Nscale
by Nscale
Vertically integrated European AI cloud, from owned data centres to serverless inference
Nscale is a London-headquartered AI cloud that owns and co-locates its own data centres and sells GPU capacity, managed training platforms and serverless inference on top of them. It targets organisations training or serving large models at scale, particularly those that want European capacity and sovereignty rather than a US hyperscaler region.
Nscale is a vertically integrated AI cloud built for large-scale training and inference, founded in May 2024 by Joshua Payne and Nathan Townsend and headquartered in London. It spun out of the Australian crypto-mining operator Arkon Energy and now owns or co-locates data centres in Glomfjord and Narvik in Norway, Loughton in the UK, and Texas and West Virginia in the United States, with partner facilities in Portugal, Iceland and elsewhere; sites use closed-loop liquid cooling and target a PUE of 1.1 to 1.15. The stack sells at three levels. Infrastructure services provide on-demand GPU and CPU compute, InfiniBand and RoCE fabrics and AI-optimised parallel storage. Platform services add managed Slurm for distributed training, a Kubernetes service for containerised workloads and bare-metal instances. AI services sit on top as serverless and dedicated inference endpoints behind OpenAI-compatible APIs, plus a prompt workbench and fine-tuning; the serverless platform launched in April 2025 and is billed per million tokens across models including Meta's Llama, Alibaba's Qwen and DeepSeek. Control Center, observability tooling and a Radar usage API handle orchestration and reporting. What makes Nscale significant is contractual rather than technical: in October 2025 it agreed to supply Microsoft with roughly 200,000 NVIDIA GB300 GPUs, including about 104,000 in a 240MW Texas campus, and it works with NVIDIA and OpenAI on the UK Stargate project and with BT on 14MW across three UK sites. Funding has been extraordinary — a $155M Series A in December 2024, a $1.1B Series B in September 2025 backed by NVIDIA, Nokia, Dell and Aker, a $433M pre-Series C SAFE, and a $2B Series C in March 2026 at a $14.6 billion valuation, with Sheryl Sandberg, Susan Decker and Nick Clegg joining the board.
The infrastructure or ML platform leader who needs large contiguous GPU clusters with InfiniBand for training runs, and wants that capacity sited in Europe for latency, energy cost or data-sovereignty reasons.
Access to reserved multi-thousand-GPU NVIDIA capacity in owned, liquid-cooled European data centres, with managed Slurm and Kubernetes above it so the cluster is usable without building the orchestration layer yourself.
At a Glance
- Category
- Infrastructure & Cloud
- Pricing
- Usage-based, Contact for pricing
- Target Market
- CTOs, CIOs, ML Platform Engineers, Heads of Infrastructure
- Deployment
- Cloud-first, API-based, Multi-cloud
- Founded
- 2024
- Headquarters
- London, United Kingdom
Key Features
- ✓Owned and co-located AI data centres
Facilities in Norway, the UK and the US with closed-loop liquid cooling targeting 1.1–1.15 PUE, giving control over power cost and siting rather than reselling someone else's racks.
- ✓Bare-metal and on-demand GPU compute
Direct access to NVIDIA GPU fleets with InfiniBand and RoCE interconnect and parallel storage engineered for predictable throughput during long training runs.
- ✓Managed Slurm
Distributed training scheduling delivered as a service, removing the substantial work of standing up and operating a Slurm cluster on raw GPU nodes.
- ✓Kubernetes service
Containerised AI workloads run on managed Kubernetes over the same GPU fabric, so inference services and training jobs share one substrate.
- ✓Serverless inference with OpenAI-compatible APIs
Launched April 2025, it serves Llama, Qwen and DeepSeek models billed per million tokens with no idle capacity to pay for.
- ✓Control Center, observability and Radar API
Environment management, monitoring and a usage-and-events API so consumption across clusters and endpoints can be tracked and charged back internally.
- ✓Fine-tuning and prompt workbench
Managed fine-tuning and an interactive prompt environment let teams move from experimentation to production without provisioning separate infrastructure.
Capabilities
Use Cases
- •Frontier and foundation model training
Reserve thousands of interconnected GPUs for multi-week pretraining runs where cluster topology and storage throughput determine wall-clock time.
- •Sovereign European AI capacity
Run training and inference inside EU, UK or Nordic facilities so regulated data never leaves the jurisdiction it is governed by.
- •Hyperscaler capacity offtake
Microsoft contracted roughly 200,000 NVIDIA GB300 GPUs from Nscale, including a 240MW Texas campus, to add capacity faster than building it directly.
- •Production inference on open-weight models
Serve Llama, Qwen or DeepSeek behind OpenAI-compatible endpoints and pay per million tokens instead of holding reserved GPUs idle.
- •Fine-tuning open models on proprietary data
Use managed fine-tuning against in-region data, then promote the resulting model to a dedicated inference endpoint on the same platform.
Ideal For
Best For
- ✓Large-scale distributed model training that needs contiguous GPU clusters on InfiniBand or RoCE rather than scattered on-demand instances
- ✓European organisations that need AI compute physically located in the EU, UK or Nordics for sovereignty, latency or energy-cost reasons
- ✓Teams wanting managed Slurm or Kubernetes over bare-metal GPUs without building the scheduling and storage layer themselves
- ✓Serverless inference on open-weight models such as Llama, Qwen and DeepSeek through OpenAI-compatible APIs with per-token billing
- ✓Multi-year reserved capacity deals where a hyperscaler cannot commit the GPU volume on the required timeline
Not Ideal For
- ✗Small teams or individual developers wanting transparent self-serve pricing — Nscale publishes no GPU rate card and large capacity is negotiated, unlike Lambda or RunPod
- ✗Buyers who need a broad managed-service catalogue beyond compute; this is AI infrastructure, not a general cloud with databases, identity and analytics services
- ✗Risk-averse procurement that requires a long operating history. The company is barely two years old, spun out of crypto mining, and Hacker News commenters have flagged construction delays — one noting a UK site scheduled to go live in 2026 was still 'a scaffolding yard in Essex' that March
- ✗Organisations needing a published SOC 2 or ISO 27001 attestation before signing, which Nscale does not advertise publicly
Integrations
Deployment
Market Analysis
Pros
- ✓Owns its data centres rather than reselling, so power, cooling and buildout are under its own control — Norwegian hydro sites at 1.1–1.15 PUE are a genuine cost advantage
- ✓Capital and demand are both proven at scale: a $2B Series C in March 2026 at $14.6B valuation, and a roughly 200,000-GPU GB300 supply agreement with Microsoft
- ✓Strategic backing from NVIDIA, Nokia, Dell and Aker means preferential access to hardware supply that smaller neoclouds struggle to secure
- ✓Covers the full stack from bare metal through managed Slurm and Kubernetes to token-billed serverless inference, so training and serving live with one vendor
- ✓European and UK capacity is a real differentiator for organisations with data-sovereignty constraints
Cons
- ✗Delivery is running behind the announcements. Hacker News commenters say Nscale 'is very behind on their planned datacenter, and will miss the deadline', with one calling the UK supercomputer project 'an obvious scam' after finding a site due live in 2026 still 'a scaffolding yard in Essex' that March
- ✗Headline investment figures are softer than they read — a UK government $2.5bn commitment was described as 'not a formal contract, rather an intention to commit capital'
- ✗No published GPU rate card, so self-serve buyers cannot price it against CoreWeave, Lambda or RunPod without a sales conversation
- ✗Barely two years old and spun out of a crypto-mining operator; the Financial Times has questioned its 'British' framing given Australian origins and founder
- ✗No public SOC 2, ISO 27001 or HIPAA attestation, which is a gap for regulated buyers evaluating it against a hyperscaler
- ✗Concentration risk cuts both ways — a business anchored on a small number of very large contracts is exposed if one of them moves
Pricing
Serverless Inference
Usage-based, per 1M tokens
- ✓Pay-as-you-go per million input and output tokens
- ✓Llama, Qwen and DeepSeek models
- ✓OpenAI-compatible APIs
- ✓Image models priced by size and steps
- ✓No idle capacity charges
Dedicated inference and platform services
Contact for pricing
- ✓Dedicated endpoints
- ✓Managed Slurm and Kubernetes
- ✓Bare-metal GPU instances
- ✓InfiniBand/RoCE networking and parallel storage
Reserved GPU capacity
Contact for pricing
- ✓Multi-year contracted clusters
- ✓Dedicated or co-located data centre capacity
- ✓Custom networking and storage design
- ✓Negotiated commercial terms
Only the serverless inference tier has public, self-serve economics: pay-as-you-go per million tokens covering both input and output for chat, multimodal, language and code models, with image models priced by output size and step count. Everything else — bare metal, managed Slurm, Kubernetes, dedicated endpoints and reserved clusters — is quote-based, and Nscale publishes no GPU hourly rate card, so there is no way to compare it against Lambda or CoreWeave without engaging sales. The economically meaningful business is multi-year contracted capacity such as the Microsoft GB300 agreement, which implies commitment levels and lead times, not on-demand consumption.
Security & Compliance
Sources
This page was written from 6 sources, 3 on domains other than nscale.com.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
WEKA NeuralMesh
Microservices storage and memory fabric built for AI training and inference at exabyte scale
Tsuga
Bring-your-own-cloud observability that keeps telemetry, and its cost, inside your own AWS account
OpenObserve
Open-source, Rust-based observability on object storage — logs, metrics, traces and LLM telemetry in one binary
Coralogix
AI-native observability that queries logs, metrics and traces in place — index-free, in your own S3 bucket