TORONTO — Smaller cloud providers are fashioning themselves as AI delivery services, going beyond processing power to make it easier for businesses to use generative technology. 

Along with providing access to in-demand chips, some newer entrants help businesses choose, adapt and run open-source large language models (LLMs), simplifying the still complex and expensive task of using AI. That brings them into closer competition with the cloud arms of tech giants as demand for compute continues to grow. 

“There’s no question enterprises are playing with generative AI,” said Adam Hendin, CEO of Toronto-based Radium, but, he argues, it’s hard for them to deploy the tech.

Radium helps clients create their own spinoff models by fine-tuning freely available systems on internal data. Once ready, customers can use the models via plug-ins, and are charged when the system is pinged by a chatbot or another generative tool. 

Radium is “effectively creating private [LLMs] for enterprise in a completely automated way,” said Hendin. The firm’s software is designed to replace the work for which businesses might otherwise need dedicated machine learning, data science and hardware engineering specialists. The system also shows clients how much energy the data centre used to produce their modified model, and its carbon footprint. 

Founded in 2020, Radium currently lets clients build on top of models from Meta’s Llama family, as well as French provider Mistral. Hendin said it worked with those firms to optimize the technology so it’s more efficient, cheaper and more secure when deployed on Radium’s hardware. The company has data centre sites filled with Nvidia chips in Toronto, New York and Santa Clara, and is set to launch shortly in Denver.   

Calgary-based Denvr Dataworks also helps clients run open-source models. “The vast majority of the market is looking for an on-demand, usage-based service,” said CEO Geoff Gordon. That means customers don’t have to make decisions about their hardware or learn how to work with it.

While training LLMs from scratch requires huge amounts of processing power, most firms building AI products just need to adapt existing ones to their own data and use cases. Denvr lets clients work with models from Meta, DeepSeek and other providers, but can also host custom systems. It turns inference, the process through which an AI system produces an outcome like a chat response, into a service. 

The firm, founded in 2017, has three compute clusters in the U.S. and one in Canada. Customers include AI startups and developers, as well as companies that are building the technology into their products.  

Both Radium and Denvr also rent access to dedicated or shared servers and chips that clients can use to train and run models directly.  

The upstarts aren’t the only ones offering these kinds of AI services. The cloud arms of Amazon, Google and Microsoft all offer their own models alongside paid products from corporate and open-source providers. Collectively, Big Tech firms in the U.S. are spending tens of billions on new AI infrastructure, and launching more off-the-shelf AI tools.

Even with such significant competition, Canadian cloud providers still see plenty of room to grow, and lots of customers to serve. “We still have the luxury of scarcity,” said Gordon, noting that despite all the Big Tech data centres, demand still outstrips capacity, especially for model training. “We don’t really have a scenario where we’re getting ultra-competitive against some of these big companies.”

Denvr is one of a few companies in Canada with advanced AI capabilities, according to Gordon, making it attractive to clients that want to store and process information locally to maintain. “Anything that’s sensitive around data, we can win those deals,” he said.