AI is changing the world. Who pays the bill?

16th AUGUST 2026

Our thoughts on AI, and why we are cautious on behalf of investors despite the technology’s undeniable potential.

Mark Twain is said to have coined the line: “History does not repeat itself, but it rhymes.” Whether the subject is politics, war, economics or bubbles, Twain has a point, as always, and perhaps also when it comes to AI. In this letter, we share our thoughts on AI and explain why we are cautious on behalf of investors, despite the technology’s undeniable potential.

AI – who captures the value?

We invest exclusively in companies whose qualities support robust, long-term value creation. Common traits across our portfolio are growing, recurring and non-cyclical revenue. A natural consequence is that there are many industries we stay away from. One of them, for now, is the AI supply chain, which has generated massive media coverage and outsized returns.

The technology has already transformed many key routines and processes in our daily lives, both at work and at home. This has driven sharp share-price gains for the companies at the forefront of its development. Even so, we have yet to find investment opportunities in the AI build-out that have the characteristics we look for. The potential of AI as a technology is large, but the conditions for predictable, long-term value creation for shareholders in AI companies are questionable. History supports our concern, so let us take a walk down memory lane.

Transformative technology vs. investor returns

The euphoria around AI stems from the fact that the technology has already had a major impact on society and will continue to shape how we work and live. However, 200 years of transformative technological progress show that society-changing technology is by no means synonymous with attractive returns for those who invest in it. Historically, the sure winner has (fortunately) been society, but often at the risk of heavy losses for those who financed the build-out of the new technology.

The historical examples behind our scepticism and restraint when it comes to investing in new technology range from the British railway bubble of the 1840s to the internet infrastructure build-out at the start of this millennium.

In 1840s Britain, railways were seen as an infrastructure breakthrough that enabled faster transport and opened up new markets. Enthusiasm turned into an investment frenzy in which hundreds of railway companies were founded. As a result, many projects were planned simultaneously, often on optimistic traffic assumptions. Investment grew explosively from 1843 to 1846, but one-third of the projects approved during the bubble were never built, and by 1850 investment had fallen 70% from its 1846 peak. Indices of British railway shares from the period show that prices never regained their bubble peak within that century, underlining the dismal returns investors earned, even though the railways were here to stay.

In the late 1990s, the arrival of the internet was seen as the start of a new economic era. Traffic grew rapidly, but not as rapidly as investment in the infrastructure behind it. Vast sums went into fibre-optic cable, network equipment and server capacity. Many players invested in parallel because they wanted to be first and biggest. In the US, capacity utilisation among internet infrastructure providers reached almost 90% in 2000.

Over the following two years, utilisation fell to 50%, and the share of companies in the sector in financial distress rose from a few percent to 30%. In the years leading up to 2000, the sector was a gold mine for investors, with almost every internet-related stock posting large gains. But anyone who invested at the peak of the euphoria, measured by the Nasdaq-100, had to wait 15 years to get back into positive territory. Throughout those 15 years, internet usage continued to grow explosively. Once again, the technology changed the world, and once again it was no golden ticket to strong long-term returns for the vast majority of investors who funded its build-out. We could go on with similar examples, such as the telegraph (and later the telephone), aviation, automobiles, radio, shale oil and more.

Does any of this sound familiar today? We think so. The common denominator for promising technologies that subsequently delivered poor investor returns is the massive flow of risk-seeking, hopeful capital pouring into booming sectors and industries – be it railways, the internet or AI.

A recurring pattern is that heavy investment is required today to reap uncertain gains tomorrow. The result is a race to build the most infrastructure in order to defend market share, despite uncertain or absent future returns on that capital.

In AI, one only has to look at the broad range of Western AI products – ChatGPT, Claude, Gemini, Copilot, Grok, Mistral and others – and Chinese models – DeepSeek, Kimi, GLM, Yi and Qwen – to see that competition in this part of the industry is already extremely fierce and price-sensitive.

Staggering sums

History shows that heavy capital commitment in unpredictable, competitive industries is a recipe for poor long-term returns. At present, the switching costs of moving from one AI model to another are very low, which means price competition will in all likelihood intensify over time.

Over time, this will affect the companies building data centres. If data-centre profitability disappoints and their owners stop investing, it is usually the suppliers dependent on recurring product sales that are hit hardest – including NVIDIA, SK Hynix and Micron.

The biggest near-term issue we see is that the AI build-out depends heavily on revenue that has yet to materialise. Goldman Sachs now expects total AI capex of approximately USD 1,000 billion in 2027 – almost four times the annual investment in renewable energy across Europe, North America and Australia combined. For 2026–2031, Goldman Sachs expects cumulative AI capex of approximately USD 7,600 billion.

Unlike railways or internet cables, this is equipment that will very likely become obsolete within a few years: roughly two-thirds of the investment in a data centre consists of assets with an expected useful life of just 3–7 years, which makes near-term end-customer revenue essential. On a rough, simplified estimate, USD 1,000 billion of annual capex must generate revenue of around USD 1,700 billion to break even at EBITDA margins of 60% – margins that may be hard to sustain as competition intensifies. And that covers only the data centres themselves.

Customers in the real economy must also fund developers’ salaries and marketing, plus a margin for the intermediaries building applications on top – so actual end demand would need to be closer to USD 2,500 billion per year. For comparison, the combined run-rate ARR (annual recurring revenue) of OpenAI and Anthropic is around USD 100 billion.

Even if these sky-high expectations are met, AI companies are priced today such that even strong results will not necessarily translate into exceptional returns.

NVIDIA is the world’s most valuable company, with a market capitalisation of around USD 5,400 billion. At the company’s current margins, that valuation requires NVIDIA to generate cumulative revenue on the order of USD 10,000 billion over the rest of its life simply for investors buying the shares today not to lose money. To keep the arithmetic simple, we have ignored the discount rate, which makes the figure (very) conservative in NVIDIA’s favour; at a normal required rate of return, the required revenue would be significantly higher. Since NVIDIA’s chips typically account for roughly 50% of data-centre investment, this implies that around USD 20,000 billion must be invested in data centres for the numbers to add up.

On the surface, the stock trades at “only” 34 times trailing earnings and roughly 22 times forward earnings – multiples that, in isolation, may look attractive for a company that has grown revenue eightfold over the past three financial years. But how durable is NVIDIA’s revenue and earnings growth?

Three points are worth highlighting. First, NVIDIA’s EBIT margins of over 60% are highly unusual for a hardware company, and that level attracts competition – both from other chipmakers and even from NVIDIA’s own customers, the hyperscalers. Perhaps not in the short term, but as the American economist Bruce Greenwald once put it: “In the long run, everything is a toaster.” NVIDIA’s technology is far ahead of the rest, but its market position gives many stakeholders strong economic incentives to change that. Second, customer concentration is extreme: a handful of hyperscalers account for the bulk of revenue, and their capex budgets can change quickly. Third, a growing share of this investment is funded not by operating cash flow but by debt and by lease and guarantee structures that will only hit the hyperscalers’ balance sheets in a few years. Goldman Sachs expects roughly one-third of hyperscaler capex in 2027 to be debt-financed, and Moody’s puts the five largest hyperscalers’ not-yet-commenced data-centre lease commitments alone at USD 662 billion, equivalent to 113% of their adjusted debt. In 2026, this figure is likely to be materially higher. NVIDIA’s earnings therefore rest on a concentrated customer base that is more cyclical and more financing-dependent than the apparent stability of its order book suggests.

But this is about more than a single stock. It is about an entire industry that depends on the music continuing to play, on margins staying high, and on returns on capital across the value chain proving satisfactory. So far the music is playing, and as shown below, AI-related stocks make up a growing share of the major indices.

Chinese models are taking a growing slice of the pie

The thesis has always been that AI models would be expensive to train but highly profitable to rent out for a range of tasks. But as switching costs between models have fallen close to zero, and Chinese open-weight models now match most of the performance of their US counterparts at roughly one-fifth – sometimes one-tenth – of the price, the most sophisticated customers are cutting their AI costs even as their AI usage grows.

The crypto exchange Coinbase has disclosed that it has almost halved its AI costs despite a sharp increase in usage, by routing simple tasks to cheap models such as GLM and Kimi and ensuring that repeated queries barely touch a model. The development in its AI token consumption and the associated cost is shown below; the colours represent different AI services, although their names are not disclosed.

Shopify, Siemens and Airbnb have moved workflows onto open Chinese models such as Qwen, and one of our own portfolio companies, RELX, has said it can migrate to a new model in a matter of hours.

The problem, in other words, is that AI adoption could plausibly triple or quadruple with no increase in total AI spend. That is excellent for society, but dangerous for the revenue and profit assumptions embedded in share prices.

The expectation was always that OpenAI or Anthropic would invest hundreds of billions of dollars in training models and then receive a long and “secure” stream of cash flows as the models were used year after year. That may prove to have been wishful thinking.

AI becomes consumer surplus

Which industry players will be able to defend their market positions over the long term and generate value-creating returns, we will leave to others to guess.

In our view, predictability is close to zero. As in previous periods of technological progress, we expect the economic gains to accrue largely to society, as the ruthless competition of capitalism will ensure that most companies pass any efficiency gains on to customers in the form of lower prices.

The companies that benefit most from AI will likely be businesses with already well-protected, scaled business models. These companies can use the technology to drive internal efficiencies and redeploy resources to higher-value activities, without competition in their industry forcing them to pass most of the gains on to customers.

Put differently: if OpenAI and Anthropic cannot charge a high price for everyday use of AI models, the value of AI accrues to users and consumers rather than to producers.

AI panic and data businesses

In our view, a good example of an industry that will be able to retain part of the gains from AI is one the market has, curiously, chosen to punish: owners of proprietary, non-replicable data.

A decade ago, the mere mention of Amazon entering a market could wipe 30% off the market value of that industry’s companies. We are currently seeing the same pattern with AI. When Anthropic launched a superficial legal plug-in, it triggered sharp share-price declines among companies providing legal data and solutions to law firms. Among them was one of our portfolio companies, RELX, which fell 14% on the day of the news.

RELX’s business model is built largely around unique, proprietary datasets which, combined with the right tools, create irreplaceable, deeply embedded data solutions for customers in fields including science, finance and law. Legal accounts for approximately 13% of RELX’s total earnings and consists of a deeply integrated ecosystem of legal data and mission-critical tools.

Every panic, like every bubble, contains a grain of truth. AI will lower barriers to entry in many industries and cause bankruptcies across sectors. It does not follow, however, that every industry AI touches immediately becomes less valuable.

Quite the opposite. In our view, data and AI models are complementary goods that increase each other’s value – but the data can be owned and proprietary, whereas AI models already appear to be a commodity. At the same time, AI models are deeply dependent on data and its quality:

Models are trained on data. The classic “garbage in, garbage out” very much applies.
AI requires up-to-date data. The best data assets are continuously updated, not static. They cannot be “stolen” once, as the model would quickly become outdated.

For owners of the right data, we also see several arguments that AI will, on balance, create more business opportunities and revenue streams:

Better analytical tools make data more valuable, not less. Card transaction data, for example, is only useful against fraud when combined with the right tools and algorithms. AI can enhance the value of data and thereby strengthen the already considerable pricing power of those who own it.
Software-for-labour increases data consumption. As work shifts from people to models, more decisions are made more often on more data, making data quality and integration even more business-critical.

On this basis, we see good prospects for RELX to continue delivering value-creating earnings growth and to maintain its competitive position and robust foundation, also in a world with AI. Combined with the substantial share-price decline over the past year, driven by AI panic, this has made RELX one of the fund’s larger positions.

Popcorn at the ready

We know that the future, as always, will look markedly different from the present, and in our view heavy investment in AI carries high uncertainty and intensifying competition. Our investment approach is uncompromising, and we will not invest on the basis of hope or rosy forecasts from investment banks’ AI research. When everyone is talking about it, alarm bells should ring.

We hold several companies in the portfolio that, like RELX, benefit strongly from AI thanks to their well-protected business models. They simply do not make money from building out industry capacity. We own a small, distinctive cluster of quality companies that we understand in depth. Together they represent less than 1% of the global equity market and operate across many different industries and countries. They have strong, durable competitive advantages in lucrative niche industries, and across the portfolio our average holding generates more than 80% of its earnings from non-cyclical revenue streams and has a history of almost 100 years.

What they all have in common is that we believe in the resilience of their business models, and the predictability of their highly profitable growth trajectories is rare among the many thousands of listed companies. Our aim is therefore to own a few carefully selected companies for a very long time, harvest the effect of compounding growth, and take care not to overpay.

If you have been patient enough to read this far, thank you for your time. As a token of our thanks, here are a few fun facts about AI to take with you:

This year, AI investment by the large hyperscalers (Microsoft, Alphabet, Amazon, Meta and Oracle) is expected to reach approximately USD 600 billion. That exceeds the inflation-adjusted cost of the Manhattan Project, the Apollo programme and the International Space Station (ISS) combined (~USD 450 billion) – and by a wide margin.
A one-gigawatt data centre covers the area of roughly 600 football pitches, and five data centres of that size are opening this year alone.
AI’s energy demand is almost unfathomable. The UK’s largest power station, Drax, can supply 5 million households with its 2.6 GW of capacity… equivalent to the power needs of two such data centres.
The water demand is no drop in the ocean either. A one-gigawatt data centre requires approximately 20 million litres of fresh water a day for cooling – equivalent to the daily consumption of around 100,000 households.

There is approximately 100 gigawatts of capacity in operation globally, and it is expected to double by 2030. We have the popcorn out and are enjoying it… as long as there is still power for the microwave.

Kind regards,

Steffen Jørgensen & Matthias Herskind
Portfolio Managers, Meritum Investments

You can also follow our Substack, where we regularly publish research notes and other content.

This newsletter has been prepared by Meritum Investments ApS and does not constitute investment advice or a recommendation to buy or sell securities. Assessments and forward-looking estimates are subject to uncertainty and are stated as of 16 August 2026. We own shares in RELX and may trade in the companies mentioned without notice. Past performance is no guarantee of future results, and investing involves a risk of loss.

Want to read an example of one of our investments?

Click here