Subscribe
Share
Share
Embed
This month, James Tulloch is joined by Valeria Moore to discuss the Artificial Intelligence investment cycle, from infrastructure spending and enterprise adoption to power constraints, regulation and security.
The episode examines whether strong demand can sustain current levels of AI investment (03:16), how businesses are choosing between different models (07:49) and why the next phase of adoption may favour cloud providers (05:36). The discussion also considers the risks surrounding cyclicality, capital expenditure, supply bottlenecks and the returns generated by AI projects (10:10).
This podcast is brought to you by TrinityBridge - trinitybridge.com - and features James Tulloch, Head of Investment Specialists, in conversation with Valeria Moore, Deputy Head of Equity Research.
We've created this podcast to set out possible approaches. Please do not view it as financial advice, or its content as investment recommendations. Just because an investment or an investment strategy has performed well in the past, it does not mean it will continue to do so. Our predictions are based on information that is currently available, however events and markets can and do change rapidly.
Creators and Guests
Host
What is Insight Talks?
Our in-house experts share our views on the current market conditions facing investors. Brought to you by TrinityBridge.
We've created this podcast to set out possible approaches. Please do not view it as financial advice or its content as investment recommendations. Just because an investment or an investment strategy has performed well in the past does not mean that it will continue to do so. Our predictions are based on information that is currently available. However, events and markets can and do change rapidly.
Speaker 2:Hello, and a very warm welcome to the latest episode of Insight Talks, the monthly podcast from Trinity Bridge. I'm James Tulloch, and I'm joined this month by our deputy head of equity research, Valeria Moore. Valeria, great to see you. Thanks very much indeed for joining me today.
Speaker 3:Very glad to be here. Thank you, James.
Speaker 2:So today we're going to try and explore the prevailing artificial intelligence or AI environment. Now AI has obviously dominated market headlines in recent years as the potentially revolutionary technology takes hold. But in the near term and medium term, its utility, the speed of its adoption, and its ultimate socioeconomic impacts are all a matter for for for debate. Indeed, it's also even been mentioned as something which might pose a threat to humanity. So not an easy topic to discuss in a short podcast, Valeria, but let's try and make sense of the investment backdrop at least.
Speaker 2:Because I guess it's still the case that we are really still in that transition period between the creation and development of this technology and and its ultimate deployment, aren't we?
Speaker 3:Yes. So we're facing an AI revolution now with agentic AI coming fast with unprecedented innovation. And we have to see it in the the the cycle. Each IT era, investment cycle is gonna last a decade. So we are only two, three years into it.
Speaker 3:Mhmm. And as it is complex to build a data center, it requires three years time, and there are bottlenecks. Now, we are, in the build out of this infrastructure, which which is complex. Mhmm. And we've still, yet to see the the full deployment and the returns.
Speaker 3:So that's what is creating complexity and anxiety. And what I can say is that the investment made are huge, but so also we already have had the first fruits of the deployment in terms of revenue growth and efficiency. There is still a lot to be seen. And as always with semiconductors, it's a cyclical industry, so cyclicality risk needs to be monitored on a daily basis. Things change.
Speaker 2:Some of those issues that you mentioned are reasons why there are question marks about the ultimate, near term, at least, monetization of, of AI. Now equity markets have continued to to march higher, generally led by, AI related companies. But there has been more recently some volatility in markets in this space. So might it be the case that investors are becoming a little bit more discerning? And ultimately, I guess the question is whether the level of of earnings and or in earnings growth that we've seen of late is is sustainable.
Speaker 3:Absolutely. That's that's a that's a dilemma and the debate. So let's start by considering the benefits. And the the benefits we've we've seen in the semiconductor companies, the one producing the the chips to make the cloud and then the data center efficient and possible, They've all benefited. They're one providing the tools to make the semiconductors.
Speaker 3:And the the demand and and and backlogs are still strong and have increased considerably. The hyperscalers that benefited too, they had to invest in CapEx, but they had accumulated healthy cash flow to to do so, and this is an investment case. They're seeing their revenues accelerating in growth. They have a strong backlog, and they are starting also to adopt AI internally to use efficiencies and to minimize, to manage the operating cost. Enterprises are starting to use it too, but at the beginning, enterprises are cautious, and they only gradually move.
Speaker 3:So if we see the valuation and for the hyperscalers, we compare, for example, the price earning ratio with the growth, these ratios are at a minimum level historically because investors are discounting all the risks and not fully considering the deployment and the benefit that still are to come. So we we still need an enterprise penetration and and the use of large learning model to to increase and the benefit to to to show in a larger way.
Speaker 2:And and what about the the costs of of producing this technology across the AI ecosystem, Valeria? So from the chip makers to the hyperscalers building out the AI infrastructure to those companies developing the models which end users interact with.
Speaker 3:Of course. So the the cost is complex. We have the cost of of tokens. And as mentioned, they are coming down. And and this is driven by by by the cost cuts for from the model providers, but also by the fact that they are improving the efficiency in newer model generation.
Speaker 3:And the cost of the semiconductor side, here, there is a supply constraint, and so costs are are increasing, and and and that is an issue. But it's also it's also offering the opportunity to apply innovation. And and and, for example, the hyperscalers, the one who invest in the cloud infrastructure and and offer the the cloud services, they can do efficiencies internally and, for example, develop specific cost of chips that allow significant savings when they are used at a scale for a specific workload. So costs are high with the silver lining, which is innovation allowing efficiencies at at scale. And, ultimately, the the enterprises...
Speaker 3:We need to see the enterprises adoption increase, and that that is gradual because they they have IT budgets that are limited. They want to experiment, but they are also limited by by the budgets. And and and so we need time for for for that adoption to increase. And then also, we need time for more efficiency to be built into the equation and for the CapEx cycle to give way to widespread adoption. It will be a choppy path.
Speaker 2:Sure. And one of the reasons it might be a choppy path is a sort of level of uncertainty about ultimate end user demand. I mean, might it be the case that end users are becoming a little bit more discerning about the AI models that they adopt? We hear a lot these days about small language models versus large language models or open source models versus closed source models. Might be the case that some companies, maybe non technology companies, decide that less sophisticated, less expensive models are are sufficient at least in the near term, and therefore, the the level or extent of of AI adoption might be being overestimated a little.
Speaker 3:So this is very fascinating. We we are this year in a period period where Genetic AI is is really flourishing. And we have seen many large learning models and and also agents in Genetic AI coming and being adopted by enterprises. I think we can also see enterprises being good at rationalizing the adoption of models for different uses. A very large foundation models can be...
Speaker 3:Which is costly, can be used where it really is is needed. And and then everyday problems can be solved by another open open source model that is available and and and less expensive. So there is this orchestration of different needs, and models keep leapfrogging each other, and there is a lot of competition. They keep improving, but also become more efficient, and the cost of of the token decreases. So the token...
Speaker 3:What what is the token? The the token is the commercial unit of the agentic AI by by which revenues and usual... Use... Usage is is calculated. So as these costs come down, enterprises usage and benefit can increase together with their efficient attempts, and then we have a further penetration of of the use of of of model, which can make us, the enterprises, more efficient.
Speaker 2:Now as you touch on the level of capital expenditure, particularly amongst hyperscalers, has been astronomical. Is it the case that cost pressures may lead hyperscalers to reconsider the level of of capital expenditure they can continue with?
Speaker 3:CapEx are high, and they're still growing because we are at the beginning of the investment phase. Hyperscaler have been through other phases. They have previously accumulated cash flow, and now they can use those cash flow to invest. They are very cautious, and they will only consider those investment that have an appropriate return on investment. And so when that doesn't happen, then they will not invest.
Speaker 3:But I'm sure so far, we we are at the very beginning, and and they're finding this... That use case is profitable. And and and so we we see that also in the backlog that is supporting the the revenues and the demand. And and financing, I mean, they are very large, and and they finance out of cash flow, but also out of debt. And I think that is important, the financing, to be monitored because in in...
Speaker 3:Initially, they're large scale. They'll be okay. But when we have to monitor the the smaller project and the neo cloud and how they support the finances. So this is an important risk to be monitored.
Speaker 2:Yeah. Yeah. Absolutely. And and you touched on it earlier, Valeria, but one of the key themes that comes up often is the the sheer amount of energy required to to power AI. Should we see widespread adoption?
Speaker 2:You know, is it even the case that broad based adoption of AI, should it come about, could be realistically powered as things stand?
Speaker 3:Power is one of the the three main bottleneck Yeah. At the moment. Yeah. And and we need we need we need time to to solve the bottleneck for data center to be built. What I want to say is that efficiency also comes into the equation, and then we have innovation at the chip level to reduce the consumption of power, improve the efficiency on a unit level.
Speaker 3:So there there is more compute also in the equation that will will solve the power dilemma and better architecture also. So the equipment that will go into the data center to to allow a higher scale of power to be directed from the grid to the data center with lower dispersion. So an important part of the element.
Speaker 2:Okay. And finally, Valeria, what about the issue of regulation? There's been some calls in certain quarters for greater levels of regulation here, a pacing of the speed of development. Potentially, there's been some high profile cases of AI models in testing environments managing to escape those supposedly closed environments. And as I said earlier, even some suggestion that AI could pose a threat to humanity.
Speaker 2:So is there a need for greater regulation here, and what might that mean for the pace of development and and adoption?
Speaker 3:Of course. Regulation is paramount and also being discussed. Would like to highlight two aspects of of regulation. First is is the demand for regulation that currently is is being addressed and and highlighted. It is not new, and it is it is needed.
Speaker 3:But it's it's being applied in a way that is not detrimental to to the growth, may limit the AI growth. But, ultimately, it will add the controls that increase the demand for compute because the connectivity, the connection would be monitored. All the flow, the communication between agents that talk to each other will be controlled by other chips. And all of that will improve the demand for for compute, for for chips, for other solutions and allow the development of AI in a safer way if addressed appropriately. That is no easy task.
Speaker 3:The second element that I want to highlight is security. So there is awareness of this, and this is important. There is a higher demand and awareness of cybersecurity software to control the the agents. And I see enterprises being conscious of that, then they apply strict rules when they use agents, a zero tolerance, a zero risk approach so that an agency is allowed only to do a task for which they are absolutely sure there are no other risks. So there is a lot of demand on software that will control agents.
Speaker 2:Okay. Okay. We'll have to watch with care. Valeria, thank you very much indeed. We'll have to leave things there.
Speaker 2:It's a topic I'm sure we could talk about for hours. I've absolutely no doubt we will revisit it again in the future. But for the time being, great to get your insights. Thank you very much indeed.
Speaker 3:Thank you, James.
Speaker 2:And thanks, as ever, to everyone for listening. I do hope you found the podcast informative and interesting. As always, if you'd like to know more, there's plenty more information available on the Insight section of our website, trinitybridge.com, on this topic and many others. So do pay that a visit. But for the time being, thank you very much, and goodbye.