DeepSeek Shows Why Canada's AI Strategy Is Off-Track

Canada can't compete in the current paradigm - we should be looking to reshape it instead, as DeepSeek has done. Plus some quick hit reading recommendations.

After last week’s brief hiatus, today I look at what the new, low-cost AI model from Chinese firm DeepSeek says about Canada’s AI strategy. I then highlight a few quick hit recommendations, some new and some a bit older, that are still worth reading.

Chasing The Pack Will Only Leave Us Always Chasing

Photo by Igor Omilaev on Unsplash

If you have been following AI news, you will have seen the revelation of DeepSeek’s new AI model. It seems to have been built for far less cost than the most advanced AI models from US firms, is cheaper to use than them, and is as good, if not better, in many benchmarks.

This has created shockwaves in financial markets. On Monday, Nvidia shares fell 17%, shaving over half a trillion dollars off its market cap. TSMC and Oracle also lost 14%, and other big tech shares fell. Canadian firms have been caught up, too, with Celestica, a company that builds data centre equipment, losing 28% and Power Corp, an energy company, losing almost 17%.

DeepSeek’s model reportedly only cost US$5.6 million to train, took two months, and it was done without the latest cutting-edge Nvidia chips that other firms deem essential for training new models.

It is difficult to overstate how much this upends things. As AI reporter Karen Hao has argued, “The biggest lesson to be drawn from DeepSeek is the huge cracks it illustrates with the current dominant paradigm of AI development.”

For Hao, DeepSeek’s model has “demonstrated that scaling up AI models relentlessly, a paradigm OpenAI introduced & champions, is not the only, and far from the best, way to develop AI.” While this approach can lead to better performance, it also comes with huge externalities, with their massive power and water needs, which are helping accelerate emissions and the climate crisis. What DeepSeek shows, though, is that the trade-off of faster development for short-term costs that OpenAI and other big tech firms “frame as wholly necessary is actually not.”

This is game-changing. As Hao argues:

Of course, there is much more to the DeepSeek story. The geopolitics of the US-China AI rivalry, what DeepSeek’s business model actually is, Chinese censorship, what this means for AI safety, and the value of deep moats vs. open-source models - DeepSeek’s breakthrough creates lots of questions and has lots of ramifications. If you want to understand it more fully, then Casey Newton’s Platformer piece is worth a read for a rundown of it all. 

However, I want to focus on the question of investing in more efficient methods. I think the DeepSeek example highlights the flaws in Canada’s recent approach to AI, particularly the $2 billion Canadian Sovereign AI Compute Strategy. As I’ve written about before, in the context of the OpenAI paradigm of massively scaling compute costs, “$2 billion is a drop in the water even if you are focused on just supporting one leading-edge company,” which the Strategy isn’t - it is spread across a huge range of the value chain and different stage firms.

The result is that we have been setting ourselves up to play by the rules of a game that we’re never going to win with that kind of commitment. We’ll only ever be chasing the pack and never leading with that kind of approach.

Instead, as DeepSeek has demonstrated, if you’re behind, then you need to change the rules of the game.

To do this, you need to return to first principles and articulate a clear vision of your goals.

Going back to first principles means, for me, actively considering how to build an inclusive and sustainable society. This shouldn’t be an afterthought, a hoped-for outcome as we plow money into innovation and where we create huge negative externalities for the environment or inequality and then look to correct them after the fact. Instead, we should intentionally tip the scales towards innovations that are in line with the outcomes we want to see.

When it comes to AI, that could mean, as I argued in October, “priority funding for firms that are working to reduce the computing power needed to run large models or that are focused on technology to make data centers more efficient.” It could also mean trying to find and support a Canadian DeepSeek that delivers an efficient yet powerful model.

An Efficient AI Challenge Prize

Instead of throwing $2 billion over five years at building compute capacity in a race we’re destined to lose, what if that $2 billion was used as a challenge prize for Canadian firms that can outperform leading benchmarks with minimal power and water usage?

You could put aside some of this as direct financial rewards for the people behind the firm(s) who are able to meet the challenge (not just management but all the team). But then you could also seek to use the prize to supercharge success. A substantial portion could be earmarked as a tax-free prize to support ongoing R&D and help them stay ahead of the pack. You could also require that another substantial portion is dedicated to marketing and sales - something that Canadian start-ups under-prioritize compared to US firms, as Charles Plant has argued, holding them back from scaling.

If you then combine this with other policies that support the development and diffusion of this technology (such as procurement policies that prioritize AI efficiency and safety), then you potentially begin to get at a more holistic AI strategy that leads to outcomes we want to see. If we can’t compete in a straight-up compute subsidy race (which we clearly can’t), then where can we deploy what are still substantial sums of money to change the rules of the game and help put Canada ahead?

A challenge prize could tick these boxes. We certainly should not be going down the path of trying to own all parts of the AI value chain and not backing up with the resources needed to make a difference. That will just doom us to forever chasing the pack.

Quick Hits

Canada must hit the U.S. where it hurts most: its lucrative patents - An interesting argument from Richard Gold, the director of McGill University’s Centre for Intellectual Property Policy, that we should be targeting US IP as a response to Trump tariffs. This would certainly be a strategy that would hurt the US, and hurt some major backers of the President in Silicon Valley more than tariffs on goods would.

Inside Canadian tech’s not-so-quiet shift to the right - Speaking of Silicon Valley, much like in the US, Canada’s tech leaders have been moving to the right. This long read by Catherine McIntyre, Laura Osman, and Murad Hemmadi in The Logic explores the growing disillusionment of tech leaders towards the Liberal government and their shift towards supporting, and funding, the Conservative Party. I think there are plenty of legitimate criticisms that those executives have raised towards Liberal policies over the years, such as their lack of vision and their failures in execution.

However, it raises flags for me when the views of a select, incredibly rich, and incredibly male group of people are given the space to speak on behalf of the “innovation-sector” and “innovation community” as a whole. Tech firms don’t represent all types of innovation, and tech executives don’t disinterestedly represent the interests of the sector and its various workers and stakeholders. They certainly are important and influential actors, but they also have incentives that would lead them to certain models of innovation and tax policy, and represent a certain type of vision of what innovation in Canada should be.

Better strategies in government - Another piece on policy coherence in the UK, this time by Nesta’s CEO,

Ravi Gurumurthy

. Gurumurthy argues that “Compared with 15 years ago, the [strategy] capacity available to this government is weak and underpowered. There is no equivalent to the Prime Minister’s Strategy Unit, the Social Exclusion Unit that I was part of, or the Office of Climate Change that I set up – a shared resource across departments that drew up the Climate Change Act.”

Those units had precious ingredients for making good strategy. They were not involved in day-to-day policy delivery so had the headspace and independence to think critically. Projects lasted for 6-9 months so teams had the time to do proper analysis. The units had a large number of outsiders – practitioners, researchers, and strategy consultants who brought skills and expertise. Because of where they were located and who was in them, their ethos was different. Instead of being focused on how to satisfy key stakeholders, the approach was to look at issues from first principles.

I’m not sure what the state of this kind of strategy function is in Canadian governments, federally and provincially but my sense is we also currently lack these kind of capabilities.

Gurumurthy goes into detail on how to improve government strategy. It is a must-read for those interested in building state capacity.

Coasian hecks, or, when the people in charge can’t change things either - Closely related to the issue of state capacity is this 2023 piece from Sean Boots. I had one comment on last week’s newsletter relating to Michael Wernick’s recommendations on government reform that asked whether “it is systematically difficult/impossible to advance this type of thinking about the public service when you are in a senior role within the system.” Boots’s piece explores exactly this, asking why various former Clerks, Wernick included, didn’t “do something about these issues when they were in charge?” For Boots, this is because “the public service is an example of an organization that has been structured in a way that prevents it from changing itself.” Boots explores the issue in-depth and also provides some strategies for operating in a system where no one is “clearly empowered to structurally change the environment itself.” Got to love some actionable advice!