Global Equity Investor insight | February 2025

Jeremy Richardson

Hello, this is Jeremy Richardson from the RBC Global Equity Team here with another update. Now, ordinarily when I sit down to do these little videos, it's a reprise of what's happened over the previous month, but I kind of feel at the moment there's just really one topic that we should be focusing on in particular, which is the news that DeepSeek, a Chinese AI startup, has unlocked a way of being able to deliver reasoning based models at a fraction of the cost and time than some of the original reasoning models that we were so excited by just a few months ago when they got released by the likes of ChatGPT and others.

So, the innovations that DeepSeek have been able to put out, really stem from two things. Firstly, being able to train a small model very quickly, it does this using a new methodology where an established large model essentially teaches a smaller model what it needs to know by putting in place the correct incentives. This is proving to be highly effective and requires a fraction of the cost. It's sort of like piggybacking, on some of the established work that has been going on within the industry.

The second innovation that they've been able to deliver is dramatically reducing the cost of using these models, what the technologists call inferencing. And they're doing this by, a ‘mixture of experts’ approach. And so for those of you who are familiar with the, ‘Who Wants to Be a Millionaire?’ program, it's a little bit like if you get a really tricky question and picking up the phone and let's say it's about sports, and phoning the person you know who's really hot on their Olympic history say, they're much more likely to know the answer to that question than you know, your friend who instead is really interested in science.

So knowing who to ask can get you to the right answer quickly. And the ‘mixture of experts’ approach does that. So, assuming you've got a big address book, instead of asking everybody in that address book, which is what the traditional inferencing approach would do, it identifies the right people to ask and asks them. Turns out it doesn't diminish the quality of the answer, but it actually saves you a lot less time dialling up friends who don't know the answer. So, you get equivalent quality of responses at a fraction of the cost. By the way, they've also been able to do this on what the technologists thought would be quite old fashioned technology, which in itself is a considerable achievement.

So where does this leave the industry? Where does this leave the outlook for technology companies and importantly, for investors?

Well, I think let's think about this in terms of the technology stack. So, the whole thing starts off with a number of large technology companies who are building large language models. So these are going on in the very big data centres and they will be costing billions of dollars, and people will need small nuclear power stations in order to just keep them going. Well, it seems really hard if you're going to spend billions of dollars, but that people can do this cheaper, that you'll be able to charge as much as you thought originally were with all the work that you've been doing.

So, some of these companies are going to require new business models in order to justify the amount of capital expenditure they're putting into data centres, etc. Then you've got the chip companies themselves, this is the likes of Nvidia and AMD and so on. You know, some of these continue to do marvellous things, but their customers are using their chips partly for training, where again, the economics of that is perhaps a little bit more in question and inferencing, which the costs of which are dramatically falling.

So maybe that changes the way in which you get to a revenue number. Revenue is volume times price. Maybe volume goes up dramatically, but price falls quite considerably. So how does that impact the financials of some of these chip companies? That's a big question for investors to debate at the moment.

Then you've got the cloud computing companies that use the likes of Microsoft, Alphabet and Amazon, who are actually running these data centres, doing all of this inferencing. Well, if inferencing becomes really cheap, and people are doing a lot more of it, those data centres are going to become incredibly busy. And that's probably quite good news for these guys, because it means they can charge more. And if you can use older technology, their cost should also be, leaner than people were originally thinking. So, balance that’s probably put that in the positive margin.

And then the last two constituencies of software companies and consumers, well, for software companies, some software companies may actually do really, really well out of this. If you're running artificial intelligence, the costs of artificial intelligence may have dramatically fallen as a result of this technological step. Which means that hopefully your profits will improve. But actually, if AI becomes super cheap to deploy and super cheap to use, then everybody's going to end up with that, in which case it becomes a platform technology.

The analogy here may be like, electricity, for example, or maybe air travel where, you know, yes, it's really snazzy and neat to be able to fly across the Atlantic in a matter of mere hours, but the industry has been built on actually efficiency and actually lowering the average cost per mile per seat flown. And so maybe that's the model that we need to think about here, who has got a competitive advantage in their software model that enables them to be able to hang on to the efficiency savings that artificial intelligence offers and allow that to translate that through to its stakeholders, including investors.

And then the final community is consumers. And these are the people who are really going to win from this, because I think for most of us, AI is something that we dabble with. Maybe we use it in a work setting, maybe we might have a subscription. But it's probably not the sort of thing that we use 24/7 and partly the reason for that is because of the cost. It gets throttled, based on the number of tokens we want to use or maybe we have to get charged for a monthly subscription, say, but if actually AI becomes omnipotent, universally found, then actually, consumers would be able to have it 24/7 in their pocket on their smartphones.

In fact, we might even be able to run it natively on our computers or even on old fashioned phones, and not even have to be connected to the internet all the time. Just imagine what sort of a wave of innovation that could get, at least from that type of a situation. It's a little bit like, you know, moving from dial up modems, back in sort of internet version 1.0, to actually sort of the world of the smartphone, you know, who would have thought back then when we were sort of plugging in our computers at the end of the day to catch up on email, that we could have, that it would lead to things like Airbnb and Uber as business models.

So, the future is bright. There's a lot of opportunity as a result of this big technological step. The question I think for investors, is where within that value stack, most of the value is going to reside. And so, as we've been thinking about allocations within the industry and within our portfolio, one of the things that you've seen us do over the last few months and quarters is actually try to gradually begin to transfer capital, away up that value stack towards the some of the business models that we, consider are actually going to end up as being net beneficiaries.

And I hope that's been of interest to you. And I look forward to catching up with you again soon.