It’s important for citizens to understand the relative strengths and weaknesses of the Canadian and US economy before they decide that the country needs a ‘deal’. I’ve brought together a bunch of background info that I haven’t heard much of from the legacy media. I think it suggests Canada is doing a whole lot better than America right now. https://billhulet.substack.com/p/my-hot-take-on-the-current-trade
We don’t need the killer app. we just need one that holds its own. $7B is still a lot pf capitalization. If we grow that, it will suffice for the Canadian market. But apart from making it big in the investment market (like Red Hat Linux did in the dot.com era) your AI megaliths do not show a lot of financialization. They don’'t seem to be producing a solid return on the investment dollar. For all its hype, very few have demonstrated how to actually make money by using it. Great for aggregating research, but how do you make money from that research? Facebook makes their money by paid advertising and paid promotions. How do you translate that model into AI? For META, Google, and Amazon, it is just a different way to do what they are already doing.
This isn’t really a matter of killer apps. It’s a general purpose technology. And, to be clear, Cohere is a good business and they’re doing well focusing on enterprise. I’m not knocking them. Still, they’re on a similar timeline to profitability as Anthropic, at wildly different scale. Cohere was around $250M ARR last year. Anthropic is on track for something like $60B ARR this year and $140B ARR next year. The idea that profitability isn’t there is just wrong. And that’s just the model providers, but if you look at Cohere’s business model that has enterprises invest in sourcing their own instead of getting inference from Cohere, who are they getting the compute from? They’re either buying GPUs/TPUs made by a US tech giant, or they’re renting compute from a US tech giant, or if they’re actually renting compute from a Canadian tech giant then that tech giant is buying GPUs/TPUSls from a US tech giant. No matter how you slice it, US tech is winning.
But here is where the opportunity for Canada lies.
https://www.buildcanada.com/memos/semiconductor
By that logic, the end benefit goes to Taiwan, because that is where Nvidia outsources their chips to.
Of course, TSMC is also doing incredibly well out of manufacturing, but not as well as Nvidia which still has like 75% margins on their GPUs and nets like 50% on them. Not sure what Google’s margins are on their TPUs, but I’m sure they’re also high and they are making huge returns on serving inference.
As to the Canadian compound semiconductor fab, it seems great and may be very worthwhile, but again it’s pretty small potatoes in the industry:
Nvidia’s profits last year were 10× the upper bound of what the potential Canadian fab is promoting as a possible 10-year gross impact on GDP.
The industry is a point where capital demands and returns are so extreme that GPU access is having to be sold like bonds to the biggest asset managers on the planet just to feed the CAPEX to meet demand, and that’s a big part of why the frontier is a two-horse race that we’re not in. Again, that’s not to say Canadian businesses can’t make money during all this, but the players that are going to dominate this space are not coming from Canada.
An old adage in the industry says that the first GPU out of the fab plant costs $1 billion dollars but the nest one costs $5. The actual making of the chip itself, since so many of them can be made from the same die, is not that costly. The total cost of designing and developing the GPU is enormous and a significant chunk of the profit from each GPU sold goes to developing the net model. Although Nvidia’s balance sheet looks good, it is only as good as the next iteration.