cooled systems do not. Second, capacity headlines are not capacity. When a provider quotes you future GPU availability, ask what is energised today and what depends on a site that is still being built.
This is where an infrastructure partner earns its keep: sourcing proven hardware on a known lead time, mixing generations sensibly across training, inference and CPU workloads, and buying back kit as you move up. We can quote UK availability on specific GPU server configurations today rather than against a 2028 energisation date.
Related infrastructure
Category: Market · Vendor: NVIDIA · Technology: Vera Rubin NVL72, GPU cloud capacity, datacentre power · Last verified: 26 Sep 2026 · ~2 min read
Sourcing this kind of infrastructure? Talk to ETON about availability, lead time and pricing.
