Steve Bleiberg
Welcome to the Actively Speaking podcast, where TDP portfolio managers and their expert guests tackle current topics concerning capital markets and portfolio management. Join us for a fresh and insightful discussion from the perspective of an active manager.
Justin Howell
Good afternoon. Welcome to the TD Epoch Actively Speaking podcast. I'm Justin Houle, portfolio manager on the US Fundamental Value Strategies. And today I'm joined by Victor Anthony who is our technology analyst. And we're going to be discussing the basics of artificial intelligence and some of the trends within the industry. Good afternoon. Victor.
Victor Anthony
How are you? Good afternoon. I'm great. Thank you.
Justin Howell
You've been a technology analyst on Wall Street for 20 plus years. So let's start at the beginning. Why don't you talk about the ChatGPT launch? Why was the technology perceived as a major breakthrough? How did the early adoption compared to the internet and other popular tech applications? So how does this compare to other tech trends that you've seen?
Victor Anthony
Yeah, sure. So first, AI is not new. The technology has been around for several decades, led by companies such as IBM. It was an early pioneer in the field. And you have Google, who made the acquisition of DeepMind, a London based company. Microsoft has been working on what I could decades to establish an AI unit in 2013. Amazon has been focused on deploying AI both as retail and cloud services units, for quite some time now.
And then you have the academic institutions. Dartmouth in particular, has been at the forefront of AI, behind the scenes dating back since the 1950s. So it's not Newton new technologies. But before ChatGPT, most of the advance, I was often siloed in enterprises and, and academic labs like Dartmouth. What happened is OpenAI, in November 2022, unveiled to the world, ChatGPT embodied to the forefront that technology was perceived as a major breakthrough, not just because of the technical complexity, also because you're able to, compose text, brainstorm, answer complex questions.
And, you delivered, human like and natural language generation in it, easy to use and consumer interface. And so now you talked about it in terms of the timeline to adoption. It's probably one of the fastest technologies, adoption wise, that we've seen over the past hundred years. And if you compare it to, Facebook, it took Facebook about 4.5 years to meet to reach 100 users.
Instagram over 2.5 years. It took TikTok about two years. It took ChatGPT two months to cross that that that threshold. So in that sense, it took the public embrace of gen AI as measured by adoption rates, significantly less time to get interaction than the internet or personal computers. Now, the second part of your question, in terms of, compared to other major tech trends, you know, you know, when I look at the dotcom era in the 90s, what's different now?
With, the AI supercycle, when I look at the 1990s and dotcom, that era was really about, the new connectivity layer, which is internet, and you build on new, distribution models like e-commerce and, web portals and advertising was built on top of that. But the bottleneck that was really bandwidth and user habits had to change significantly.
Then you go over to the mobile era in 2007, plus, that was about establishing new access points and form factors, which is a smartphone. The bottleneck that was really just often hardware speed and OS maturity. Now, the AI supercycle, which, ushered in by the looked at ChatGPT in 2022. So you're looking beyond that. And that's about building a new foundational utility intelligence and compute, with the eventual goal of achieving AGI, which is artificial general intelligence.
And the unprecedented part of this whole cycle is that you have a massive CapEx build out before any apps or any sort of utilities. Could be built on top of that.
Justin Howell
Since the launch of ChatGPT, how have the other technology companies responded from both a competitive and capital expenditure perspective?
Victor Anthony
Yeah. So following ChatGPT is launch, other technology companies, such as the Magnificent Seven responded aggressively, both from a competitive and capital expenditure perspective. And so the major players launched rival large language models almost immediately and integrated them, integrated them into the product ecosystems. So, for example, Google, they from what we understand, the declared code red within the organization and really looked to launch, Lem rather quickly.
They started with by changing names to Gemini, Gemini to compete directly with ChatGPT. My personal view, I think Gemini has actually caught up in terms of functionality, whether it be video, photo editing and just, responses to to text queries and Gemini just launching, a new updated version, Gemini three and basically saying there's really, they've had some major breakthroughs in mathematics and major breakthroughs in, in other areas.
And in particularly they pointed out that, the, it's almost a PhD like responses that they get from, Gemini three and vastly superior to the previous model of 2.5. Then you have meta, which launched an open source, Monocle Llama that was a, material breakthrough, breakthrough for them. But they've seen some somewhat stumble with the rollout of llama four.
And so the kind of back to the drawing board in terms of what they're looking to do with the large language models, Tesla, eventually launched and I that's the large language model. But Amazon and Microsoft are taking different approaches. You know, they've, you know, the two decided to keep their LMS proprietary within their organization in tandem instead of funding it out to, to the user base and to Amazon to their bedrock product.
They make other LMS like ChatGPT and, Gemini available to that, to their user base. And Microsoft is doing the same as well as well as integrating, you know, outside elements into the current, copilot, product base, Oracle late to the cycle there, but they're looking to make inroads. And then you have a palantir, which is really just using AI to win government contracts and not really farming out any sort of, sort of large language models of the outside world.
And in the private side, we have anthropic and protect city was coming up and trying to compete in the space as well. And on the CapEx side, you mentioned, I think the most pronounced response has been the massive spike in CapEx by the hyperscalers, which is, alphabet, Amazon, meta, Microsoft, as well as Oracle. They're actually spending close to $400 billion in CapEx this year.
That's up almost 60% year over year. And that was on top of this 60% year over year growth rate on CapEx spending 2024, estimate and consensus as well. That will add another probably 100 billion. So the 500,000,000,000 in 2026, probably that goes to 600,000,000,000 in 2027 in terms of, CapEx spent on the hyperscalers. But a growth rate, decelerates quite meaningfully.
So you're looking at roughly instead of 60% growth rates, we're looking at mid-teens growth rates over the next two years.
Justin Howell
So drilling down a little bit more on that conversation about capital expenditures and subject I'd be interested in, and I haven't heard a lot about, can you please walk us through the cost structure and timeline of building a data center and where all this CapEx is going?
Victor Anthony
So building, data center will take anywhere from two and a half to three years to build out. And, it's a multibillion dollar endeavor. Nvidia has cited, roughly a one gigawatt data center with cost anywhere from 50 to $60 billion in total to construct. I've seen other estimates going to South Side analysts, anywhere from 35 to, to 40 billion.
But it's a big number. The hardware component for most estimates. Took a look at the Bob estimate there, actually said the compute component of that is roughly 80 to 85% of the mix. AMD sees kind of a smaller mix, 75 to 80% of the mix. But the compute and hardware component is probably the biggest piece of that, massive, data center spend that we're seeing on the hyperscalers.
Next is networking and storage. That's another 10 to 15% of the mix. Then we have power infrastructure that's 5 to 10%. And the balance of that is really just a land provisioning and building provision. And that's around 5 to 10% of the mix.
Justin Howell
So how does, energy or power play into all of this? Am I infrastructure build out? And in your opinion, will that end up being a better, bigger bottleneck than actually securing the compute?
Victor Anthony
Yeah. So in the early phase of the AI super cycle, compute specifically the supply of high end GPUs like the A100 coming out of Nvidia was the primary bottleneck. There's several times that Nvidia was unable to deliver those chips to the to the hyperscalers and and other customers like OpenAI. So as the ecosystem sees more competition for chips, you know, AMD is coming out with, competitive chips, power right now is rapidly becoming more of the bottleneck by most estimates.
Today, data centers consume 1 to 2% of global electricity, and that percent is growing exponentially by the day. A modern AI data center consume 10 to 100 times the power of a legacy cloud data center requirement hundreds of megawatts, and securing that much new electrical capacity from the grid requires years of planning. New substations and transmission lines build out often hindered regulatory and physical capacity constraints, and limits faster than hardware can be delivered.
And so for new bills, the living effect is shifting from how many GPUs Nvidia and AMD can ship to how many megawatts the local utility can deliver. This dynamic favors utilities, power infrastructure companies and those investing in energy efficient cooling and next generation and power management. So, in short, the answer power is the biggest bottleneck going forward.
Justin Howell
Great. So let's move on. What who are the major players on the infrastructure side and what are their competitive positions? And more specifically, what puts Nvidia in such a great position and who is in the sole position to compete with Nvidia?
Victor Anthony
Yeah. So on the chip and compute side we have Nvidia. We have AMD, we have Broadcom with A6. We have Google with TPUs at Amazon with Trainium. Yeah Marvell. And 1 or 2 Taiwanese players on a network inside two key players. Arista Cisco Nvidia server side we have SanDisk, we have Western Digital on a cloud server side you have Amazon with AWS.
You have Microsoft with Azure. Yeah, alphabet, Google Cloud and also Oracle and IBM. So what puts Nvidia in a great position that Nvidia is in right now. Largest market cap company in the world. Now number one they have hardware dominance. And that's what the Cuda parallel computing platform that they've developed over the past two decades. Second is a software moat with the cutest software stack.
Third is supply control. Nvidia has privatized secure and advanced foundry capacity with TSMC in particular, and that gives it a tight grip on the supply of the most in-demand chips. And, lastly, they really have a strong strategic focus, which is Nvidia's roadmap of, chips that graze the black walls as well as Ruben. They make sure those are fully aligned with the massive scale, networking and power demands of the next generation LMS.
Now, second to that is AMD, which is playing catch up for the mi 5G series, which is planned for the middle of 2026. It will be the first time that AMD has implements a scaled up architecture, similar to what we see with Nvidia. And, AMD expects to mi 450 chip solutions to be at least on par with the Nvidia's open platform for both training and inference workloads.
So despite the fact that AMD has been playing catch up, you know, I think and most in the industry are thinking that they'll come close to matching the solutions that Nvidia has on the, on the market today.
Justin Howell
Is there any way you can quantify how big of a theme AI has been in the stock market, and how much AI related companies represent other major U.S. stock market indices today?
Victor Anthony
Yeah, sure. In 2025 AI related sectors, that's information, information technology, communication services, and to a lesser extent, industrials and utilities have accounted for near 75% of the total return on the S&P 500 and the corresponding year, 80% of the EPS accretion. And eight of the ten largest companies in the S&P 500 today. AI related to exceptions are Berkshire Hathaway and JPMorgan and those eight, approximate 38% of the S&P 500 today.
Simply put, AI stocks are driving the stock market returns. All year long.
Justin Howell
So we already talked about the AI infrastructure players, but maybe we could talk a little bit about the companies that are developing the consumer business applications. What are their competitive positions? And specifically how is this artificial intelligence technology impacting the Mag seven companies?
Victor Anthony
The incumbent productivity search side of this, you do have Microsoft. You have alphabet. You know, meta as well. Yeah. Massive advantages in terms of distribution and integrated data modes. So Microsoft has 365. Google has workspace search social platforms. They integrated AI into existing services to drive of modernization. That gives them sort of a competitive edge. Pure play models like OpenAI anthropic perplexity, these guys are focused on building out the best foundation models and user facing agents and the need for strategic partnerships, but for them to build out their infrastructure, for them to achieve their goals over time.
And then you have the vertical, the enterprise, players such as Salesforce, ServiceNow, Adobe, Palantir, you know, those are focused on embedding AI, genetic AI in particular across the organizations. And, you know, providing the customers of, agentless AI solutions over the time frame. So, but the AI supercycle is causing a lot of, I think, re acceleration in, a replatforming of the core Max seven businesses.
So Microsoft alphabet transforming the cloud businesses, Azure GCP from general purpose utility AI centric platforms that command in higher margin AI services revenue Copilot subscriptions. This drives higher, more defensible revenue. AWS using its market leading cloud position to offer competitive foundational models and specialized chips. AI is the next battleground for the cloud market. Share. I think I I better using AI to dramatically improve content ranking and advertising targeting.
And now to build new hardware products Ray-Ban sunglasses, Nvidia and Tesla. You know, everyone knows what they do in terms of AI. You know, Tesla in particular is really just focusing AI outside of the AI, really robotics and, the autonomous driving, businesses. And so across all these different businesses, you have, you know, different moats, different competitive advantages, but they all use an AI across their core businesses.
And in terms of new new setups, in terms of the cloud, to offer compute capacity to the customers.
Justin Howell
Turning back towards the AI CapEx, where do you think that can eventually go over the next several years? What ending are we in this build out of the AI infrastructure? And then the question I have more specifically is, if you look at it from the time of the Netscape IPO to the peak of the stock market or the internet bubble in March 2000, it was about five years.
Do you think that's a reasonable timeline for this AI CapEx cycle and stock market cycle as well?
Victor Anthony
So I mentioned earlier that, you know, this year, you know, the hyperscalers will spend close to 400 billion. That goes to 500,000,000,000 in 2026. And, probably goes to another 600,000,000,000 in 2000 and, 27 and continues on. I just think this, cycle just continues on for now of, multiple different years. Nvidia, in particular, has called out roughly 3 to $4 trillion of US spend over the next decade.
And incidentally, TSMC's TSMC has actually agrees with them and backs that number, and generally noted recently that the AI server CapEx outlook of mid 40% kegger is actually low, given the exponential growth in token usage. And so we should actually believe that, you know, that number, that 3 to $4 trillion number over the next several years? You know, Jensen in particular thinks we're in in three of a ten year buildout cycle.
I tend to agree with him. You know, when I take a look at where that spend is been allocated. So you have, the sovereign build out. That's about $1.5 trillion of spend. And that's the US, European countries, Middle Eastern countries. If in some African countries are looking to build out their own sort of AI factories, then you look at, there's about another $1 trillion of existing data centers that need to be retrofitted for accelerated compute.
Then you have the hyperscalers themselves. You know, we've walked through the numbers looking to spend over $1 trillion over the next five years. And then you have even enterprises are looking to build out their own sort of mini AI factories. A lot of telecom companies, a lot of, consumer based companies, a lot of automobile manufacturers looking to build out their own sort of AI factories.
And so, you know, I think that takes us into into the next decade. So I would peg it around in three and four. I do think we may have digestion periods over time. You know, I was talking to the CEO of AMD and Isuzu in a Saturday conversation after their analyst day, and actually whether or not do you think this is a bubble?
And her response was that definitely thinks that she actually thinks investors as well as, as the media, a more short term focus, you know, when she talks to customers, they're talking three, four, five, six, seven years out in terms of the build cycle of deposit, AMD units of supply to the customer base. So actually so she actually sees a very, very long tail, all these companies telling you that demand is greater than supply.
And so, there's a long tail to this. I think the build out is going to take us multiple years. But yes, I do think there's going to be digestion periods along the way where, you know, there may be some pause spending. Lisa Sue, she mentioned that the one caveat that she sees is if the economy dovetails into a recession and she definitely thinks, you know, the hyperscalers businesses, based on advertising, e-commerce, largely very economically, sensitive businesses, they will, ultimately be forced to cut back on the CapEx band temporarily.
And so we may see that, but but over time, I just think this is up until the right going forward. And in terms of your next comparison, you know, I do think history does rhyme. And I'll be the last person to tell you that this time is different. I won't tell you that, but I think the current cycle that we're in right now is characterized by several entrenched ecosystem players that are operating at scale.
There are strong network effects, huge user bases, strong balance sheets, the highly profitable, and you have strong free cash flow generation. And you know, I think that's that's materially in significantly different from what we saw years ago, you know comparison to your Netscape example. And so so effectively what you have is a better, more superior businesses that exist now versus years ago and past bubbles of past cycles.
Justin Howell
How do you think about the returns on invested capital for all of these capital expenditures? And since our investing philosophy is based on free cash flow, what do you think are the long term implications for that metric? At some of the key players we have discussed this afternoon?
Victor Anthony
So I do think the CapEx build is creating a defensive moat, an entirely new revenue streams and high margin revenue streams for a lot of these companies that to invest in against it. You know, Microsoft Copilot for example, by the Atlantic, I you know, if I can deliver significant productivity gains, against that spend for millions of customers, which they do have, I think, you know, these new services can be priced at a premium, leading to high returns on invested capital and, on the AI infrastructure built right now.
I was kind of seeing some of that right now in terms of meta, I just look at the core business. They lay it on AI across all of the different social media platforms that meta owns, and that has benefited them significantly. This year. Advertising Rpu is on pace to accelerate to 16% year over year in 2025, and that's versus a 4% uptick in 2024.
This is one example of, you know, I then you do have the hyperscalers who are generally returned from from listen to capacity. And so so there's revenue of revenues being generated right now. At the end of the day, if they're able to generate cost efficiencies and remove significant costs from the business models, and we start to see examples of that right now, you know, Amazon in particular, make a massive reduction in force.
Other companies have reduced have met has reduced headcount significantly this year as well. And so you're going to get the efficiencies laid on plus high margin revenues. Then you could get the, returns on invested capital that investors demand from that spend. Now, the variant view is if there's, you know, model commoditization and, there's a competitive need for these companies to keep up with their rivals, then, you know, you could probably see negative returns on invested capital.
So my view, it remains to be seen these companies have gone through different cycles in the past, and I've covered them over the past two decades. And, they've been proven to withstand and adapt to new technologies and make significant money, generate significant returns in the past. And I think they could do it again, but ultimately remains to be seen how this evolves and how the competitive landscape evolves.
Justin Howell
Thank you for that. And so my final question is what are the signs that you will watch for this becoming a bubble and or that the CapEx cycle is beginning to roll over?
Victor Anthony
Good question. Excellent question. I think what I've been looking for is commentary from Amazon, meta, Google and Microsoft opening stating that they plan to ease CapEx spending or spending growth is expected to decelerate meaningfully if they say that on the conference calls. That's one sorry, any sort of reversal of the current language on the earnings call.
Second, if you no longer hear demand is greater than supply on its earnings conference calls, that's another two. I'll be watching for a very, very closely. There's a spate of, IPOs coming to market. They're unprofitable to me. That's just you know, I've seen this before in the past two decades that that to me, ultimately, that's what probably winners probably say, that this is a bubble that needs to pop at some point.
And so those are the three things I probably would look for as evidence that that this is becoming a bubble. And that's beginning to roll over.
Justin Howell
Great. Thank you. Victor. That was very helpful. And obviously AI is, the topic de jour in the stock market today, and we could talk about it forever. But I think in the interest of time, we will leave it there. Thank you for coming on the podcast today.
Victor Anthony
Thank you.
Justin Howell
We'll see you again for conversation in the near future.
Steve Bleiberg
Remember to subscribe to Actively Speaking on Apple Podcasts, Amazon Podcasts, Spotify or Pandora. You can find all of our previous episodes and additional content on our website, www.tdgis.com.
The information contained herein is for information purposes only. The information has been drawn from sources believed to be reliable. Graphs and charts are used for illustrative purposes only and do not reflect future values or future performance of any investment. The information does not provide financial, legal, tax or investment advice. Particular investment, tax or trading strategies should be evaluated relative to each individual's objectives and risk tolerance.
This material is not an offer to any person in any jurisdiction where unlawful or unauthorized. These materials have not been reviewed by and are not registered with any securities or other regulatory authority. In jurisdictions where we operate. Any general discussion or opinions contained within these materials regarding securities or market conditions represent our view or the view over the source cited, unless otherwise indicated.
Such view is as of the date noted and is subject to change. Information about the portfolio holdings, asset allocation or diversification is historical and is subject to change. This document may contain forward looking statements or FLS. FLS reflect current expectations and projections about future events and or outcomes based on data currently available. Such expectations and projections may be incorrect in the future, as events which were not anticipated or considered in their formulation may occur and lead to results that differ materially from those expressed or implied.
FLS are not guarantees of future performance, and reliance on FLS should be avoided. No investment strategy or risk management technique can guarantee returns or eliminate risk in any market environment. Any characteristics, guidelines, constraints or other information provided for this material is representative of the investment strategy and is provided for illustrative purposes only. They may change at any time and may differ for a specific account.
Each client account is individually managed. Actual holdings will vary for each client, and there is no guarantee that a particular client's account will have the same characteristics as described herein. Any information about the holdings asset allocation or sector diversification is historical and is not an indication of future performance or any future portfolio composition, which will vary. Portfolio holdings are representative of the strategy, are subject to change at any time, and are not a recommendation to buy or sell a security.
The securities identified and described do not represent all of the securities purchased, sold or recommended for the portfolio. It should not be assumed that an investment in these securities or sectors was or will be profitable. The securities disclosed may or may not be a current investment in any strategy. Any reference to a specific company listed here in does not constitute a recommendation to buy, sell or hold securities of such company, nor does it constitute a recommendation to invest directly in any such company.
Any projections, targets or estimates in this presentation are forward looking statements and are based on our internal research, analysis and assumptions. There can be no assurances that such projections, targets or estimates will occur and the actual results may be materially different. Additional information about our assumptions is available upon request. Other events which were not taken into account in formulating such projections, targets or estimates may occur and may significantly affect the returns or performance.
