David Clinton

Based mostly on their 2024 budget, the federal government has promised $2.4 billion in support of artificial intelligence (A.I.) innovation and research. Given the potential importance of the A.I. sector and the universal expectation that modern governments should support private business development, this doesn’t sound all that crazy.

But does this particular implementation of that role actually make sense? After all, the global A.I. industry is currently suffering existential convulsions, with hundreds of billions of dollars worth of sector dominance regularly shifting back and forth between the big corporate players. And I’m not sure any major provider has yet built a demonstrably profitable model. Is Canada in a realistic position to compete on this playing field and, if we are, should we really want to?

First of all, it’s worth examining the planned spending itself.

• $2 billion over five years was committed to the Canadian Sovereign A.I. Compute Strategy, which targets public and private infrastructure for increasing A.I. compute capacity, including public supercomputing facilities.
• $200 million has been earmarked for the Regional Artificial Intelligence Initiative (RAII) via Regional Development Agencies intended to boost A.I. startups.
• $100 million to boost productivity is going to the National Research Council Canada’s A.I. Assist Program
• The Canadian A.I. Safety Institute will receive $50 million

In their goals, the $300 million going to those RAII and NRC programs don’t seem substantially different from existing industry support programs like SR&ED. So there’s really nothing much to say about them.

And I wish the poor folk at the Canadian A.I. Safety Institute the best of luck. Their goals might (or might not) be laudable, but I personally don’t see any chance they’ll be successful. Once A.I. models come on line, it’s only a matter of time before users will figure out how to make them do whatever they want.

But I’m really interested in that $2 billion for infrastructure and compute capacity. The first red flag here has to be our access to sufficient power generation.

Canada currently generates more electrical power than we need, but that’s changing fast. To increase capacity to meet government EV mandates, decarbonization goals, and population growth could require doubling our capacity. And that’s before we try to bring A.I. super computers online. Just for context, Amazon, Microsoft, Google, and Oracle all have plans to build their own nuclear reactors to power their data centers. These things require an enormous amount of power.

I’m not sure I see a path to success here. Plowing money into A.I. compute infrastructure while promoting zero emissions policies that’ll ensure your infrastructure can never be powered isn’t smart.

However, the larger problem here may be the current state of the A.I. industry itself. All the frantic scrambling we’re seeing among investors and governments desperate to buy into the current gold rush is mostly focused on the astronomical investment returns that are possible.

There’s nothing wrong with that in principle. But “astronomical investment returns” are also possible by betting on extreme long shots at the race track or shorting equity positions in the Big Five Canadian banks. Not every “possible” investment is appropriate for government policymakers.

Right now the big players (OpenAI, Anthropic, etc.) are struggling to turn a profit. Sure, they regularly manage to build new models that drop the cost of an inference token by ten times. But those new models consume ten or a hundred times more tokens responding to each request. And flat-rate monthly customers regularly increase the volume and complexity of their requests. At this point, there’s apparently no easy way out of this trap.

Since business customers and power users – the most profitable parts of the market – insist on using only the newest and most powerful models while resisting pay-as-you-go contracts, profit margins aren’t scaling. Reportedly, OpenAI is betting on commoditizing its chat services and making its money from advertising. But it’s also working to drive Anthropic and the others out of business by competing head-to-head for the enterprise API business with low prices.

In other words, this is a highly volatile and competitive industry where it’s nearly impossible to visualize what success might even look like with confidence.

Is A.I. potentially world-changing? Yes it is. Could building A.I. compute infrastructure make some investors wildly wealthy? Yes it could. But is it the kind of gamble that’s suitable for public funds?

Perhaps not.