DeepInfra has opened its first data center outside the United States, choosing Toronto for a facility that will run more than 1,000 Nvidia Blackwell B300 GPUs across 1.7 megawatts of capacity.
The Palo Alto-based company runs an AI inference cloud, meaning businesses rent access to its GPUs to run already-trained models rather than train new ones from scratch. Toronto marks DeepInfra’s ninth data center overall, but the first one it has built beyond US borders, and the timing lines up closely with a $107 million Series B round the company closed just before the announcement.
“Enterprises are moving from experimentation to production at unprecedented speed, and that shift demands infrastructure that is both scalable and globally distributed,” said Nikola Borisov, DeepInfra’s CEO and co-founder. He framed Toronto as the first step in a broader push outside the US, one aimed at letting customers run AI workloads physically closer to their own users and data.
That distinction matters more than it might seem. As companies move AI tools from pilot projects into actual production systems, latency and data location start to carry real weight, especially for businesses whose customers or regulatory obligations sit outside the US. A Canadian facility gives DeepInfra a way to serve that demand without asking customers to route everything through American servers.
Toronto, for its part, has become Canada‘s biggest data center hub, and DeepInfra is stepping into fairly crowded company there. Digital Realty, Equinix, Cologix, Hut 8, and several other operators already run facilities in the market, so DeepInfra is likely leasing space at an existing site rather than building its own from the ground up, though the company hasn’t confirmed those details yet.
DeepInfra has grown quickly since it launched in 2022, raising an $18 million Series A round in April 2025 before this year’s much larger Series B. The company currently runs on Nvidia’s Blackwell chips and has said it plans to add Nvidia’s upcoming Vera Rubin GPUs once they become available, positioning itself to keep pace with whichever hardware generation customers want to run their workloads on next.
