Are We Entering Bubble Territory?

Artificial intelligence may transform the world—but have markets already priced decades of promise into today’s valuations?

Matthew Jones | Co-Founder & Precious Metals Analyst
Britannia Bullion | 2 September 2026

Across the world, vast industrial structures are rising from the ground. They are usually called data centres, but that description now feels inadequate. These are not simply buildings used to store photographs, emails and company records.

They are AI power plants.

They consume land, water, electricity, advanced microchips and extraordinary quantities of capital. Their promised output is not energy, but intelligence: automated decisions, digital workers, scientific discoveries, autonomous machines and perhaps, eventually, millions of physical robots.

The potential is enormous. But one uncomfortable question hangs over the entire construction boom:

Who will ultimately pay for all of it—and how much will they need to pay?

That is not an argument against artificial intelligence. AI may prove to be every bit as revolutionary as its supporters claim. But history repeatedly teaches us that a revolutionary technology and a successful investment are not necessarily the same thing.

The technology can win while investors lose.

A Multitrillion-Dollar Leap of Faith

The sums involved are becoming difficult to comprehend.

McKinsey estimates that approximately $5.2 trillion may need to be invested in AI-ready data centres by 2030. Including conventional computing infrastructure, the total rises to around $6.7 trillion.

Bain & Company calculates that supporting the projected level of AI computing could require approximately $2 trillion of new annual revenue by 2030. Even after allowing for potential savings generated by AI, Bain still identifies an annual shortfall of around $800 billion.

Venture capitalist David Cahn originally described the mismatch between infrastructure spending and end-user income as AI’s “$200 billion question”. It became a $600 billion question in 2024, an $840 billion question in 2025 and, according to his updated calculation, a $1.5 trillion question in 2026.

These estimates use different assumptions and should not be treated as precise forecasts. Yet they all point towards the same concern: investment is accelerating much faster than clearly identifiable end-customer revenue.

It is possible that demand will catch up. But it now needs to catch up on an almost unimaginable scale.

Where Is the End Customer?

AI is already generating genuine revenue. Companies pay for cloud computing, software subscriptions, coding tools, automated customer service and access to advanced models.

Amazon, for example, said in April that its AI services had reached an annualised revenue rate of more than $15 billion. It also argued that a substantial portion of future AWS capacity is supported by customer commitments. That is meaningful commercial demand.

However, Amazon simultaneously projected approximately $200 billion of capital expenditure for 2026, mainly focused on AI. Capital expenditure creates assets that can produce income for several years, so comparing one year’s spending directly with one year’s revenue is imperfect. Nevertheless, the enormous difference demonstrates how much future growth is already being assumed. Reuters

There is another important distinction: revenue generated inside the AI industry is not necessarily the same as revenue arriving from the final customer.

A chip manufacturer sells processors to a data-centre operator and records revenue. The data-centre operator rents computing capacity to an AI company and records revenue. The AI company may fund that rental using capital raised from investors. Several businesses have now reported sales—but the chain has not truly closed until an outside customer pays enough for an AI product or service to support everybody above it.

This is why questions about circular financing matter. Reuters recently reported growing investor scrutiny of arrangements in which Nvidia supported the financing of AI-cloud customers and, under one proposed model, offered to rent back capacity those customers could not sell. Such arrangements may help a young industry scale, but they can also make underlying demand more difficult to assess. Reuters

The real question is therefore not whether money is moving through the AI ecosystem. It plainly is.

The question is whether enough money is entering from outside it.

What Are They Planning to Sell Us?

The industry has several potential answers.

AI could be monetised through cloud computing, business subscriptions, digital assistants, advertising, search, medical research, defence, software development and autonomous vehicles. Companies may also justify investment through cost savings rather than new sales—particularly if AI allows them to complete more work with fewer employees.

But even these savings must eventually appear as measurable improvements in profits or productivity. Economic value created for society does not automatically become cash flow for the company that paid for the data centre.

The largest possible consumer market may eventually be physical AI: robots capable of working in factories, caring for older people, completing household tasks or performing jobs that presently require human labour.

This could explain why today’s infrastructure is being built so far ahead of visible demand. The industry may not merely be preparing to sell us software. It may be preparing for an entirely new generation of intelligent machines.

But here, too, the assumptions are considerable.

A technically capable robot is not automatically a commercially acceptable robot. Cost, reliability, safety, privacy, insurance, maintenance and regulation will all affect adoption. Consumers may require years to become comfortable allowing autonomous machines into their homes.

There may also be a deeper social obstacle. If large numbers of people come to associate AI with losing their jobs, damaging their children’s employment prospects or weakening their financial security, will they enthusiastically welcome an AI-powered robot into the family home?

Technological readiness and social acceptance do not necessarily arrive together.

If acceptance takes five years longer than anticipated, the consequences could be serious. Data-centre buildings may remain useful for decades, but the advanced processors inside them can become obsolete far more quickly. Microsoft, for example, currently estimates useful lives of approximately two to six years for computer equipment. The clock begins ticking long before the hoped-for customer arrives. Microsoft Annual Report

When Fear Becomes a Sales Tool

The marketing surrounding AI is unusually powerful because it sells hope and fear simultaneously.

We are told that AI could cure diseases, transform productivity and create extraordinary abundance. At the same time, we are warned that it could replace entire professions, become impossible to control or allow one company—or one country—to dominate the future.

Many of these warnings may be sincere. But they also create an extraordinarily effective commercial message:

Invest now, because the cost of being left behind could be existential.

That creates a corporate prisoner’s dilemma. Every individual company may rationally conclude that it cannot afford to spend less than its competitors. Yet collectively, the industry may build far more capacity than customers are ready to purchase.

Chief executives are unlikely to be criticised for joining the AI race. They may be severely criticised if they miss it. Under those conditions, the normal question—“Will this investment produce an acceptable return?”—is gradually replaced by a very different one:

“Can we afford not to participate?”

That is fertile ground for a bubble.

From Dot-Com to Dot-Bomb

The comparison with the technology boom of the late 1990s is difficult to ignore.

The internet was real. It transformed communication, entertainment, shopping, banking and almost every other area of modern life. The optimists were correct about the technology.

But many investors were disastrously wrong about the price they should pay for it.

Capital flooded into companies with exciting names, enormous ambitions and little evidence of sustainable profits. Infrastructure was built ahead of demand. Valuations rose because investors feared missing the future. Institutions funded the early expansion; public enthusiasm then provided an opportunity for early backers to realise gains.

The suffix moved rapidly from dot-com to dot-bomb.

Today’s situation is not identical. Microsoft, Amazon, Alphabet and Meta are enormously profitable businesses with real customers, substantial cash flows and established products. This is not simply a collection of newly listed companies with little more than a website and a presentation.

But strong companies can still overinvest. Real demand can still be overestimated. Exceptional businesses can still become poor investments when purchased at excessive valuations.

Most importantly, AI does not need to fail for an AI investment bubble to burst.

The internet survived its crash. Many internet companies and their investors did not. AI could follow precisely the same path: an enduring technological revolution interrupted by a brutal destruction of speculative capital.

Who Is Left Holding the Risk?

The financial structure around AI is also changing.

What began with venture capital and the retained profits of technology giants is spreading into private credit, infrastructure funds and public bond markets. In August, Nvidia announced partnerships with six major financial institutions intended to create financing platforms targeting more than $500 billion of third-party capital for AI infrastructure. Reuters

US technology companies are also becoming major issuers in European debt markets. According to Reuters, companies including Amazon, Alphabet and Microsoft now account for almost 10% of gross new corporate bond issuance in the eurozone as they raise capital for AI investment.

This matters because risk does not disappear when it leaves a venture-capital fund. It migrates.

It can migrate into corporate bonds, pension funds, infrastructure vehicles and index-tracking portfolios. Ordinary savers may acquire significant exposure to the AI investment cycle without consciously deciding to do so, simply because a small number of technology companies represent such a large proportion of major market indices.

If the optimistic forecasts prove correct, that exposure may be rewarding. If revenue arrives late or margins disappoint, the eventual adjustment could extend far beyond a handful of speculative AI start-ups.

Gold While the Dust Settles

None of this means investors must reject technology or abandon equities. It means they should recognise how much of today’s market value depends upon uncertain cash flows arriving many years into the future.

Until the economics become clearer, physical gold may offer a useful counterweight.

Gold does not have a revenue stream—and it does not pretend to. It is not valued as a growth company and does not require a chief executive to convert trillions of dollars of capital expenditure into future profits.

It has no product launch to complete, no subscription target to reach, no data centre to refinance and no consumer adoption curve to forecast. Physical gold is a tangible monetary asset whose ownership does not depend upon another company meeting its earnings projections.

That does not mean gold rises every time technology shares fall. Gold prices can decline, sometimes sharply, and holding it involves its own considerations. Its role is not to replace every growth investment or to guarantee a profit.

Its role is diversification.

For investors whose pensions, funds and share portfolios have become increasingly concentrated in the same small group of technology companies, an allocation to physical gold can place part of their wealth outside that single financial narrative.

If AI fulfils every promise, a balanced investor can still participate through equities and the broader economy. If valuations reset before the promised income arrives, gold offers an asset that does not need the AI business model to succeed.

This is not a choice between progress and the past. It is a choice not to place every part of one’s financial future on the same version of tomorrow.

The Technology May Win. The Trade May Lose.

Artificial intelligence may transform medicine, industry, education, transport and the nature of work itself. It may ultimately justify investment on a scale that currently appears extraordinary.

But “may” is carrying an enormous financial burden.

Trillions are being committed before the final revenue model is clear. Infrastructure is being constructed before many customers know what they are prepared to buy. Valuations increasingly depend not merely upon AI succeeding, but upon it succeeding quickly, profitably and on a scale rarely witnessed in commercial history.

Perhaps it will.

But if it does not—if adoption is slower, competition pushes prices lower, consumers resist or the promised productivity gains prove difficult to capture—the technology can continue advancing while the investment bubble deflates around it.

That is why the question is not whether AI is real.

It plainly is.

The question is whether the price being paid for its future has become detached from the money it can realistically earn.

AI may be the next internet—and still experience its own dot-com crash before delivering its full promise. Until the dust settles, holding part of one’s wealth in something tangible, finite and independent of those promises may prove to be the more intelligent strategy.

About the Author

Matthew Jones is Co-founder and Precious Metals Analyst at Britannia Bullion. This article represents his personal analysis and opinion. It is intended for general information and does not constitute personal financial advice or a recommendation to buy or sell any asset.

Investment in physical gold is unregulated in the UK and is not protected by the FSCS or Financial Ombudsman Service. Its value can rise or fall, and ownership, custody, insurance and storage arrangements must be properly understood.