Nscale offers AMD AI chips-as-a-service in green data center

In the competitive realm of artificial intelligence, dominated by Nvidia’s formidable GPUs, a bold newcomer is betting on a different contender.

Nscale, a little-known GPU cloud provider based in Norway, has launched a new AI Cloud Services platform built around AMD’s MI300X accelerators. The move is a bold gambit to challenge Nvidia’s dominance and offer customers a potentially cheaper and greener alternative for running AI workloads in the cloud.

For buyers of cloud services and builders of AI models, Nscale’s embrace of AMD could signal a small shift in the competitive landscape.

“AMD may offer a cost advantage over Nvidia,” said University of Pennsylvania engineering professor Benjamin Lee. “Data center operators and their clients optimize the total cost of ownership, which includes both hardware purchases costs and operating energy costs, and AMD’s hardware costs may be lower than Nvidia’s. AMD may offer competitive or comparable performance for inference, the computation required when a trained model responds to a user prompt.”

Nscale’s vertical integration, spanning from its modular data centers to its high-performance compute clusters and its use of natural cooling solutions and low-cost renewable power, could potentially result in customer cost savings. However, the company must provide evidence of significant cost savings and performance advantages to attract buyers away from established providers.

Challenges and Opportunities in the AI Infrastructure Market

For AI model builders and cloud service buyers, Nscale’s offering represents a new option in the market. However, observers say the AMD MI300X accelerators and ROCm open software ecosystem have yet to establish themselves as a widely adopted alternative to Nvidia’s offerings regarding performance, ease of use, and developer support. 

Lee also discussed the implications for various stakeholders in the AI industry. “Developers must understand more precisely the nature of the AI computation. General data processing, training, and inference are different types of computation that stress different types of hardware in a data center server. Increasingly, developers might use one type of server for model training and another for inference,” he said. “In addition to AMD’s offering, AI developers will increasingly see data centers deploy other custom chips for inference from Intel, Microsoft, Google, Meta, and others.”

When asked how this move differs from other AI-focused computing infrastructures provided by hyperscalers like Azure, AWS, or GCP, Lee pointed to AMD’s long-standing efforts in creating and popularizing the ROCm software ecosystem. “Whether AMD’s chips will gain traction depends on whether ROCm provides sufficient support for inference computations compared to hyperscaler alternatives,” he noted.

Olivier Blanchard, Research Director from The Futurum Group, suggested several factors that may have influenced Nscale’s decision to work with AMD. “Nscale already has a good working relationship with AMD and decided to strengthen it by choosing their GPUs over NVIDIA’s,” he explained. Additionally, Blanchard pointed out that there might be a cost-benefit, as “NVIDIA GPUs tend to price high.”

Supply chain considerations could also have played a role in Nscale’s choice. “It could also be a supply chain decision: Nvidia GPUs can come with six-month lead times, and high demand could create supply bottlenecks. It is possible that AMD can provide better lead times and a lower risk of supply chain disruptions,” Blanchard noted.