Talking Heads – AI investment: the mega-caps ‘arms race’ is on

This is an audio transcript of the Talking Heads podcast episode AI investment: the mega-caps ‘arms race’ is on

Daniel Morris: Welcome, Derek, and thanks for joining me.

Derek Glynn: Hi, Daniel. Thanks for having me, glad to be here.

DM: It feels like if we’re not talking about politics these days, we’re talking about AI [artificial intelligence]. Investors are aware of the significant returns from the US NASDAQ index this year. More recently it’s fallen back, but to put that into perspective, just back to the levels at the end of June, which was already up 18% for the year. I think investors would be pretty happy with those gains.

 What’s been driving it has been enthusiasm around the potential for AI’s positive impact on corporate profits and the broader economy. Then inevitably there’s the discussion as to whether that enthusiasm is warranted, given what happened in the late nineties and early 2000; no-one wants to see a repeat of that.

So what’s happening with AI and where do you see that potential? Let’s start with generative AI and maybe you could distinguish what is different about generative AI as opposed to AI more broadly. And how could generative AI usher in a new wave of innovation in the economy?

DG: Generative AI is a technology that can generate content from a text-based prompt. Technically, these are large language models designed to mimic the way humans think. They’re trained on a vast amount of data and were originally designed to predict the next word in a sequence. But with more and more data and the right type of training, they’ve been able to reach advanced levels of reasoning.

My conviction in the potential of the technology to result in more economic innovation and productivity has grown over time. The technology itself has improved rapidly. Not only do the answers appear [to be] more accurate and relevant due to new training techniques, but the way in which we interact with the models has improved.

State-of-the-art models today can now incorporate images, audio and soon video. It’s a much richer experience and a better interface for the user. It feels more natural when communicating with the model. These formats also unlock new uses.

For instance, language translations, say from English to French and vice versa, is close to seamless, using the model as the translator and intermediary. We also have many early proof points that suggests productivity benefits for the economy. Developers, for example, can use generative AI tools to help them write code so that part of the code is automatically generated by AI.

Another important innovation is the amount of data that can be ingested by the models. The best models now have the capacity to reason across the equivalent of 10 to 20 novels. This helps developers and others with long-form uses.

Software is growing as a share of total GDP, so productivity improvements in this area tend to have an outsized impact on the economy more broadly. From an enterprise adoption perspective, we’re seeing more and more companies become interested in evaluating the technology, even in heavily regulated industries that have sensitive data, like financial services and healthcare. We’re seeing companies move with a sense of urgency to adopt AI solutions.

DM: We see the major cloud service providers planning to accelerate their spending to support their AI efforts and build out data centres. What are the implications of this spending? More importantly, do you think they can earn a good return on the capital?

DG: We think there is an AI arms race developing between the mega-cap technology companies that are investing in general AI. We estimate the major cloud service providers in the US could spend over USD 150 billion combined this year. If we include a couple other mega-cap technology companies that are building out AI infrastructure, that number could eclipse USD 200 billion. That’s a big number and would represent more than 50% year-over-year growth.

As you alluded to, a key topic of debate in the financial markets is whether these companies can earn a good return on their capex. Some investors worry that this trajectory could result in a lot of wasted spend. Many pessimists tend to draw analogies to the dot.com era of the late nineties and early 2000s, a boom-and-bust period when companies overbuilt global networks of fibre optic cables and eventually faced financial difficulties.

I don’t think that’s a good analogy to what we’re seeing today, and I think it’s unlikely that something similar would play out for the major cloud service providers. There are some important differences.

First, many of today’s mega-cap tech stocks already have strong core businesses and various means to fund their capital expenditure without necessarily incurring a lot of debt. These core businesses tend to generate high amounts of free cash flow. They’re often supported by numerous competitive advantages like network effects or economies of scale. So, we believe their ability to generate cash is durable.

They also tend to have strong balance sheets with large amounts of cash. So, they have various avenues of funding these initiatives in low cost-of-capital ways. As it relates to the return on capital specifically, I think they’ll be good, but there will [likely] be variations across the cloud service providers.

They should all benefit from activities in the cloud, like storage, computing resources and renting out access to graphics processing units for use in AI applications. However, the best positioned will be those that can use generative AI to develop a new product or service that can be monetised. The highest returns will be generated by those that can expand beyond the infrastructure layer and develop AI-enabled applications that become widely adopted.

DM: We’ve been talking about the cloud service providers. What industries and types of companies do you see as well positioned to benefit from generative AI? Who will be the winners and the losers?

DG: There are many industries that are benefiting from these high levels of spend and from the technology itself. For instance, the early winners include many of the semiconductor and semiconductor equipment companies. They are fundamentally enabling this technology. We also think there will be opportunities for companies in non-technology sectors that could benefit from the buildout of data centres.

For instance, industrials companies that provide cooling solutions could benefit, as could alternative energy companies like solar, because many of the cloud service providers have net zero commitments that they need to meet.

Cybersecurity companies could also stand to benefit. Bad actors could access these powerful new AI tools, and that’s essentially a new threat vector that needs to be protected. An interesting area here is in machine identity and authentication, as AI models will increasingly act autonomously and interact with each other. Those machines need to be authenticated so that the right models are talking to each other and are protected.

We also believe those with proprietary data more generally are well positioned because they can integrate generative AI to develop new products that can be monetised or to improve retention rates with existing customers.

In terms of risks, we are wary of companies that could get disrupted or disintermediated, particularly if they don’t adapt quickly enough. Some software businesses could face more competition, particularly if they don’t have proprietary data or have other barriers to entry.

In other cases, for some industries like e-commerce, AI features will essentially be commoditised. These features won’t necessarily help companies differentiate themselves from the competition. They’ll just be common features of the product or service. It’s only a matter of time before every major e-commerce site has an AI-powered copilot that assists consumers with their shopping and helps them find the right items and answer questions about those items. But those features won’t necessarily be monetisable over and above their existing businesses.

DM: to the last question: what about the future? We appreciate how rapid the advancements have been around AI over the last year, even if it’s been many years in the making. What can we expect to see over the next few years?

DG: We’ll see continued refinement in the ways we interact with these models. There’s still progress to be made in the audio and video formats.

For example, another trend will be much broader enterprise adoption likely to occur in 2025 and beyond. Although companies are evaluating and beginning to adopt it today, it just takes time to get their data estates in order and to get comfortable with [the] security, governance, compliance and privacy standards of generative AI.

In time, I would expect to see improvement in categories that require very advanced levels of reasoning, such as some higher levels of mathematics and in long-term planning. Generally, the latter could be an important unlock for one day attaining artificial general intelligence or AGI, which is loosely defined as an AI model that has attained or surpassed human levels of reasoning.

Finally, I would expect greater clarity in terms of regulations around AI and a continued focus on the risks. There is still more work to be done on potential issues like copyright, data security and privacy, misinformation and AI safety more broadly.

DM: Derek, thank you very much for joining me.

DG: Thanks, Daniel. Glad to be here.