Has the AI boom got ahead of itself?

Jeremy: As 2026 draws to a close, what have we learnt about artificial intelligence's potential, its costs, and the businesses being built around it? When we last spoke to today's guests, they argued that artificial intelligence could drive decades of growth. They also pointed to a world in which digital employees would start taking on more of our work.

Now it's time to revisit those expectations. Which developments this year have strengthened their conviction, and which have given them reason to reconsider?

How do they distinguish a promising technology from a business that can turn that promise into lasting profits? With Americans just having voted in the US midterms, we'll ask what the political backdrop could mean for the AI investment boom, whether we're in a bubble, and what to watch out for in 2027 and the next five years.

Hello, I'm Jeremy Maggs. You're listening to No Ordinary Wednesday, and joining me from London are Investec Global Equity analysts David Smith and Zane Bezuidenhout.nGentlemen, welcome back to the show.

David, first question to you then, and if you had to choose one event or development that defined technology in 2026, what might it be, and why?

David: Jeremy, Zane and I were debating this before we came on your podcast, and I'm probably going to have to stick with 2026 was the year when useful intelligence became cheap enough that we could use it almost ubiquitously.

It is the year where it became cheap enough to get agents to start doing things for consumers.

So we've had a couple of agents out, Meta Muse, there's another one called Instinct, which I use. And while these companies still aren't making money off it, they are able to actually afford to have agents that will start doing tasks for you, like booking restaurants and filling out diary requests, which is pretty cool for me because it means I can finally actually be decent at admin.

Jeremy: So Zane, let me reverse engineer that question then. Was there something that took place this year then that didn't get as much recognition and maybe should have?

Zane: Yes, thank you, Jeremy. As Dave mentioned, we were chatting, and I think maybe it was, I think, around NVIDIA's big technology event in Taipei in Taiwan, and they announced a partnership with Microsoft around AI laptops, basically.

And their competitors, AMD, have announced a similar type of product in terms of their personal computing chip Ryzen capabilities. But basically, this means that consumers like you and I can run maybe last year's frontier models on our laptops.

And I think this is maybe an important development in that it now moves AI capabilities away from data centres and the cloud closer to us, the end users, and that is reinforcing of the adoption and proliferation of what Dave mentioned around being agentic AI.

And it has consequences for, and I'm sure we'll discuss it, the next step being physical AI, whether that's robotic autonomous vehicles

Jeremy: All right, David, when we spoke in July, if memory serves, you highlighted robotics as a longer term opportunity.

A couple of months on, what specific commercial milestone or milestones would persuade you that things are moving beyond promising demonstrations into something that is a little more scalable?

David: Jeremy, I actually think we're there. Now, we're still very early days, and that context is important. And when we think about robotics, there's probably three key buckets.

The bucket one would be autonomous vehicles. Think about the full self-drive that Tesla has. That's sort of grown three to four times in four years.

At the current trajectory, I would be very surprised if most people who buy a new car in five years don't have access to full self-drive in more than half of vehicles. So that's happening already. The second bucket is what you sort of classify as enterprise robotics. Think about a really fancy manufacturing site or a distribution centre where you'd have robots that would be doing the vast majority of the work.

Musk famously has all these pictures and videos of what happens in his factories, but you have the likes of Symbotic in the US, who fully automates or close to fully automates warehouses for the likes of Walmart, and that's growing really nicely. it's up sort of 10 times on revenue in the space of four or five years, and we think the traction of that's going to continue to accelerate.

The last one is one that most people think of, though, and that's humanoid robots. So I'm guessing most people listening to the podcast have seen context or videos where you see robots having robot Olympics or being in people's houses where you'd be vacuuming, and that's starting to happen.

The problem is that the complexity of the fringe cases, think about how when you pick up a glass on a table to clean underneath it, the pressure you use on that glass could break it.

So there's a range of software that needs to be happening in the background, so these robots know when to do things. There are actually interestingly companies right now in San Francisco that'll come to your house and clean it for $30 an hour with a robot. It's very early, but the problem is they have to have somebody in the background actually controlling the robot in certain instances.

So I suspect that's probably at least five years out, but I'd be very surprised if we have a chat in five years' time where you don't have at least one friend that has a fully autonomous robot in their house.

Jeremy: All right, David, Alphabet suffered a sharp loss in market value following its latest results, despite strong growth in Google Cloud.

Let me ask you this. What worried investors more? Was it the cost of its AI ambitions or the threat that AI poses to the search advertising business that helps fund them?

David: Jeremy, you're talking to a tech investor, so the last week or two has been particularly painful. I feel like I've got a couple of scars on my back from the gyrations in the market, so it definitely was Alphabet, but it's happened to more than Alphabet.

It was one of the worst weeks for tech that we've had this year. But Alphabet in particular had two issues. One is that the Google search business, the one that you traditionally type into a browser, that had one of the slowest growth rates we've seen in a while. It decelerated from 16% the prior quarter to 12%, and it suddenly got investors very nervous that AI is going to eat into their search business.

You won't have to go into Google anymore to ask questions because you can just ask your AI to do it instead. The other bit is they are spending a lot of money on their cloud business. That grew amazingly. It grew north of 100% year on year. But people don't tend to like to... When they're a little bit nervous of the market, they don't like to invest in businesses that are free cash flow negative, and they're probably going to remain that way for at least a year or so

Jeremy: So what would you need to see then in the next set of results to judge whether AI is strengthening Google's business or eroding its earnings power?

David: So Zane and I spend a lot of time trying to answer that exact question. We are fairly comfortable that Google's going to be one of the major winners in AI in the next five years. They've got this full tech stack approach where they have their own chips, they have their own cloud, they have whole bunch of users they can drive it to, they've got the models.

They're basically one of the only full stack approaches towards AI, and we think that's coming through in a number of ways. So what's really interesting as you talked, well, I spoke about the Google Search revenue growth, but what management highlights on their recent call is they actually had the fastest growth in searches they've had in more than five years.

A lot of those searches were in AI mode. So we are firm believers that they're going to be a winner, but it's just going to take a little time before they pull the monetisation lever.

Jeremy: And David, Americans voting in the midterms, or they've just voted, what would you be watching as the outcome consequences become clearer, and how could the balance of power in Congress then affect AI investment, maybe through spending decisions or regulation or even borrowing costs?

David: Jeremy, so I think the number one concern that tech investors have is that you have some sort of wobble in terms of supply of data centres coming online. So that would have a ripple effect throughout the supply chain, the chip makers, the model makers, if you just can't produce enough units of intelligence or tokens because you just don't have enough data centres coming online.

So the worry always is that you have some sort of regulatory overview from whichever body in whichever country that slows down the rollout of data centres. So that's probably front of mind for everyone. It tends not to be at state level, at a higher level. In the US, it tends to be very specific to states or counties, rather.

So it would be a case of do you see a change in politicians in charge slowing down the rollout of data centres and making it more difficult?

Jeremy: So Zane, just a quick follow-up then to what David has been saying, how much difference can Washington make to an investment boom being driven by private companies?

Zane: I think not just specific to Washington, but any government, I think regulation is necessary to some degree, especially for nascent industries, and particularly one of this nature, where we're dealing in the currency of intelligence. the impacts can potentially blunten or dampen short-term growth.

But if done correctly, and that's caveated with a big if done correctly, then they could make the growth, although blunted in the short term, more durable and ultimately drive long-term adoption to a technology. I always like to use the example when motor vehicles were invented, rules of the road weren't invented coincidentally that had to be regulated for everyone's safety, and- When it is safe to use, that drives further adoption. So if done correctly, regulation could actually be an enabler of adoption and proliferation of the technology going forward.

Jeremy: Dave, you and Zane previously highlighted resistance to data centres and the prospect of higher electricity bills.

So here's my question: how should the cost of expanding local infrastructure be shared between tech companies and communities?

David: There's a massive drive of NIMBYism right now. For those who don't know the acronym, NIMBYism stands for not in my backyard. You're spot on. There's a resistance to it, and there are a number of reasons for it.

One, if you live very close to a data centre, they tend to have huge amounts of generators in the background, and those generators make a lot of noise. So there is no doubt there's some of that, but there's also a lot of concern around the usage of water or electricity. The data shows it's actually not as bad as often feared, and often happens, it's blown up in the press.

But to get round those issues, what I think should absolutely be happening is that you have direct investment from the data centre companies into the communities to help with infrastructure and services around the area. But there's actually some really cool data around things like if you use the data centre to balance the grid in a regional area, you can actually cut electricity prices by 10%.

It does, however, require coordination with the state bodies, and that's often the trick. But to get round these issues where people do push back on data centres in areas, I think it's almost going to become a prerequisite.

Jeremy: So Zane, let me ask you this then. What should companies be expected to demonstrate about their water use before expanding into water-stressed areas?

Zane: Been quite a little bit maybe of, misinformation around this. Yes, data centres do use water, but a lot of that is recycled within their systems, so they're not big consumers of net new water resources. Yes, new water resources are needed, but also because of that ability to recycle water mainly for cooling within the system, their consumption isn't large.

And I've seen studies where they put the water use in comparison just to the golf courses in the United States, and the data centre water consumption pales to the golf course water consumption in the United States. So in terms of it seems to have been maybe a political sticking point, but reality is far different from what the narrative might suggest.

Jeremy: And Zane, if AI takes over the junior tasks through which people traditionally learnt, how can employers, I wonder, develop the next generation of specialists without the expertise to challenge its answers rather than simply rely on them?

Zane: Yeah, it's a tough one. We do see it, we do hear it anecdotally, even in our industry, around the number of interns that are absorbed by corporations is lessening.

That being said, I think, like with any technology shift, and I remember quite clearly the world that my parents ran their business in South Africa versus the one that I walked into was chalk and cheese. an electronic typewriter was a modern invention in their time, and use of fax. I think in terms of the doing bit, the workflow bit, that maybe gets taken away, but you still need an understanding of workflow processes.

And I think that the experience side shifts more towards judgment, management, and delegation. And I think it's also just a bigger question, not only around what they learn in the early stages of their careers, but also what the education system brings to young people for the age of AI, and, probably a lot of reinvention that is required in that regard.

Jeremy: You're listening to No Ordinary Wednesday on Investec Focus Radio with David Smith and Zane Bezuidenhout. Now, after the break, we're going to explore how businesses can judge AI's reliability, examine the economics of AI companies, and look ahead to 2027 and beyond. First, discover Investec Wealth and Investment Global Leaders portfolio.

Managed by an experienced team in South Africa and the UK, the portfolio invests in 30 to 50 higconviction global companies with enduring competitive advantages, taking a long-term approach to capital growth. To find out more, visit investec.com. The Global Leaders portfolio invests exclusively in equities and carries a high degree of risk.

Jeremy: The value of investments can go down as well as up, and investors may get back less than they invested. Past performance is not a reliable indicator of future results. Now, let's get back to the conversation.

All right, Zane, back to you, and how much of the difficulty then in adopting AI comes down to a company's own data being fragmented, outdated, or poorly organised?

And I wonder whether businesses are underestimating that work.

Zane: I think we are in this period of shifting from, at least in a corporate sense, and I know we are as a business, shifting from experimentation to implementation. And what that does highlight is what needs to be done in terms of going back to, Dave's first comment around AI being useful, and that's agentic AI.

It's technically only as useful as the data it has to work with and the understanding and the context of the business that it's operating within. There are some companies that assist with this. On the public side, there's a company called Snowflake that helps corporations get their data ready. There's a private company called Databricks that, competes with Snowflake.

 If you think about, data, if you just think about a spreadsheet for that matter, and you have today's date in one cell, but then you have it formatted in a range of different ways, maybe with a comma in there that makes it unreadable. It really is scrubbing the data and making it interpretable by your artificial intelligent agents and the models that you put on top of it.

So this is a very necessary part for corporations to maximise the usefulness of AI that when they're integrated with their workflows as an organization.

Jeremy: All right, Zane, let me ask you about operational context that you've just referred to. How then should businesses test AI systems that can give a convincing answer even when they're wrong?

Are industries benchmarks then measuring what matters in everyday business use, I wonder?

Zane: I think this is where subjectivity comes into it. So yes, models get measured by, scientific scores and intelligence scores that are there as a standardised measure of capability. But really what, consumers in their personal capacity or what corporations in an enterprise capacity want is usefulness.

They want outcomes, and they want the most correct outcome for a given scenario. And, quite frankly, in a human environment, it's not a case of optimising, but more sometimes a case of satisficing.

So it's about finding a good enough solution to get the outcome that is most desirable. And I think how businesses go about this is, yes, there has to be strategic leadership and direction set in place, but ultimately the context of these workflows are by the users.

So people like Dave and I, when it comes to researching companies, we know best how to measure the outcome that we want, and it might not be the perfect outcome, and through lots of iterations it will improve. So I think it's measuring AI around its usefulness in the eyes of the end user, which is us as employees in the organization.

Jeremy: Now XMTP Labs Chief Executive Shane Mack describing how an AI agent he used for personal and professional tasks posted his bank account information in a company Slack channel. He said the agent appeared to be doing what it thought it should do.

So David, how should businesses decide then what an AI agent is allowed to access and to do, and when it must stop and ask a human for approval?

I sense this is a very important question.

David: Jeremy, you are 100% right. I would first make the argument that you should absolutely be separating your personal life from your work life, whether it's on an agentic level or otherwise. 

I feel like that's sort of the what we all learned around the internet, trying to separate your two lives and not use the same devices or the same systems applies here as well.

But Toby Lütke who's the CEO of Shopify, made a really interesting comment, and his view is that AI can do amazing things and will continue to do amazing things. What it can't do is take responsibility.

In my mind, we have moments where an output is created or something is used that is going to impact a decision, a workflow that is not automated or normal.

You absolutely need a human in the loop to check. And I think that's probably how it plays out, is that you're going to have all these systems which are automated, but you're going to have to continually have them checked, one, by other agents, but two, ultimately by a human to make sure that the errors that come through, just like you wouldn't trust a junior to publish a company's view to clients without some sort of oversight, the same is going to be true for agents.

Zane: If I may chime in. What Dave said is spot on and what we are seeing, particularly on the certain areas within the software stack, and where not all software companies are threatened by AI, is I think there's a growing appreciation, particularly at an enterprise level. the industry's calling this the harness layer, and this is really around cybersecurity.

And one can see what the share prices of cybersecurity companies have done since Anthropic launched their Mythos model and also Project Glasswing off the back of that to select users to patch work where it's revealed certain cracks in their security architecture. There's also observability, so watching these agents as they move.

And then there's an orchestration element with governance also built in, and that's company policies. So it is almost wrapping these agents up in cotton wool to try and ensure that they behave how they should behave and ultimately, like David said, and then a human in the loop, depending on the use case and the workflow, just to make sure that none of that wrapping has fallen by the wayside and that the agent is doing something it shouldn't.

As you said, a very important question, and the industry's increasingly talking around this harness layer in a way to protect users, companies, individuals from agents misbehaving in pursuit of achieving an outcome.

Jeremy: Now, before we discuss individual businesses, I just want to remind you this is analysis, not investment advice.

Jeremy: Listeners should do their own research.

David, back to you, and Anthropic's preliminary IPO prospectus gives us a closer look at the business behind Claude. Very interested to know what you've learned from it. What did you find compelling? And anything there that might have given you pause?

David: It's one of the most exciting growth stories you're going to ever see as a potential investor.

The sheer run rate at how fast this company has grown from having no revenue in 2022 to being at an annual run rate of, call it 70 odd billion dollars right now, just phenomenal. The second bit, which is quite encouraging, is they have 11,900 clients who spend an annualised amount with them of $100,000 or more.

So the breadth of their customer base is really big. The bits that worry both Zane and I are more around long-term moats. If you, if you're talking around defensibility of your business model in the long run, and how... What quality you have to protect yourself. What we've seen the last couple of weeks, the likes of Google and Meta have come out with amazing frontier-end models that can compete with the best models for OpenAI, Anthropic, never mind what's happening in China, and they're having, increasingly having really good models of their own.

So I have some concerns around the fact that the model there actually just becomes fairly normalized in terms of what it can do relative to other models, that they all become amazing, and it actually just becomes a commodity. That keeps us awake at night. Do you need to invest in the model there, or can you just choose other parts of the ecosystem?

Jeremy: So Zane, let me widen the aperture slightly. The persistent question on investors' minds is, are we in an AI bubble? And I'm wondering if the technology can live up to its promise while investors still overpay for the businesses building it

Zane: Jeremy. Yes. That's certainly been the question on market's mind, particularly in the month of October, when we've seen quite a big sell-off in this tech space.

And I would say investors are not overpaying, especially on the public side. And there's maybe some bifurcation here. there's been a lot of valuation momentum in private companies, and what we've seen in the likes of... We've been talking about Anthropic or its, like, competitor, OpenAI, or a few others in that space.

We've seen their valuations ratchet up quite meaningfully. in the public space, we've seen similar growth, but what we've observed is actually that growth has been led by earnings. So it's been earnings, fundamental earnings, that have driven that growth appreciation in these companies' valuations.

And if you look at a, the poster child of artificial intelligence being NVIDIA, you can buy it on a forward 14 times PE as we speak at the moment. some of the companies in the memory space, they're trading on a 4P of two point nine times. So I think the market environment now is that investors have been overly skeptical, where I think the sentiment's gone the other way and fearing that the earnings power of these companies cannot persist for much longer.

Based on the work that Dave and I have done, we've basically got NVIDIA going to shape its financials in terms of margins and returns back to where it was pre-AI after '28. So almost the party's over after '28.

And when Dave and I step back and we look 10 years out and we look at the technological shift, this is really a fundamental shift in how we as human beings do things with readily available intelligence at our fingertips.

And we think that as a driver, when we're looking back in 10 years' time, I think the world would be demanding more compute power in the form of more intelligent tokens than what we currently are, and that kind of gives us longer term confidence for where we are at the moment

Jeremy: So David, a quick follow-up.

What would tell you that spending on AI infrastructure is running ahead of likely demand? And if expectations fall or borrowing costs stay high, which businesses are most exposed, in your opinion?

David: This cycle is quite different to previous cycles when you've had lots of CapEx. Think about what happened with the internet fibre rollout, or if you had with railway networks, you go back a lot further.

Those bits of infrastructure, you had to build out a whole chain. So think about connecting New York to Chicago by rail network. So you have to build an entire rail network, which took a long time, took a lot of cash before you even got to a stage where people could build a carriage to run on the network.

What's different now is that as soon as people turn on data centres, there's a massive backlog. Those chips light up immediately, so there's no dark chips at the moment out there where you've got excess compute that people don't want to use. In actual fact, if you look at the ratio of backlogs for the cloud players relative to their revenue, it's almost never been higher.

We are comfortable right now that there's a mismatch in demand and supply. One of the data points we track, it's not going to be a perfect predictor, but it's useful anyway, is the spot prices of renting a GPU, and what we see there is the prices continue to trend upwards. In other words, the demand for those GPUs are stripping supply, and those who actually have them online can charge more and more each month.

We do track that, but fundamentally we think we are quite some way away from having too much capacity. If anything, we think we will have an undersupply of capacity for way longer than the market expects. But when we get to the point, 'cause there will ultimately be a point where we have too much, the entire supply chain is going to come under pressure.

The very high margins you're getting from the chip and memory players will definitely be pressured, but we think we are much further away than many people expect.

Jeremy: All right. A question to both of you, and let's look ahead five years, 2031. If AI assistants become commonplace, what happens to businesses that currently depend on customers visiting their websites, comparing products, or dealing directly with their staff?

And where could South African businesses find an opportunity in that change? David, do you want to pick that up?

David: The starting point is you need to be fit for the new world. So I think that there's a fairly high chance what will happen is your agents as a company will talk to your customers' agents, so you need to prep for that because the, a lot of engagement will happen on that side.

What you tend to find, and I've tried this out with my own agent, Instinct, is that I still want a range of choice. So I recently booked a flight to New York to visit a friend, and as part of that process I said, "Please find me a bunch of flights. This is when I want to fly, this is the time I want to arrive by," and it gave me a variety of options and I chose.

So you still will have that final point where you might have the person go and pull the trigger, or your product will be on a shortlist where you are. What that means is you probably need to have background checks to make sure that your products are competitive price-wise, and that they are not left behind, and that they are fit for purpose when an agent comes knocking on your website.

Jeremy: Zane, anything to add to that?

Zane: Yeah. I think it's also an exciting time for businesses. we discussed this a lot internally. To enhance your productivity, it's teaming up with the agents. agents will become colleagues. The agents that Dave and I use in the workplace are there to assist us to enhance our productivity.

 So economically, I think where you find economies that have been maybe constrained by a certain labour force or, how much the labour force is skilled or unskilled, you're basically putting the world's intelligence in the hands of employees, and they understand what their business is trying to achieve.

And as Dave said, yes, there will be agent-to-agent business, but also for your customer, at the end of the day, is choice. we already live in that world where, dealing with an airline, you're probably dealing with a chatbot or an automated call centre up until a point where maybe the query is a little bit technical or is, unique or bespoke that requires human intervention.

And I think that's how this will evolve and progress, is kind of the more remedial tasks will shift to a more automated level, and then with increasingly, complex or bespoke scenarios shifting to humans. And then AI capabilities will enhance where it's supporting the human employees to enhance productivity for the businesses.

So we do view this as a productivity enhancement over the long term.

Jeremy: So Zane, let me stay with you, in assessing the different layers of the AI value then, would you rather own the company selling the picks and shovels, the companies building the AI models, or the businesses deploying AI?

Zane: I think where we are in the current market environment, I think it's all right to be in the picks and shovels still.

We think there's durability to this build-out. It's not going to be in a straight line, and there are going to be bumps along the way. And we think the market is paying investors for that at the moment, where things have sold off in the month of October.

In terms of how we, our lens, as long-term investors for our clients, I think we, over the long term would look to the businesses that are ultimately deploying AI, and a lot of these are your established tech businesses, your cloud service providers, or some of them, the hyperscalers.

These are companies that are building an asset, a long-term asset, in the form of these data centres, which basically they are renting out intelligence to the world. They have existing moats, and they are using this technological shift to enhance their earnings power and those moats going forward.

So picks and shovels, yes, for now, but in terms of durability, we do like the big names sitting at the centre and being your cloud service providers.

Jeremy: All right, gentlemen, we're coming to the end of this fascinating conversation, and David, let me come to you first. Give us one prediction for tech in 2027 that we can come back and test.

What do you think is going to change, and how will we know it's happened?

David: I think the spread of consumer agents is going to be one of the things that will surprise you. So Metamuse and Instinct, the two apps I talked about, and there's some access. You can't even get them outside of Japan and the US at the moment on the Meta side.

Instinct is only invite. But my suspicion is that more than half the people you know in a year will have access to an AI agent that will do a range of tasks for them.

Jeremy: And Zane, your prediction, and which widely anticipated breakthrough possibly do you think will still be waiting to happen, shall we say, 12 months from now?

Zane: I think I'd be guessing. the world in terms of anticipated breakthroughs, there's been a lot of chatter outside of the AI world, but loosely connected in the quantum realm, and the effectiveness and the usefulness and the readiness primarily of that. From what we hear, it does feel like things have been getting closer, but also the commentary around that, people have been saying that for some time.

But it really does feel in terms of quantum computing that we'll still be waiting, but there'll be some progress that's been made.

Jeremy: Gentlemen, our first conversation was fascinating, and let me tell you both, this was even more so. David Smith, Zane Bezuidenhout, thank you both, and we are going to keep those predictions for the, next time we converse.

Thank you for listening, and for more conversations on the forces shaping business and the economy, follow Investec Focus Radio on Spotify, Apple Podcasts, or YouTube. Until next time, goodbye.

Disclaimer: The views expressed are those of the contributors at the time of publication and do not necessarily represent the views of the firm and should not be taken as advice or recommendations. Investec Limited and subsidiaries authorized financial service providers, registered credit providers, and long-term insurer.