AI boom continues despite market volatility

Text transcript

Welcome to Advisor to Go, brought to you by CIBC Asset Management, a podcast bringing advisors the latest financial insights and developments from our subject-matter experts themselves. 

Robertson Velez, portfolio manager for the CIBC Global Technology Fund with CIBC 

Let me give you an update on the tech sector. 

So, 2024 was a fantastic year for technology. I started the year very optimistic about the growth of AI into 2025, and I still feel very resolutely that that view is correct. But to say that 2025 year-to-date has been volatile in tech would be an understatement.  

One event that is specific to the technology sector, that has had a profound impact on the industry, is the introduction of DeepSeek in January, which I’d like to talk about in some detail because it has been so impactful. 

Let me set the context for DeepSeek. As you all probably have heard, DeepSeek is a Chinese AI lab company. They released DeepSeek-R1 in late January, with performance that was comparable to the latest models from AI leaders such as Open AI and Meta. But the key point was that DeepSeek claims to have trained this model at a significantly lower cost. This news tanked the AI market on fears of disruption. And there’s been an ongoing debate about what this means for the industry. 

So, the crux of the concern around DeepSeek is the claim that they can be trained for 10 times less than their competitors. I’m not disputing the performance claims or the training costs claims, but I think there’s a lot of understanding about the comparability of these figures. DeepSeek’s claim is that they trained their model for $5.5 million, whereas the training cost for OpenAI’s ChatGPT 4.O was estimated at $60 million — a 10 times improvement. 

However it’s unclear what is or is not included in this training cost. The 5.5 million figure is the GPU cost of the pre-training run, which is only a portion of the total cost of the model. And it’s widely believed that DeepSeek has about 50,000 GPUs that they’ve acquired at a cost of 1.6 billion. So, its hardware investment is similar to other AI peers. 

In addition, DeepSeek-R1 uses other learnings from other AI models that have come before. It is widely believed to have used distillation techniques to train DeepSeek model, based on Meta’s Llama and OpenAI’s ChatGPT models. So, nothing necessarily wrong with these techniques, but if they’ve leveraged training from other models, shouldn’t those costs be included in its training costs? 

So, then, the next argument is that, regardless of the true costs of DeepSeek, DeepSeek has demonstrated that the incremental cost of training the model is significantly less, by a factor of 10 times, compared to Open AI and others. So, again, this is misleading. The DeepSeek-R1 model was released in January 2025, whereas the open AI ChatGPT 4.O-model was released in May 2024, which is close to a year different. Estimates put algorithmic progress at about four times per year. So if we were to compare the models trained in the same timeframe, the difference would be closer to two to three times, not 10 times. 

This is not to take away from the progress that DeepSeek has achieved. But the performance improvements versus current compatible models are much more modest than what the market seems to believe. 

Putting aside the question of whether DeepSeek is really that much better, the cost of training really has come down significantly for the whole industry. And the argument now is that this will bring down CapEx spend from the hyperscalers like Microsoft, Meta and Alphabet, which are spending 10s of billions in CapEx on AI. 

My view is that the hyperscalers are not spending this CapEx to reach a specific AI goal. They’re spending this much because they want to maintain AI leadership and don’t want to fall behind their peers. So, if the industry achieves improvements in efficiency — which is a normal course event — all the hyperscalers will incorporate these improvements without reducing CapEx spend. And recent commentary from all these hyperscalers, in their earnings calls, support this view. They’re increasing their CapEx spend, not reducing it. 

So, who are the beneficiaries of these AI efficiencies? The stock market reaction to DeepSeek implies that the market believes efficiencies in AI will lead to less demand for AI, because what we can achieve with less cost will be good enough. I completely disagree with this view. And I think history backs me up.  

There’s an economic concept called Jevons Paradox, which states that efficiency gains lead to increased consumption, not less.  

So, in the 70s, for example, when the cost of fuel came down with the introduction of fuel-efficient vehicles, people drove more, not less. In AI, efficiencies that reduce the cost of creating an inference will accelerate the rate of adoption of AI, not reduce it. 

In my view, the original thesis of AI adoption remains completely intact. And I continue to add to positions in AI [with] infrastructure names like Nvidia and Broadcom. And I hold the AI cloud providers, Microsoft and Google. And I’m positioned for software names that will benefit directly from the adoption of AI, such as ServiceNow, Salesforce, etc. And I would estimate that about 80% of the technology holdings in the fund is exposed to this AI theme. 

* * * 

My outlook for 2025 is positive for technology. And, as I’ve laid out [in] my AI thesis, I think we are still at the early stages of AI adoption, which I think will be transformative for the industry. 

So, let me provide a very broad context for how I look at technology. In 1965, Gordon Moore, the co-founder of Intel, made the observation that the number of transistors that would fit on an integrated circuit doubles every one to two years. That prediction has held true for the last six decades or so. Why is this important? Because every major inflection point in technology has been driven by Moore’s law, from personal computers, to the Internet, to the smartphone. Technological capability leads to new applications,not the other way around. 

Over the last few years, the pace of Moore’s law has slowed down. Even though we’re still moving down to smaller process nodes, the cost is much higher, making it much longer to double performance for the same cost. And this is where accelerated computing comes in. 

Accelerated computing is a type of processing that allows for calculations and processes to run concurrently, rather than successfully, resulting in greater efficiency and faster calculations. So it optimizes every layer of computational processing, allowing for the most impactful opportunities, such as generative AI, machine learning, drug discovery, robotics, climate modeling, etc. 

Accelerated computing is why Nvidia, not Intel, is so important in this new industry. The performance of Nvidia’s parallel computing systems has increased 1000 times in eight years and has replaced Moore’s law as the driver of the new AI industry. 

And how big is this opportunity? Nvidia estimates that the long-term available market opportunity could be one trillion, split between data centre systems, autonomous vehicles industrial applications. I think this estimate is conservative. So the first take away is this: that we have to size the opportunity right. AI is a huge potential market and we’re still in early stages of the adoption cycle. So I’m very optimistic that we will continue this trend for 2025. 

* * * 

So, what other areas in tech provide opportunities? 

I think large-cap tech names will continue to outperform. And there are new opportunities coming up as AI adoption broadens out. 

One area of focus for us is in enterprise software names. So, for example, ServiceNow, Salesforce, and many others. As infrastructure is built up, the main beneficiaries of AI will be companies that have access to enterprise data, and that can use AI to turn that data into useful actions. The biggest challenge that enterprises have is migrating their data into a structure that is useful for AI. We believe that the market for these data-migration tools might actually be bigger than the AI opportunity itself. 

So, for example, ServiceNow, which is a workflow management software for large enterprises, helps provide access to enterprise data that can be turned into useful insights or action. Salesforce offers customer relationship management software that not only provides access to enterprise data to drive insights and actions, it offers new products. For example, agentic AI where AI can actually take over the word processing. 

* * * 

I see growth and value as two sides of the same coin. We’re looking for companies that are trading at prices less than their worth, taking all factors including competitive advantages and growth rates, into account. 

So I think the idea of switching between growth and value is not completely the right framework. 

For example, one of the biggest myths in tech is that the gains are spread out fairly among all the players, which could not be further from the truth. In the PC era, 80% of industry operating profits went to Intel and Microsoft. In the smartphone era, 80% of industry operating profits went to Apple. So don’t bet against the leaders in the industry just because they’re big, since large companies usually have more sustainable competitive advantages, and this is reflected in their higher valuation. 

The big tend to get bigger. And there may be a time where we scale back positions in some of these names because the competitive advantages no longer seen as compelling. But we do that when the facts change, not because the stock has climbed to great heights. In tech, we usually are working at great heights. Instead, what we do is we ask ourselves what assumptions are embedded in the current prices of stocks today that may not be true tomorrow. 

I’ll give you an example. Among broad-based semiconductor stocks, such as Texas Instruments, we’re seeing one of the longest inventory corrections on record, partly because of the inventory hoarding behaviour post-pandemic as industrial and automotive customers dealt with semiconductor shortages, and partly because of uncertainty in the current environment due to the new U.S. administration potential new tariff policies. But in our analysis, we have to ask ourselves, how likely are these conditions to persist indefinitely? Inventory has to be drawn down eventually. And whatever tariff policies are ultimately in place, there will be less uncertainty. So production can resume at whatever levels make sense, which is likely higher than current levels. So, we have many examples of such opportunities where we make investments in companies based on expectations of more favourable conditions. 

We don’t buy companies, necessarily, because they’re cheap. But rather, we buy them because we think that, going forward, the expectations and the conditions will change such that they will not be as cheap. 

Part of our strategy — which I think is also our main competitive advantage — is to invest in a few names that we understand very well. We try to recognize the things that we don’t know or can’t know. And within that concentrated portfolio, we diversify into many areas within the sector that is predicated on different themes. 

So for example, in this talk, we’ve talked about investing in AI infrastructure, the cloud providers, the AI software vendors, broad-based semis — all are themes in which we believe we have a competitive advantage in understanding the businesses that will drive outperformance.

**

This program is intended for Advisor Use Only. The views expressed in this material are the views of CIBC Asset Management Inc., as of the date of publication unless otherwise indicated, and are subject to change at any time. CIBC Asset Management Inc. does not undertake any obligation or responsibility to update such opinions. This material is provided for general informational purposes only and does not constitute financial, investment, tax, legal or accounting advice, it should not be relied upon in that regard or be considered predictive of any future market performance, nor does it constitute an offer or solicitation to buy or sell any securities referred to. Individual circumstances and current events are critical to sound investment planning; anyone wishing to act on this material should consult with their advisor. Forward-looking statements include statements that are predictive in nature, that depend upon or refer to future events or conditions, or that include words such as “expects”, “anticipates”, “intends”, “plans”, “believes”, “estimates”, or other similar wording. In addition, any statements that may be made concerning future performance, strategies, or prospects and possible future actions taken by the fund, are also forward-looking statements. Forward-looking statements are not guarantees of future performance. These statements involve known and unknown risks, uncertainties, and other factors that may cause the actual results and achievements of the fund to differ materially from those expressed or implied by such statements. Such factors include, but are not limited to: general economic, market, and business conditions; fluctuations in securities prices, interest rates, and foreign currency exchange rates; changes in government regulations; and catastrophic events. The above list of important factors that may affect future results is not exhaustive. Before making any investment decisions, we encourage you to consider these and other factors carefully. CIBC Asset Management Inc. does not undertake, and specifically disclaims, any obligation to update or revise any forward-looking statements, whether as a result of new information, future developments, or otherwise prior to the release of the next management report of fund performance. Past performance may not be repeated and is not indicative of future results. The material and/or its contents may not be reproduced without the express written consent of CIBC Asset Management Inc. ^® The CIBC logo and “CIBC Asset Management” are registered trademarks of CIBC, used under license.