Highlights:

  • In terms of software, a specialized infrastructure orchestration platform powers Nscale’s hardware. It combines Slurm, a well-liked open-source platform for data center infrastructure management, with Kubernetes.
  • Data centers with a combined power consumption of 300 megawatts are currently part of Nscale’s building pipeline.

A London-based startup Nscale Ltd. raised USD 155 million to expand its infrastructure base. The company develops data centers optimized for artificial intelligence operations.

The Series A funding round was announced recently, with Sandton Capital Partners leading the investment. The round also saw participation from Kestrel 0x1, Blue Sky Capital Managers, and Florence Capital.

A few weeks after an AI cluster constructed by Nscale was included on the Top500 list of the most potent supercomputers in the world, the funding milestone was reached. The Svartisen Cluster made it to the 156th position with an optimum performance of 12.38 petaflops and 66,528 cores.

The system was put up by Nscale using servers that each contained six Advanced Micro Devices Inc. chips: four MI250X machine learning accelerators and two central processor units. In addition to 128 gigabytes of internal RAM for storing the data from AI models, the MI250X has two graphics cards built on a six-nanometer technology.

Nscale used Broadcom Inc. chips to build an Ethernet network that connects the servers. By using a technology known as RoCE, the network may transfer data between two computers without using either machine’s CPU, which speeds up traffic flow. Additionally, RoCE automates processes like rerouting packets to alternate connections and identifying overburdened network lines.

In terms of software, a specialized infrastructure orchestration platform powers Nscale’s hardware. It combines Slurm, a well-liked open-source platform for data center infrastructure management, with Kubernetes.

Slurm and Kubernetes both automate the process of deciding which server in a cluster should handle which job. They are different in a few ways, though. Because of its self-healing system, Kubernetes can automatically recover from some kinds of faults. In response, Slurm supports a network technology called MPI, which has a high degree of efficiency in transferring data between various AI task components.

In Glomfjord, a Norwegian town inside the Arctic Circle, Nscale constructed the Svartisen Cluster. The system is housed in a data center, which is directly connected to the infrastructure of internet providers via a fiber-optic connection and receives power from a nearby hydroelectric dam. Because of its double redundancy, the cable can function even in the event that some crucial parts fail.

Customers can use the company’s infrastructure in a number of ways. It offers AI training clusters and an inference service that adjusts hardware resources automatically according to workload requirements. Options for bare-metal infrastructure are also available, allowing customers to further alter the software stack that drives their deployments.

Clients can get their AI models or access from Nscale’s algorithm library. The business claims to have a preset compiler toolset for converting user workloads into a format that its servers can handle effectively.

Data centers with a combined power consumption of 300 megawatts are currently part of Nscale’s building pipeline. The company’s Glomfjord operation uses ten times as much electricity as that. Nscale plans to expand its pipeline by 1,000 megawatts with the help of the Series A fundraising round.

“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 center, the supercluster and the cloud environment end-to-end for our customers.”

The company will construct new data centers throughout Europe and North America. It announced that next year, 120 megawatts of data center capacity will be constructed. Nscale plans to launch a public cloud offering in the first quarter of 2025 that is targeted at training and inference workloads, and the new infrastructure will help power this offering.