Nscale Ltd., a London startup that builds data centers optimized for artificial intelligence workloads, has raised $155 million to grow its infrastructure footprint.

The Series A round was announced this morning. Sandton Capital Partners led the investment with participation from Kestrel 0x1, Blue Sky Capital Managers and Florence Capital.

The funding milestone comes a few weeks after an AI cluster built by Nscale made it into the Top500 ranking of the world’s most powerful supercomputers. The Svartisen Cluster nabbed the 156th spot with 12.38 petaflops of maximum performance and 66,528 cores.

Nscale assembled the system from servers that each include six Advanced Micro Devices Inc. chips: two central processing units and four MI250X machine learning accelerators. The MI250X includes two graphics cards based on a six-nanometer process, as well as 128 gigabytes of onboard RAM for storing AI models’ data.

The servers are linked together by an Ethernet network that Nscale built using Broadcom Inc. silicon. The network implements a technology called RoCE can move data between two machines in a manner that bypasses their respective CPUs, which speeds up the flow of traffic. RoCE also automates tasks such as detecting overloaded network links and redirecting packets to other connections.

At the software level, Nscale’s hardware is powered by a custom infrastructure orchestration platform. It integrates Kubernetes with Slurm, a popular open-source framework for managing data center infrastructure.

Both Kubernetes and Slurm automate the task of determining what workload should run on which server in a cluster. However, they differ in several ways. Kubernetes has a self-healing mechanism, which allows it to recover automatically from certain types of malfunctions. Slurm, in turn, supports a network technology called MPI that can move data between different components of an AI workload with a high degree of efficiency. 

Nscale built the Svartisen Cluster in Glomfjord, a Norwegian village located inside the Arctic Circle. The data center (pictured) that houses the system draws power from a nearby hydroelectric dam and sits directly on a fiber-optic cable that links it to internet providers’ infrastructure. The cable features double redundancy, which means it can continue operating even if multiple important components fail.

The company makes its infrastructure available to customers in multiple ways. It provides AI training clusters as well as an inference service that automatically adds or removes hardware resources based on workload demand. There are also bare-metal infrastructure options, which enable users to more extensively customize the software stack that powers their deployments.

Customers may download AI models from an algorithm library provided by Nscale or bring their own. According to the company, there’s a preconfigured compiler toolkit for turning user workloads into a form that can run efficiently on its servers. 

Currently, Nscale’s construction pipeline comprises data centers with aggregate power consumption of 300 megawatts. That’s 10 times the amount of electricity used by the company’s Glomfjord facility. Using the Series A funding round announced today, Nscale will grow its pipeline by 1,000 megawatts. 

“The largest risk to the market’s ability to scale is the large contiguous tranches of electricity required to power these large GPU superclusters,” said Nscale Chief Executive Officer Joshua Payne. “Nscale has a 1.3GW pipeline of sites in our portfolio, which allows us to design from the ground up, the data centre, the supercluster and the cloud environment end-to-end for our customers.”

The company will build the new data centers in North America and Europe. According to the company, 120 megawatts’ worth of data center capacity is scheduled to be built next year. The new infrastructure will help Nscale power a planned public cloud offering focused on training and inference workloads that is set to launch in the first quarter of 2025. 

Photo: Nscale