OTTAWA — Canada is missing crucial infrastructure necessary to be an AI power, and risks falling behind other advanced economies if it doesn’t secure a lot more compute capacity, according to a new report from the Dais, a Toronto Metropolitan University think tank. 

Here’s what you need to know:

Chips and servers: “Compute” is the tech-industry term for the processing power, storage and other infrastructure on which software runs. Artificial intelligence models, particularly the multi-purpose ones that power generative tools like ChatGPT, require large amounts to train and continuing quantities to run. But specialized chips or cloud services are expensive and in high demand. 

Figure this: The Dais estimates that Canada had 41 petaflops of compute capacity in academic or public-sector hands as of November 2023. The U.K. has about 82, while the U.S. has nearly 3,726; both have a lot more on the way.

Research organizations and AI executives say Canada needs more domestic capacity to ensure scientists and startups can afford and access the compute they need to make technological breakthroughs and develop new products. Industry players have been pushing Ottawa for a strategy.

The report suggests three ways Canada could catch up, or at least contend, on AI infrastructure: 

• Governments could bulk-buy or subsidize compute capacity from cloud providers for researchers and small firms. Program uptake would also allow policymakers to gauge the need for AI infrastructure in Canada, said report co-author Graham Dobbs, a senior economist at the Dais. As The Logic first reported, federal officials have discussed compute with the cloud giants.
• Canada should form a consortium with similarly inclined governments to jointly purchase capacity. Ottawa and London agreed to cooperate on the issue last month, although there’s no firm plan yet. Innovation Minister François-Philippe Champagne said last week that the G7 are considering pooling computing resources.  
• Ottawa could stock up on AI-specialized hardware to build more supercomputers at home. But that could prove expensive—chips don’t come cheap, and countries that make them may try to keep them. The U.K., for example, is spending  £1.7 billion ($2.89 billion) to build more public AI infrastructure.

The different programs “can all be started at the same time,” said Dobbs. The AI ecosystem will be looking to next month’s federal budget to see if any of them will.