On AI: the Age of Extremes 

We seem to be at that stage of the market cycle where extreme views dominate. This holds true for many things, but we feel it is especially evident in the current AI debate. On one side are the AGI¹ evangelists who see an imminent singularity that will allow whichever company that gets there first to dominate across most industries. On the other side are the doomsayers who see a bubble, overinvestment and low returns. But could it be that both views are partially right? In our opinion, AI is real and will touch most things – but the investable outcomes are likely more modest and more diffuse than the headlines imply.  

AI is a truly general-purpose technology. Like the invention of the automobile a century ago and containerisation 50 years later, AI will change the way we work and reshape entire industries. However, history also shows that much of the economic value added from these technological innovations accrue to consumers and society. Prices tend to fall as capacity increases, and companies that adapt their business models tend to gain share. Investors do well when they back firms that are able to convert cost deflation into durable competitive advantages. They do less well when they back the builders of generic infrastructure once competition and capital catch up.  

Warren Buffett, in a 2001 speech to students at the University of Georgia, used the analogy of the auto industry at the beginning of the 20^th century to describe how most fast-growing businesses fail to deliver the desired returns over the long term. While the invention of the automobile and its subsequent impact on society were significant, understanding a business’s economic characteristics is different from predicting the success of an industry. Out of the 2,000 American auto companies that existed in the 1920s, only three survived into the next century and – to Buffett’s point – their performance over time has been questionable.  

The same point was made in a recent essay by venture capitalist veteran Jerry Neumann, AI Will Not Make You Rich (Colossus, Sept 2025). Drawing on the history of containerisation, he argued that certain technological revolutions – though transformative for society – had largely failed to generate lasting wealth for investors.  

Containerisation radically lowered the cost and complexity of global trade, accelerated globalisation and lifted productivity across industries. But while it reshaped the world economy, almost none of the companies that built or operated the infrastructure became sustainably profitable. Competition intensified, capacity was overbuilt, and returns were competed away. The real winners were the firms that adapted their business models to exploit the new logistics paradigm – companies like Walmart and IKEA, which used cheap and predictable shipping to build global scale and lower prices for consumers.  

AI could prove similar. The societal impact will likely be profound, but the profits may accrue to the users or those that redesign processes, distribution and products around the new technology – not those that supplied the early infrastructure. As with containerisation, the early infrastructure phase has attracted enormous sums of capital expenditure (capex) ahead of the establishment of sustainable demand. Today every hyperscaler is building its own data centres, clusters and model infrastructure in parallel, each betting that scale or first-mover advantage will ultimately translate into dominance.  For the equipment suppliers in the value chain, this looks like a golden age – multiple customers racing to outspend one another, all drawing on the same supply chain. But this phase is inherently unsustainable. It’s a bit like every company laying its own set of railway tracks between the same two cities, even though there’s only demand for one line – maybe two at most.

The reality of today’s AI race is that most model capabilities are converging. Open-source competition is narrowing the gap and customers increasingly treat models as interchangeable. If this continues, we should expect a gradual shakeout, as one by one the weaker players drop out of the capex race when monetisation fails to keep pace with cost. The beneficiaries at the infrastructure layer – the chipmakers, memory suppliers and component manufacturers – will feel the slowdown first.   

AI could of course be different to the innovative technologies of the past. However, in our view the current level of capex spending can only be justified if one of two things happen:  

1. One company reaches true AGI – a form of machine cognition that can reason, plan and generalise across domains. Or: 
2. One or more players succeed in building a network effect or high switching cost that locks in customers and generates pricing power.  

This is essentially the “winner-takes-most” scenario that underpins today’s hyperscaler capex boom. Each of the big players – OpenAI/Microsoft, Google, Anthropic, Amazon and Meta – is investing as if AI dominance is attainable. The logic is simple: if intelligence becomes a scarce and tradable resource, then whoever gets there first captures the profit pool.  

Both of the above outcomes remain possible, but in our opinion, they are not likely to be achieved in the foreseeable future. The distinction between cognition and statistical inference is central to our view. Today’s large language models are extraordinary tools of inference – they identify patterns, correlations and likely continuations of text or code across massive datasets. But that is not the same as cognition. They predict; they do not think.  

If AGI is truly about cognition – the ability to reason from first principles, to generate original ideas, or to act autonomously with persistent understanding of the what and the why – then the current architecture seems fundamentally limited. Scaling up parameters and computation power may improve performance, but it does not necessarily move us closer to the AGI end-goal. In other words, while the models may get better at predicting the next word, they still won’t be able to understand the world.  

That distinction matters for investors. The “winner-takes-most” thesis assumes one company will control the key models that everyone depends on, giving it monopoly-like economics and the ability to continue the capex spending. But if the underlying technology remains statistical rather than cognitive, competition will stay intense, open-source models will narrow the gap and switching costs will remain low. In that case, AI could still be massively deflationary and societally transformative, but the economics will resemble container shipping rather than a natural monopoly – high volumes, low margins and most of the economic value added captured downstream by users and consumers rather than the builders of the infrastructure.  

We are not expecting an imminent “over-the-cliff” moment, but we think it makes sense to think carefully about where in the value chain one is exposed – whether that be in training, inference or productisation. The three stages are economically distinct and will not be affected equally as the cycle matures.  

Training – the phase where models are built and refined – is the most capital intensive and cyclical. Demand is currently inflated by parallel model development across multiple players, each racing to train larger versions of broadly similar architectures. This layer has attracted most of the incremental spending and is where valuations have already rerated the most. If the industry consolidates around a smaller number of models, or training efficiency improves, spending here could slow sharply.  

Inference – the application of trained models to generate outputs – should grow as usage scales but is likely to become increasingly price-competitive, as it is the stage most at risk of commoditisation. Performance gaps are narrowing, switching costs are low and open-source models are proliferating, while efficiency gains, falling costs and standardised interfaces make differentiation difficult and pricing power hard to sustain.

Productisation – embedding AI into consumer devices, enterprise software and physical systems – is still in the early stages, but it should prove to be structurally more durable. It depends less on model scale and more on user adoption and workflow redesign. This is where stickier economics emerge, as companies integrate AI into their products and processes, creating switching costs and recurring revenue streams.  

The market has so far rewarded the training layer where growth has been front-loaded, but as capex normalises, we expect the focus to shift from capacity expansion to productivity gains – from those funding the infrastructure to those using it effectively.  

Following Neumann’s logic, the risks are therefore highest at the model-training layer, where it is still unclear how the spenders will effectively recoup their investments. From a capex perspective, the longer-term value and more durable opportunities lie further downstream in the productisation layer – companies geared to the eventual proliferation of AI into consumer devices, robots, cars and other end-markets should be better placed.  
 

With this backdrop, we have been thinking about how these trends could affect the portfolio. At a headline level, we believe AI adoption will be a long-term tailwind for most of our holdings, albeit indirectly, by enhancing productivity, improving customer experiences, and in many cases allowing our holdings to build even stronger moats around their businesses.  

In the near term, however, we see risks to the current capex boom and have started to position the portfolio accordingly. We have direct exposure to AI-related capex spending through TSMC and SK Hynix, and to some extent Mediatek. The first two will be affected if AI capex slows meaningfully at the training and inference layers, but TSMC should still benefit from the broader diffusion of AI into end-devices and industrial applications. Its natural monopoly at the leading edge of technology, and close integration with key customers, make it a structural beneficiary of productisation, which we believe will accelerate in the coming years even if near-term growth in training-related demand normalises. 

Mediatek’s exposure comes mainly through its custom application specific integrated circuit (ASIC) business, which designs the chips used in AI accelerators and edge devices. While a downturn in AI spending would likely affect this segment, it remains a relatively small share of revenue – roughly mid-single digits by our estimate – and should be manageable within the context of the broader smartphone and connectivity business. Overall, our largest concern is with SK Hynix, which has been a major beneficiary of the current training-driven high bandwidth memory (HBM) upcycle. We have used the recent share price strength to trim our position, reflecting our view that this part of the value chain is most vulnerable if spending on large-scale training infrastructure begins to moderate. 

Indirectly, however, we expect most of our holdings to benefit from growing AI adoption. Many of our software, internet and technology companies – together accounting for roughly 50% of the portfolio – should see cost deflation and improved customer engagement. In practical terms, AI can help them automate support functions, personalise services, refine product recommendations and target advertising more precisely. The effect may be gradual, but over time these applications should strengthen their profitability and deepen competitive moats. 

One such example is Totvs, the Brazilian enterprise resource planning (ERP) software company we have held in the portfolio since 2022. In a recent meeting, CEO Dennis Herszkowicz described AI adoption as a long-term structural opportunity, potentially more significant for Totvs than the earlier transformative shift from on-premises to cloud, which helped lift group margins by around 10% over five years.²

Totvs’ advantage lies in its proprietary client data and local domain expertise. Its software underpins core operational and financial processes across a large installed base, giving it access to high-quality, structured datasets that can be used to train models specific to each industry vertical. This should make its solutions more accurate and more deeply integrated into client workflows, increasing switching costs and reducing the risk of displacement. 

Once developed, these AI modules can be scaled across its customer base at minimal incremental cost, which should support gradual margin expansion as adoption broadens. At the same time, new AI functionality adds a monetisable layer within existing systems, raising revenue per client without proportionate cost increases. In short, AI should help Totvs strengthen its competitive position while improving returns on capital. 

Tencent is another example of how AI can strengthen existing franchises rather than create entirely new ones. While it is not building foundation models at the same scale as the US hyperscalers, Tencent has unique advantages: a massive user base across WeChat, Games, and Payments; rich first-party data; and integrated advertising and content ecosystems. 

AI allows Tencent to sharpen the focus of the targeted ads across its network – understanding context, intent and consumption patterns with far greater precision. This should translate into higher ad yields, particularly as AI enables dynamic creative optimisation (adjusting visuals and copy for each user). Beyond monetisation, AI tools have also improved engineering efficiencies across the organisation. Code generation and automated testing have reportedly shortened development cycles by 20–30% in certain product teams, helping Tencent deploy new features faster and at lower cost. 

In gaming, Tencent can use AI to personalise player experiences and generate in-game content dynamically, thereby extending product life cycles. Across its enterprise cloud and mini-program ecosystem, AI enhances the productivity of its developers and enables it to offer more responsive, data-driven services to its small and medium-sized business clients.

The cumulative effect of AI integration across Tencent’s business should be higher productivity, more relevant content and stronger engagement – the kind of incremental compounding that rarely grabs headlines but adds to its long-term franchise value. 
 

We think the picture for emerging markets looks supportive from both an absolute and relative basis. Valuations are attractive, currencies are cheap, and unlike the US – which may face a few lean years that could weigh on the dollar and prompt global investors to seek alternatives – emerging markets appear poised for a stronger growth outlook after what has been a few pedestrian years. Moreover, we believe that the possibility of a “deal” between China and the US could lead to a reassessment of the notion of China being “uninvestable” in some parts of the world, potentially driving renewed interest in emerging markets. After all, China remains the largest market in a global emerging market context. 

From a bottom-up perspective, we expect the portfolio’s aggregate earnings and free cash flow to grow by mid-teens annually over the next two years. We’re currently paying a 5% free cash flow yield for that growth – that is a discount to long-term average valuations and is not building in any rerating assumptions. This makes us optimistic about the return potential from here. We believe the combination of reasonable valuations, improving fundamentals and strong underlying businesses provides a solid foundation for attractive long-term returns.

As always, we appreciate your continued support. If you have any questions about the strategy, our approach, or specific holdings, we would be happy to discuss them further. 

References

¹ Artificial General Intelligence – a hypothetical artificial intelligence that reaches human-level general intelligence.

² Source: All company data herein retrieved from company annual reports or other such investor reports. As at 13^th October 2025 or otherwise noted.

Source: All company data herein retrieved from company annual reports or other such investor reports. Financial metrics and valuations are from FactSet and Bloomberg. As at 13^th October 2025 or otherwise noted.