My responses to the seventh theme of the Canada AI Survey, focusing on building and enabling infrastructure.
This is part 7 of “Responses to the Canada AI Survey” - a series containing my responses to the eight themes of the Canadian government’s public consultation on artificial intelligence.
Theme 7: Building and Enabling Infrastructure
Q1: Which infrastructure gaps (compute, data, connectivity) are holding back AI innovation in Canada, and what is stopping Canadian firms from building sovereign infrastructure to address them?
I don’t think compute is holding back innovation in Canada. It is a commodity which is easily accessed across borders from private enterprise such as AWS. The price of GPU-enabled compute has actually fallen for late-generation chips. There are signs of a disconnect between rental prices and purchase pricing which, for a depreciating asset, is problematic. This is a serious consideration for any public infrastructure build-out.
Data has, for a long time, been digital gold. There has been a large effort to access data in the healthcare technology space, particularly medical images after AlexNet in 2012. However, with the advent of natural language processing and now LLMs, unstructured patient data is also sought. This is data that is not available on the web. Canada should guard all its medical data from US firms if it seeks an advantage. We are behind in this respect. We actually use other country’s data in some instances for non-AI guidance, such as bone-mineral density screening. This is an ongoing effort with many barriers remaining.
All verticals have valuable data housed in silos. Data owned by Canadian firms (and ostensibly coming from Canadians or Canadian activities and processes) should be leveraged to a Canadian advantage. It may make sense for industries to cooperate and the government could have a role in establishing neutral parties to aid this cooperation, both from legal and technical (e.g. data lakes) standpoints.
Connectivity varies widely across the country, especially in rural areas. We should address this immediately. Rural connectivity projects often take many years after commitments are made to achieve the actual connections. This is too slow.
Q2: How can we ensure equitable access to AI infrastructure across regions, sectors and users (researchers, start-ups, SMEs)?
Access is currently equitable in the sense that it’s mediated by internet connectivity and money for anything in the current “LLM stack”. Outside of this stack (research), we need to follow the lead of these researchers and their academic institutions. SMEs may want to access such specialized stacks if some capability arises that isn’t already on the open market. At this point, I think it is a question of market forces - the cost of expanding this stack vs. the revenues it could create. So long as it falls short of a renewed promise of an imminent AGI which requires immediate, massive scaling, as is the case with current LLMs.
Q3: How much sovereign AI compute capacity will we need for our security and growth, and in what formats?
It’s long been clear that in a globalized world, Canada is unable to onshore everything, especially as so much focus is on resource extraction and automotive manufacturing.
At minimum, Canada should have sovereign Content Delivery Networks that are located and wholly owned by Canadians. The same for data centers.
However, it is likely unreasonable to manufacture our own chips or even replicate the data center build out happening elsewhere.
There is a potential scenario where the buildout south of the border has happened too soon and too quickly. This could be an advantage, where these resources could be accessed for cheaper than it would have cost to build them. We could even purchase them at a steep discount and power them with our own green electricity.
Lastly, I’m going to challenge the assumption that more compute equals more security. If more compute is needed because of more use and more use means more applications and each application is a new, novel attack surface, perhaps more compute equals less security.