Canada’s best-capitalized large-language model (LLM) developer, Cohere, thinks the media firestorm surrounding the launch of the latest chatbot app from Chinese artificial intelligence (AI) company DeepSeek proves its point that efficiency is just as good of a predictor of innovation as eye-watering amounts of capital.
“Innovation and efficiency, not excessive compute, is the key.”
Nick Frosst
Cohere
“People have been chasing the wrong rabbit in LLM development, thinking more compute is going to always lead to breakthroughs,” Cohere co-founder Nick Frosst said in a statement to BetaKit. “But now folks are coming around to our point of view: innovation and efficiency, not excessive compute, is the key.”
DeepSeek’s latest LLM product, released on Jan. 20, runs on older chip technology, is free to use, and is comparable in many ways to other consumer-facing models like OpenAI’s ChatGPT—but it appears to have been developed for a fraction of the cost. As a result, US AI tech stocks plunged this week, prompting Canadian AI leaders to consider how domestic companies with limited resources can compete on the world stage.
“It’s not about unlimited resources but about smart, efficient solutions,” Cohere CEO Aidan Gomez said in a statement to The Canadian Press. “We’ve believed this for a long time, but it’s finally hitting home across the industry.”
Cohere is among the most funded Canadian AI companies, having secured $500 million USD in financing at a $5.5-billion valuation last year ($687 million at $7.6 billion CAD), not to mention the $240 million CAD committed by the federal government towards a Cohere data centre. In comparison, Meta plans to spend up to $65 billion USD on AI infrastructure this year.
RELATED: Cohere Series D hits $500 million USD raised at $5.5-billion valuation
The latest iteration of DeepSeek’s model, DeepSeek R1, exceeds or is on par with the performance of OpenAI’s GPT-4o in coding and quantitative reasoning, according to the independent comparison tool Artificial Analysis. Meanwhile, the company claims it was built using a fraction of OpenAI’s capital—according to DeepSeek, just $5.6 million USD. The massive sell-off of tech stocks after this came to light wiped billions in value from key American AI chip companies such as Nvidia.
DeepSeek’s assertion has been met with industry skepticism that it spent so little to develop the model. It has also sparked existential dread that AI innovation may not require the level of capital some AI giants have raised.
In a post on X, Gomez said that the people who think DeepSeek built this model with under $6 million are wrong, but those who think Big Tech companies were humiliated by their massive AI investments have a point.
The Canadian AI value chain
Julien Billot, CEO of the federally funded innovation hub Scale AI, told BetaKit that DeepSeek’s achievement is an indication that Canadian companies still have “room to play” in the AI space, despite the dominance of American incumbents. Billot noted Cohere as an example of a Canadian company worthy of support. “That’s our best bet to exist in this AI market,” he said.
However, he added, the Canadian government and private investors need to be smart about their investments by controlling and retaining as much of the AI value chain as they can within Canada.
Canada ranks fourth globally in terms of generative AI companies per capita, according to a Deloitte report. In addition to Cohere’s $500 million USD, Canadian autonomous driving firm Waabi raised $275 million CAD and text-to-image generator Ideogram raised $80 million USD in 2024, while Toronto-based Radical Ventures closed an $800-million USD fund for AI startups.
RELATED: In 2024, Canada struggled to find its place in the global AI race
DeepSeek disrupted Silicon Valley’s narrative that bigger investments will necessarily lead to better products, Billot said. He added this is a good thing for competition, particularly in Canada, since Canadian VC funding does not match the levels of capital raised by the likes of Anthropic or OpenAI. OpenAI was estimated to spend up to $7 billion USD on AI training and inference in 2024 to power ChatGPT, according to reporting from The Information.
“The best way to control our destiny is to own as much as we can of the value share in AI,” Billot said in an interview with BetaKit.
In contrast to the statements of Cohere’s CEO, Billot emphasized that achieving that goal will require domestic investment in yet more compute. He said supporting Canadian startups and building local demand for AI infrastructure and applications will contribute to a healthy AI value chain within the country, from semiconductor chips to software applications. Solutions include building Canadian data centres and creating more Canadian intellectual property (IP), he said.
“The best way to control our destiny is to own as much as we can of the value share in AI.”
Julien Billot
Scale AI
Billot called specifically for increased government spending on AI infrastructure, such as an extension of its $2-billion Canadian Sovereign AI Compute Strategy.
French private-equity firm Jolt Capital, which recently invested in the Montreal-based semiconductor IP company Dolphin Semiconductor, countered the narrative that “America owns AI” in a LinkedIn post.
“DeepSeek AI’s feat reminds us that brute force and hundreds of billions thrown at a challenge are not always the smartest way,” the company stated, adding that Jolt Capital plans to fund non-US AI startups who “keep their hubris under control.”
Intense energy
DeepSeek’s model is apparently less energy-intensive than American incumbent models, indicating a pathway forward for mitigating the energy consumption costs of AI and its associated infrastructure.
According to a brand-new international AI safety report chaired by Mila Research Institute co-founder and AI scientist Yoshua Bengio, carbon emissions and water usage are serious safety concerns surrounding LLMs because of the huge amounts of energy consumed by data centres. Data centres are projected to become the fifth-largest consumer of electricity in the world by 2026, according to an Organization for Economic Co-operation and Development (OECD) estimate.
However, Sasha Luccioni, AI and climate lead at New York-based Hugging Face, noted that DeepSeek’s environmental impact could climb as its usage increases, despite taking less energy to train.
Billot said that Scale AI told the federal government it would be willing to fund efforts for Canadian companies to lower their energy consumption for software.
“The race for energy is a big thing,” Billot said. “Having more energy-efficient software is really critical for the industry.”
With files from Josh Scott. Feature image courtesy Web Summit via Flickr.