The AI gold rush, two years on: have our beliefs held true?

Authors: Ian Quigley, Head of Investment Strategy

Two years ago, we described AI as a gold rush. The analogy was deliberately provocative: AI represents real value, but who that value accrues to is not necessarily intuitive. The prospectors of 1849 endured enormous hardship for disappointing returns. The merchants who sold them picks, shovels and hard-wearing trousers did rather better.

Much has happened since – SpaceX’s record-breaking flotation, near-trillion-dollar valuations for companies yet to make a profit, a public plea from the industry’s leading figures to slow down, and, most recently, a single consumer app knocking billions off the value of banks, insurers and travel websites in a matter of days.

What strikes us, reviewing it all, is how little the underlying framework has needed to change. We made three arguments then. The advantage is assets and competences, not AI; competition hands the gains to customers, not shareholders; and the companies supplying AI get paid while the spending continues.

Idea one: the advantage comes from beyond AI

For most technology-led companies, the durable competitive advantage was not the AI itself but the way that it unlocks other assets and competences. Two years ago, that meant companies that know your search and purchase history, hold bespoke data sets or patents. The AI was the engine; intellectual capital was the fuel, and the fuel was the scarce part.

The market is now trying to choose which businesses have a genuinely defensible capability and those which are, in the end, mere data processors. That has proved most consequential for software companies charging a subscription to process a client’s own data and return it in a different form – workflow tools like Monday.com, for instance. AI can write code, so the barriers to entry for this market have collapsed, and their profits seem likely to follow.

Not every data-reliant business is equally exposed. RELX, perceived to be under threat from AI startups like Harvey AI, and later by the prospect of a legal research plugin for Anthropic’s language model, tells a different story. In practice, RELX has been partnering with Anthropic – AI is the tool that leverages its massive, proprietary, legally protected and hard-to-replicate data estate.

Where data is not proprietary, is where the shake-out continues. Meta’s consumer assistant, Muse, topped the U.S. app store within days of launch. Investors assumed an AI ‘agent’ – software that does things on your behalf – would strip value from businesses relying on customers not shopping around. Amazon’s response was instructive: it simply blocked Muse from its platform.

Two things stand out. This disruption came not from an AI laboratory but from an established technology giant, hinting that power is shifting from building models towards owning the customer. And the battle lines are drawn: firms that own the inventory, the customer relationship and the ability to transact have agency. A clever interface sitting on top of someone else’s business does not.

Idea two: competition hands the gains to customers, not shareholders

Our second argument was that where every firm in a competitive industry can adopt the same tools, the savings get competed away as lower prices. Splendid for customers; less so for investors.

What we could not know then was how forcefully that logic would apply to the AI builders themselves.

The industry has divided. The so-called “frontier” laboratories are locked in an arms race, each chasing an unassailable lead. Meanwhile ‘open-weight’ models – those whose inner workings are published, so anyone can run and adapt them – have stayed close behind at a fraction of the cost.

Game theory describes a ‘prisoner’s dilemma’ in which two suspects, unable to collude, end up likely to turn each other in to earn a reduced sentence. That’s the risk facing the AI labs. If none dares stop spending, for fear of being stranded between the best models and the cheapest, they risk pricing themselves into a smaller corner of the market which is price insensitive for the marginal gains that come from having the absolute best model.

We’ve cited the precedent of solar energy before: a genuinely transformative technology that produced an enormous bubble and decades of low returns because competition made high profit margins hard to sustain. The parallel is pointed: China’s determination to lead in solar destroyed Western profitability, and Chinese open-weight models are now snapping at the heels of the AI frontier. Transformative and profitable are not synonyms.

Idea three: picks and shovels – while the digging continues

Semiconductors were the obvious modern equivalent of Levi Strauss’s denim, and that call has been vindicated handsomely. What has changed is the breadth of it: data centres need extraordinary quantities of electricity, cooling, backup power and electrical components, to the point where companies that once supplied the solar industry now serve AI construction sites.

Picks and shovels only pay while the prospectors keep digging, and some are burning cash heavily – SpaceX reportedly lost close to $5 billion in 2025, and meaningful profitability for the AI laboratories remains conjecture. Being a step removed from the gold face is safer, but not the same as being safe.

As we noted two years ago, semiconductors are cyclical. Encouragingly, despite very strong share price gains, valuations already assume a marked decline in profits as the cycle matures – even while demand for AI capacity stays strong as ever. Low expectations are a friend to the investor; it is euphoric ones that cause damage.

How the market has developed

First, scale. Information technology now accounts for a record 5% of U.S. economic output, and business investment has accelerated all year. U.S. growth has been below trend for three quarters, propped up by AI investment. An economy this reliant on one spending category is sensitive to change, but the good news is that things seem to be broadening out.

Second, public markets have opened up. SpaceX’s flotation raised over $75 billion at a $1.75 trillion valuation. Anthropic and OpenAI both prepared listings at or approaching trillion-dollar valuations.

This warrants caution, because flotations are widely misunderstood. Buying shares in an established listed company means buying from another investor, under strict rules designed to give everyone the same information. Buying at a flotation means buying from founders, insiders and early backers – those who know the business best and have chosen this moment to sell.

The record reflects that asymmetry. Across thousands of U.S. listings over several decades, the average cumulative three-year return has been around 19%, while the wider market returned more than 20 percentage points more. Nearly 60% left investors with a loss after three years.¹

OpenAI has decided to delay its listing. That reflects the general sense of caution in the market, and also removes what could have been a large draw on investors’ capital (participation in IPOs is often funded by selling existing holdings).

Third, the industry began arguing for restraint. Anthropic CEO, Dario Amodei published an essay, We must pace the Frontier, calling for deliberately slowing (not stopping) development so that safety can catch up. This was prompted partly by an episode in which AI agents set about achieving their task by fooling their own evaluation systems and systematically hacked a third party (Hugging Face).

Sam Altman and Elon Musk agreed, inviting suspicion that rivals agreeing on the need for regulation are agreeing on barriers newcomers cannot afford to clear. In the dilemma described earlier, prisoners who agree a story between them get the best outcome.

But two things can be true at once. The safety concern is probably genuine as well as commercially convenient. For example, only recently OpenAI shelved its October launch of its GPT-6.1 Astra model, after it failed its own internal safety tests. Either way, we see more positives than negatives if investment moderates.

Currently, capital has been herding to frontier model training where the marginal benefits look modest. A slowdown would enable investment to fund other sectors, or go from training models to using them. Given the computing power shortage, a change in the mix looks likelier than a collapse in demand.

The current controversy

Muse triggered a wave of ‘consumer inertia’ trades, with speculators bundling banks, insurers and subscription businesses as AI losers, assuming that AI agents will shop around on customers’ behalf and shred their margins.

It’s another conclusion that needs more nuance. An agent cancelling subscriptions users didn’t know they had is entirely plausible. An agent moving savings to the best rate runs into both human and regulatory discomfort. UK insurers already operate in markets where price comparison is the norm. But there is little evidence yet that consumers will trust software to assess their circumstances, move beyond gathering quotes and choose a counterparty on their behalf.

Online travel platforms were caught in the same net, which is odder still, since they do not monetise apathy. Their strength lies in a vast network of hotel supply and their role in actually processing the transaction – neither of which a front-end interface can easily replicate. Consumers enjoy browsing for holidays.

What matters is still what matters

Two years, one record flotation and several panics later, our conclusion is where we began. The gold was real, but shared among many prospectors (fewer than 5% did).² They all needed picks and shovels, though.

Most of AI’s economic value won’t accrue to the headline-grabbing frontier models. It will go to countless businesses tidying up their data and automating unglamorous, repetitive tasks with today’s capability. They will mostly pass it on to customers as lower prices. That is why we remain deliberately measured, favouring quality and durable advantage over speculative bets on whichever bottleneck is fashionable this quarter.

The investors best placed to thrive are those who can admire a technological miracle while calmly asking: yes, but at what price? Transformative change and rewarding investment are not the same thing – and perspective is the investor’s most powerful tool.

If you’d like to talk through what these themes means for your portfolio, your wealth manager is well placed to help.

---

¹Initial Public Offerings: Updated Long-run Statistics, Jay R. Ritter & Eugene F. Brigham, August 2026

²Bureau of Economic Analysis

---

About the author

Ian Quigley

Head of Investment Strategy

Ian is Head of Investment Strategy for RBC Brewin Dolphin Ireland. Ian is a member of RBC Brewin Dolphin Ireland’s Executive Committee and sits on RBC Brewin Dolphin’s Investment Committee.

Ian has a First Class Degree in Economics from Trinity College and a Master’s in Finance from the Smurfit Business School. Ian has worked in Wealth Management since 2004, having previously worked for Investec Wealth & Investment and NCB. Ian has led the Investment Strategy team for over a decade and believes strongly in supporting the collaborative investment culture we have in the firm.

---

---