The AI frenzy gets a reality check

THE initial euphoria over artificial intelligence has hit a speed bump. For the past three years, the market has treated every dollar spent on AI as a downpayment on the future. But over the past month, it has started asking some hard questions.

Shares in the hyperscalers – Microsoft, Alphabet, Amazon and Meta – have fallen, and the selling has spread to the chipmakers and cloud providers that supply them.

The S&P 500, which is heavily weighted towards the hyperscalers and chipmaker Nvidia, is barely above its August peak. In the month to Oct 23, the Philadelphia Semiconductor Index was down 14 per cent. The biggest listed specialist AI cloud provider, CoreWeave, has lost more than 38 per cent.

Meanwhile, Treasury bonds, the classic safe haven for investors, have rallied despite rising inflation, which normally leads them to fall.

No single piece of bad news explains this performance. Rather, investors appear to be confronting a question that the euphoria had made them defer: Can the revenue from AI ever catch up with the spending?

The sums involved are staggering. The four big hyperscalers alone are on course to spend roughly US$700 billion on capital expenditure in 2026 and close to US$1 trillion in 2027, most of it on data centres and the chips that fill them.

Lessons from past manias

In sheer scale, the AI build-out invites comparison with the great infrastructure booms of the past: the railway manias of Victorian Britain and Gilded Age America, and the fibre-optic and dot-com frenzies of the late 1990s.

Those precedents were premised on the notion that transformative technologies justify heavy spending. They are better read as a warning. Each was a vast bet on future demand, and each ended in a bust.

Britain’s railway mania collapsed in 1847, wiping out thousands of investors. In America, overbuilt lines helped trigger the panics of the late 19th century. Telecommunications carriers poured roughly US$500 billion into networks in the late 1990s, and by 2002 the Nasdaq had lost almost 80 per cent of its value.

SEE ALSO

Tech stocks tumble as AI fears collide with earnings misses

In every case, the infrastructure went on to transform the economy, but the early investors who paid for it mostly did not share in the gains.

Each boom overestimated how quickly demand would grow. The problem was not the technology but the timing.

By 2001, only a small, single-digit fraction of the fibre laid in America was actually operational. The network eventually filled up – it carries today’s internet traffic – but not before one-time telecom giants WorldCom, Global Crossing and dozens of others had collapsed.

The same mismatch is visible now. Data centre construction is following an exponential curve, with each year’s plans dwarfing the last.

But AI adoption is not exponential. It seems to advance in fits and starts: first, there is a burst of enthusiasm; then, a long plateau as organisations work out how to reorganise around deploying a new tool; and then, maybe another leap.

A widely cited Massachusetts Institute of Technology study in July 2025 found that 95 per cent of organisations saw no measurable positive return on investment after six months. It is organisations, not individuals, that will drive most AI revenue.

The revenue gap

No hyperscaler is likely to be making money on generative AI. Microsoft and Amazon, the only two to disclose AI revenue, report a combined annualised revenue of US$60 billion to US$70 billion – against around US$400 billion in capital spending between them in 2026 alone.

By some estimates, covering current and planned AI capex through identifiable AI income would require something like US$2.5 trillion a year – more than the combined revenue of the entire technology sector today.

Meanwhile, valuations across much of the AI ecosystem have run far ahead of earnings. Today’s share prices make sense only if AI delivers productivity gains far larger than anything yet visible in the data. That doesn’t mean the bet is wrong.

But investors are paying now for a transformation that may arrive much later, and patience is finite. Markets need not conclude that AI is a failure to reprice it sharply. They need only decide that the payoff is further off than they had assumed.

Money that goes in circles

What further complicates the story is the way the industry finances itself. A growing share of AI demand rests on recycled capital, vendor financing and promises of future purchases.

For example, Nvidia has invested US$30 billion in OpenAI, which will in turn fill its data centres with Nvidia chips and has guaranteed up to US$105 billion in leases on a data centre campus that will host its chips for OpenAI. Nvidia holds a stake in CoreWeave, a cloud provider that buys its chips, and has agreed to purchase capacity that CoreWeave cannot sell.

Oracle has booked a contract worth around US$300 billion to supply computing power to OpenAI, a company still losing billions of dollars a year.

In such arrangements, one company’s capital expenditure becomes another’s revenue. Reported growth can stay spectacular without proving that the ecosystem has a self-sustaining economic model. When suppliers fund customers who then buy from those same suppliers, separating genuine demand from financial engineering becomes difficult.

The question is not whether demand exists – it does. The question is whether it will eventually generate enough revenue from customers outside the AI ecosystem – the banks, retailers, hospitals and manufacturers that must ultimately pay the bills – to justify the capital already committed.

Borrowing against the future

Debt adds another layer of risk. Since 2025, bond issuance has funded a rising share of AI capital spending, as even the cash-rich hyperscalers have turned to credit markets to raise money.

Hyperscaler bond issuance soared from around US$17 billion in 2024 to US$109 billion in 2025 and US$194 billion in the first half of 2026.

Meta raised US$30 billion in a single bond sale and arranged a further US$27 billion for its Hyperion data centre in Louisiana through an off-balance-sheet vehicle with private credit provider Blue Owl Capital. Oracle, more stretched than its peers, has borrowed heavily to meet its OpenAI commitments.

Credit markets have noticed. Trading in credit default swaps (CDS) on hyperscalers and their suppliers, which insure against defaults, has surged, with Oracle’s spreads serving as the bellwether of what traders now call the AI hyperscaler default trade.

The volumes of CDS linked to seven AI-linked companies traded more than US$9 billion a week in September, up more than fourfold from the US$2 billion weekly average in July.

What worries these investors is not an outright collapse. The problem is that debt has been raised against forecasts that assume AI spending will keep accelerating, while interest and lease payments are fixed. The real test of resilience, then, is not whether the cycle can survive a collapse in spending, but whether it can survive spending that merely stops accelerating.

Chips with a short shelf life

There is reason to think this cycle could be harsher on investors than its predecessors. Railway track and fibre-optic cable stayed productive for decades; later investors who bought bankrupt networks for cents on the dollar did very well out of them.

Graphics processing units (GPUs) – the chips that power AI – offer no such consolation. New chips outperform older ones so quickly that a GPU may pay its way for only two to five years. And if data centre construction is delayed, the chips that firms buy may have to remain in storage and run the risk of becoming obsolete.

Data Center Watch counts roughly US$200 billion worth of projects blocked or delayed by local opposition in the first half of 2026 alone.

Projects have also been held back by power and building shortages. Microsoft’s CEO, Satya Nadella, warned as early as November 2025 that the company might end up with “a bunch of chips sitting in inventory that I can’t plug in”.

None of this means AI is a mirage. It will almost certainly transform economies, productivity and work over the coming decades, as the railways and the internet did before it. But current valuations and investment levels assume that transformation will happen faster, and more profitably, than the evidence supports.

The web of cross-shareholdings and circular financing leaves the industry exposed to any loss of confidence, anywhere in the chain. When reality falls short of inflated expectations, as it usually does, a painful correction looks likely. The past month may prove to be only the first instalment. THE STRAITS TIMES