OTTAWA — Artificial intelligence startups say the federal government shouldn’t slow down its $2-billion plan to help them access computing capacity by placing tight restrictions on where they buy their processing power.

While chip firms push for a “buy Canadian” requirement, many AI developers argue that a large chunk of the promised federal program should go to subsidies to help pay for the predominantly U.S.-powered compute that they’re already buying. A quick rollout would be a boon for the sector today, they say, even as Ottawa works to grow the country’s infrastructure in the long term.

For startups developing their own AI models, compute is “often the number one line item on their financials,” said David Katz, a partner at Radical Ventures, a Toronto investment firm that has backed a number of prominent companies in the field.

While there are plenty of AI startups in Canada, much of the compute they’re currently using is based outside the country. Foreign cloud services are “the only ones that offer the capacity that we or other companies need,” said Colin McKay, head of Canadian policy at Waabi, a Toronto-based self-driving startup. Firms often sign multi-year contracts with those providers. 

If Ottawa wants to roll out its AI infrastructure program quickly, it may have to subsidize Canadian startups to buy compute from foreign firms. That will raise concerns elsewhere in the tech sector. Chip firms argue the program needs “buy Canadian” requirements to ensure most of the money doesn’t just flow to U.S. tech giants like Amazon, Microsoft or Nvidia. 

But AI startups say the government shouldn’t dictate what hardware firms can use, Canadian or not. “We need to train our models with the technology that we use, that works today,” said Melika Carroll, head of global government affairs at Cohere, a Toronto-based startup building large language models (LLMs) that power generative AI tools.

Montreal-based Reliant AI initially tried out different hardware providers, and picked Nvidia because its software let the startup quickly start running its algorithms, said Marc Bellemare, the firm’s co-founder and chief scientific officer. “Any program designed to make us use a different kind of chip, [that] doesn’t allow us to use what’s the easiest for us to use, has a cost.”

Reliant estimates that renting compute capacity will account for about a quarter of its total spending next year. The startup sells software for pharmaceutical companies to analyze scientific and clinical data, using open-source LLMs it modifies with its own technology. 

Ottawa should focus on helping firms that are just getting started with the compute they need to try out their ideas, according to Bellemare. In January, for example, Reliant was working on some new features using a Google Cloud program that offers free capacity to startups. But the specialized chips it needs can cost thousands of dollars a week to rent, so the firm quickly ran through the credits it received. 

“It prevents innovation if you have to worry first and foremost about that bottom line,” Bellemare said, adding that a federal subsidy program would give startups like Reliant more time to experiment. He called for a program that would let Canadian firms pool their compute purchases, as well as work together on shared priorities like security and privacy standards for the data centres they use.

Federal funding could also affect where and how firms choose to expand. Subsidies would allow venture-backed companies to “use the money we’ve raised in more specific and productive ways,” said Waabi’s McKay. 

Waabi requires much less cloud time than an LLM developer to train its AI models, which direct autonomous trucks. Still, the firm wants Ottawa to “start delivering funds to support the compute costs of startups as quickly as possible,” McKay said. 

Katz said Ottawa’s compute program should offer larger payouts to companies that are scaling up models to sell to clients, where the costs of training can be “eye-watering.” 

AI executives say the federal AI infrastructure program could also help keep firms in Canada, as governments around the world court them with incentives. Cohere’s founders have committed to grow the firm in its home country, but Ottawa’s support “is super important to be able to do this in Canada,” Carroll said. She also emphasized the need to roll out the compute program quickly. “Innovation in this area is super-fast,” she added. 

Long-term, the federal government has said it wants to build “sovereign compute” capacity that’s owned and located in Canada. Ottawa could use an initial subsidy program to demonstrate to cloud firms how much compute demand there is here and encourage them to expand, Katz said. The federal program could then provide larger payouts to developers who use Canadian data centres, he suggested.

McKay said Waabi would be open to moving to a local provider that’s part of Ottawa’s sovereign compute strategy, as long as its hardware meets the startup’s technical needs and it offers faster access or lower prices.

Not all AI firms agree that Ottawa should be spending so much on infrastructure, however. 

Companies building their own mega-models are most likely to benefit from the federal compute program, said Long Dinh, CFO of Toronto-based Ada. But he predicted they will struggle to catch up to Silicon Valley players like OpenAI and Anthropic, which have raised billions of dollars to spend on training and are already booking hundreds of millions in revenue. “I just don’t think it’s the right battle to fight at this stage,” Dinh said.

Better, he suggested, to support firms selling AI applications. For example, Ada’s AI customer service “agents” use LLMs from OpenAI and other providers. Ada includes the expense of using those models in its prices, so it’s not running up huge compute costs that it can’t pay for.

Instead of putting $2 billion into compute, Dinh called for Ottawa to spend the money to buy itself AI tools from Canadian firms, which he said could make public services better and more efficient. Startups “need a market,” said Dinh. “We need adoption more than R&D at this point.”

But Katz insisted Ottawa’s bet on AI infrastructure is the right one, because compute is the most in-demand and expensive resource for the sector right now. “We won’t be innovating if we can’t afford the chips,” he said.