For the non-technical, all you need to know about AI CapEx spend is this:

Training: Expensive, short-lived CapEx cycles

Inference: Cost compounds as trained models get real-world use

When we say training is expensive, we mean it, especially as the models being trained continue to get bigger and more capable. Take Zuck’s commentary on the next series of Meta’s Llama models:

Napkin math:

A single H100 GPU from NVIDIA goes for roughly $30K. Assuming Meta has some negotiating leverage given its scale, let’s say they bring it down to $20K a pop - that’s >$2B of GPUs alone being used to train the Llama 4 series.

Okay, so we know training costs are huge, but do they need to be?

China’s DeepSeek is making investors question that, following the release of its v3 LLM and R1 reasoning model. The models showcase performance on par with OpenAI’s leading models for a fraction of the cost:

The company claims it did its final training run in 55 days for under $6M, peanuts in comparison to the investments being made by U.S. big tech. Though there is some nuance missing from the headlines we tackle in the full piece.

So how should you interpret this? Is the AI trade dead?

In our view, no, but it’s certainly changing. The framework we want to leave you with today is that the LLMs are a commodity, and efficiency gains are a good thing.

The best way to illustrate this is through an oil analogy:

If a new process emerges and enables more efficient oil production (ie. modern fracking techniques), supply that was previously uneconomic to extract comes online and value-added products come down in price while preserving margins, driving increased demand from end-markets and in turn, increased demand for oil.

AI is undergoing a similar shift right now.

DeepSeek’s more efficient training techniques make the use of LLMs more economic, so the cost of the value-added products (software) leveraging them should come down too, and entirely new products that weren’t economically feasible before will be today.

For this to hold true, we should expect a big jump in demand for inference (real-world use of models), and we’re already seeing signs of it:

So what’s next? We think capital will flow towards the application layer, to companies already generating value with LLMs that can now do so at a fraction of the cost.

If the above link won’t work, try this: https://www.bullpen.finance/content/23

Tomato-gate wasn’t on the bingo card for 2025 but here we are, as Kraft Heinz came out publicly defending itself following Trudeau’s talk about placing tariffs on American made ketchup:

Concerned citizens later shared pictures of Heinz bottles marked with “U.S.A.”, forcing Heinz to respond that 10% of all ketchup sold in Canada last year came from south of the border.

We’ll keep an eye on this one, as something tells us Big Ketchup isn’t giving us the full truth.

Aecon (ARE) announced it reached commercial close on a progressive design-build transit project worth >$2.8B to the company.

Foran Mining (FOM) announced a $41M funding agreement with Canada’s Strategic Innovation Fund to advance its McIlvenna Bay copper project in Saskatchewan.

AIMCO (Alberta’s public pension manager) announced a small round of layoffs

Flagging NFI Group in this one as Trump wants to freeze federal grants, threatening the growth of its EV bus segment. It has been blocked by a judge after hours, but could weigh on near-term sentiment until a resolution is reached.

Canadian companies are starting to report… finally! We had Metro Inc. yesterday morning who sold off ~3% after its print. Numbers were in-line and guidance was maintained, so investors were probably looking for a beat and raise.