DeepInfra Debuts First International AI Data Center in Toronto with 1,000+ NVIDIA B300 GPUs

DeepInfra has launched a new AI data center in Toronto, marking its first expansion beyond the United States and bringing its total global footprint to nine facilities. The deployment represents a strategic shift in the company's infrastructure buildout, aligning with a broader industry pivot from AI model training toward production inference workloads.

The Toronto facility delivers 1.7 MW of critical power capacity and will host more than 1,000 NVIDIA Blackwell B300 GPUs. Once fully operational, the cluster is expected to significantly boost DeepInfra’s global inference capacity while reducing latency for customers in Canada and other international markets. The company owns and operates its GPU infrastructure, supports more than 200 open-source models, and currently processes nearly five trillion tokens per week across its managed inference platform, which serves both open and proprietary AI models.

This expansion reflects a fundamental shift in the AI infrastructure market. As enterprises move generative AI applications from pilot projects into full-scale production, demand is increasingly centered on inference rather than training. Large language model inference requires sustained GPU availability, low-latency networking, and geographically distributed infrastructure to support interactive applications, AI agents, and high-volume API services. McKinsey & Company research cited by DeepInfra projects that inference will account for more than 40% of total data center demand by 2030, growing at an estimated 35% compound annual rate.

DeepInfra said the Toronto deployment is designed specifically to address those requirements by bringing GPU resources closer to users and data while increasing overall platform capacity. The choice of NVIDIA’s Blackwell architecture further underscores the focus on inference, as B300 GPUs are designed to improve inference throughput and energy efficiency for large language models and multimodal AI workloads compared to previous Hopper-based systems. While DeepInfra did not disclose additional infrastructure details such as networking or storage architecture, deployments at this scale typically rely on high-bandwidth interconnects and distributed storage to keep GPUs fully utilized.

According to CEO and co-founder Nikola Borisov, enterprise AI has shifted from experimentation to production deployments, increasing the need for globally distributed infrastructure that delivers lower latency and scalable inference capacity. He said the Toronto cluster represents the company’s first expansion beyond the U.S. and positions compute resources closer to customers and their data. The deployment follows DeepInfra’s recent Series B funding and forms part of the company’s broader infrastructure expansion strategy. The company indicated that additional international data center deployments are under consideration as demand for GPU-based inference services continues to grow.