On 5 December 1996, Alan Greenspan asked how anyone could know when “irrational exuberance” had pushed asset prices too far. Wall Street briefly shuddered, then resumed climbing. The S&P 500 went on to gain about 105% before its March 2000 peak. The Nasdaq Composite did even better, rising from 1,300 to 5,049, or roughly 288%. Greenspan had identified the mood with eerie precision and offered almost no help with the timing.

The market even staged what looked like a finale along the way. Russia defaulted in 1998, Long-Term Capital Management came close to collapse, and the hedge fund’s $2.3 billion capital base was supporting about $107 billion of recorded positions, leverage of more than 50 times. The Nasdaq fell from 2,049 on 20 July to 1,419 on 8 October, a drop of nearly 31%. That was a real bear market, accompanied by a real credit accident and a real rescue. It still was not the end. From that October low, the Nasdaq rose another 256% in only seventeen months.

The crash had come for a dress rehearsal, then returned to the wings. A warning can be correct, a correction can be violent, and the bubble can still have another lifetime ahead.

A bubble has no statutory retirement age

The comparative Macrobond chart is seductive. Japan, housing, commodities, dot-com, crypto and AI are placed on the same starting line, which makes every mania look as though it follows a familiar itinerary: two or three breathless years, a summit, then gravity. The picture is useful for describing tempo. But it's hardly a countdown clock. 

Each line begins where an analyst, looking backwards, has decided the story began. Hindsight enjoys that luxury, but investors living through the episode unfortunately don't. Rebase enough cycles to 100, choose a convenient start date, and the curves will rhyme. The chart shows what happened after an event was labelled a bubble. It does not tell us the probability that today’s line will end at the same point.

The current AI cycle illustrates the problem. If the public launch of ChatGPT on 30 November 2022 is used as the starting gun, the boom is about three years and seven months old, not five. Start with Nvidia’s first post-ChatGPT earnings shock, the hyperscaler capital-expenditure acceleration, or the HBM shortage, and the answer changes again.

A stopwatch is therefore the flimsiest tool in the kit. Valuation, financing, capacity, customer economics and the identity of the marginal buyer are harder to assess, but there is no shortcut. Bubbles end when a detail changes the equation: funding disappears, demand fails to arrive, collateral stops supporting debt, or a supposedly scarce product becomes abundant.

Are we better equipped today?

Unfortunately, more analysts, faster data and better computing do not abolish that problem. A bubble is rarely caused by a simple shortage of information. It can survive in full view because incentives remain aligned with continuation. Portfolio managers fear leaving a benchmark winner too early. Suppliers keep expanding while orders are firm. Hyperscalers keep spending because the first company to economise may hand the frontier to a rival. The market can know that a structure is fragile and keep financing it anyway.

What better surveillance may change is the shape of the ending. Vulnerabilities can appear first as credit-spread widening, violent rotations, earnings resets and 20% or 30% corrections rather than one cinematic crash. The 1998 episode already showed that a market can correct before it breaks, then make the final excess larger.

The market is choosing the cash register

The latest market action is not a broad vote for anything carrying an AI label. It is a rush toward the businesses that can invoice the buildout now. SK Hynix’s July 2026 Nasdaq offering raised $26.5 billion, was more than seven times subscribed and followed a roughly 680% rise in the Korean shares over the previous year. The valuation multiple had fallen even as the stock soared because earnings rose faster than the price.

Goldman Sachs tells the same story with baskets rather than one company. Its AI-infrastructure basket had gained 74% in 2026 by late June, beating the equal-weight S&P 500 by 65%. The prospective productivity beneficiaries had not enjoyed anything comparable. The word “technology” is doing too much work in many sector charts. Semiconductors, networking equipment, cloud platforms and data-centre suppliers are producing the aggregate profit surge. Parts of application software are simultaneously facing price pressure, faster coding and the risk that AI agents consume features that used to command a subscription.

For the time being, Wall Street is choosing the cash register, not the promise.

Follow the free cash flow, then ask how long it lasts

The Bank of America chart below is abruptly revealing. Twelve-month forward free cash flow for the semiconductor group climbs toward $400 billion. The line for Amazon, Alphabet, Meta, Microsoft and Oracle collapses toward zero and, in the chart’s final estimate, below it. A separate Bloomberg consensus series is less extreme but points in the same direction, with combined hyperscaler free cash flow falling from about $395 billion in 2024 to roughly $50 billion in 2026. The five largest spenders are expected to commit more than $1 trillion to AI-related capital expenditure across 2025 and 2026.

Figure 2. A forecast transfer of free cash flow from hyperscalers to semiconductor companies. The precise endpoint varies by estimate, but the direction is the same. Source: BofA Global Research.

This does not mean that hundreds of billions have vanished. FCF is measured after capital expenditure. The hyperscalers are exchanging cash for chips, servers, networking, buildings and electricity infrastructure. The accounting question is easy. The economic question is more demanding: will those assets earn enough, for long enough, to beat the return investors could have obtained elsewhere?

Suppose, very roughly, that $1 trillion of investment carries a 10% required return. It must eventually generate around $100 billion of annual after-tax cash flow to clear that hurdle. An 8% return would still produce an impressive $80 billion, yet leave a $20 billion annual shortfall relative to the opportunity cost. Positive earnings are not the same thing as value creation. The dosage makes the poison.

The chart reveals who is being paid today and who is waiting to be paid tomorrow.

Semiconductor manufacturers and infrastructure suppliers recognise revenue while the factories and data centres are being built. Their customers must operate the assets, absorb depreciation, secure power, keep utilisation high and sell enough AI services to justify the bill. For the time being, hyperscaler capital expenditure has become semiconductor and infrastructure revenue. The duration of that transfer will decide whether today’s windfall becomes a durable FCF franchise or a magnificent one-off order cycle.

This is where market momentum can disguise a shrinking opportunity. A business that goes from $2 million of annual demand to $200 million has experienced something extraordinary. An investor buying after the market has already priced $200 million, plus flawless growth beyond it, owns a different proposition. The operating miracle may be real while the remaining stock-market return is mostly a contest over who will pay a higher multiple.

HBM is today’s answer, not a universal law of physics

High-bandwidth memory sits near the sweet spot of the present cycle. Micron estimated the HBM market at about $35 billion in 2025 and projected roughly $100 billion by 2028, a milestone it expected two years earlier than in its previous forecast. That is the sort of demand shock that can remake an industry.

But it can also tempt investors to treat the current architecture as permanent... Memory history advises a little humility. Intel and Micron unveiled 3D XPoint in 2015 as a new class of memory between DRAM and NAND, combining speed, persistence and density in a way that seemed capable of redrawing the hierarchy. Then Micron stopped development in 2021 after concluding that the market had not validated the product at sufficient scale. Intel wound down Optane the following year and lost $559 million.

The technology was not foolish, but the economics were. Workloads changed, software support lagged, manufacturing remained difficult and the ecosystem never reached the necessary scale. HBM could remain dominant for years, but inference architectures, custom ASICs, CXL-connected memory... until the next piece of silicon wizardry with an incomprehensible acronym reshuffles the memory mix. At some point, the most useful person in an AI equity meeting may be the memory architect rather than the spreadsheet virtuoso.

Scientific progress does not wait for a DCF model to finish recalculating.

Where are the customers’ profits?

Infrastructure must become productivity, revenue and earnings for the companies using it. Goldman Sachs found that 54% of S&P 500 companies discussed AI in the context of productivity or efficiency during first-quarter 2026 calls. Only 11% quantified an effect on a specific use case. Just 2% connected the productivity gain to earnings. MIT researchers cited in the same report estimated that only 10% to 20% of S&P 500 companies were using AI in ways that generated revenue or a meaningful economic benefit.

The FT chart sharpens the puzzle. Earnings growth in the US technology sector has accelerated, while the sectors expected to use AI heavily, including finance, manufacturing, transport, media and healthcare, have not produced a similar wave.

Figure 3. Earnings growth has surged in the US technology sector, while the expected heavy users of AI have not yet shown the same acceleration. Source: Financial Times, Absolute Strategy Research, Worldscope and LSEG.

That does not mean software as a whole is prospering. The red line is an aggregate dominated by a handful of semiconductor, cloud and platform profit pools. Microsoft and Oracle can sell compute, databases and AI tools while also using AI to reduce internal costs. Nvidia and Broadcom sell the hardware and custom silicon. Meta and Alphabet can improve advertising systems while funding enormous infrastructure programmes. Other software companies face a rougher bargain, as customers ask why a coding agent or model cannot replace a seat, a feature or an entire layer of the application stack.

On one hand, AI can enlarge the profits of the companies supplying compute while cannibalising parts of the old software pool. On the other, it can also lift output per employee before the gain appears in reported margins, especially when companies reinvest the savings in models, engineers and more compute. While the second wave is a logical consequence, it is still unproven at scale.

The BIS is watching the financing

Greenspan asked whether prices had become irrational. The Bank for International Settlements is asking what happens when internal cash generation no longer covers the investment.

The funding model is moving from self-financed expansion toward bonds, leases, project structures, private credit and off-balance-sheet vehicles. Direct-lending funds have quadrupled their exposure to AI and information technology over five years, to roughly 15% of portfolios. A note by Syz Group also mentions that credit-default-swap spreads for several AI-linked companies have widened even while their equities continue to price substantial upside.

The macro footprint has become large enough to deserve the attention. The BIS estimates that total US IT investment has reached around 5% of GDP, above the dot-com peak. Semiconductor facilities and data centres have contributed about 0.4% to annual GDP growth on average since 2022. Broader IT investment has accounted for almost half of US growth in some recent quarters.

AI Capex is no longer merely an outcome of the expansion. It is one of the expansion’s engines. The same spending that supports growth can leave a hole when it slows.

Railways were real. And so were the losses.

The BIS itself invokes canals, railways, electrification and the dot-com buildout. A genuine breakthrough can attract more capital than its eventual cash flows justify. Railways changed commerce. Telecommunications changed society. The internet survived 2000. None of those facts guaranteed an adequate return for every investor who financed the buildout.

What’s more, competition intensifies the risk. Each hyperscaler can rationally overspend from its own point of view. Underinvestment might leave it dependent on a rival’s model, cloud or chip. The industry can therefore produce an irrational total from individually rational decisions. The BIS calls this the “contest motive” and warns that it can drive the sector’s aggregate economic surplus below zero under adverse assumptions.

No CEO wants to be the first to cut the budget and discover that everyone else was right. That corporate FOMO, coupled with sunk-cost bias, can keep the boom alive even after projected returns have begun to sag.

Three clocks, one dashboard

One way to read the cycle is through three clocks. The cash-flow clock measures how long hyperscalers can fund the buildout without weakening their core businesses or accepting returns below their hurdle rates. The credit clock measures how long lenders will finance assets exposed to customer concentration, fast obsolescence and uncertain utilisation. The adoption clock measures how quickly companies outside the technology sector can convert AI into sales, lower costs and wider margins.

The first clock still has room to roam because the largest spenders entered the cycle with formidable balance sheets. The second has started to tick more audibly as debt and private structures expand. The third remains the slowest. Investors have capitalised years of prospective productivity, while only a small minority of companies can attach a clear earnings number to it.

Consensus estimates cited by Goldman put hyperscaler capital expenditure at $920 billion in 2027, with growth slowing from 84% in 2026 to 22% in 2027. Goldman thinks those estimates may be too conservative. Either way, the arithmetic contains a trap: $920 billion can be an eye-watering amount of spending and still disappoint a market that had priced an even steeper ascent.

The turn may therefore begin without a collapse in absolute investment. A slower growth rate, shorter lead times, lower HBM prices, weaker data-centre utilisation or a higher funding spread can change the valuation equation before the factories stop running.

The 46-stock problem

Hendrik Bessembinder’s century of US stock-market data adds a final complication. Of 29,754 common stocks traded from 1926 through 2025, just 46 produced half of the market’s $91 trillion in net wealth creation relative to Treasury bills. Nearly 60% destroyed wealth relative to one-month bills, and the median stock returned minus 6.9% over its lifetime.

Nvidia tops the table of companies with more than twenty years of data, with a 37.04% annualised return from January 1999 through December 2025. The ranking is not simply a 2025 invention. In Bessembinder’s earlier study ending in December 2023, Nvidia was already number one at 33.38% a year. Yet the endpoint still transforms the picture. One dollar had grown to about $1,316 by the end of 2023 and to $4,959 only two years later. The lifetime wealth multiple almost quadrupled at the end of a 27-year sample.

Figure 4. Highest annualised returns among US common stocks with more than twenty years of data, through December 2025. Source: Hendrik Bessembinder, CRSP.

The table is a photograph, not a timeless podium. It suffers from the natural seduction of hindsight, which lets the winner explain its own inevitability after the fact. Plenum Publishing, Logicon, Keurig Green Mountain and Express Scripts sit beside names that now look obvious. They were not obvious before compounding did the advertising.

That leaves an old argument with fresh teeth. Jack Bogle’s answer was to buy the haystack rather than hunt for the needle. Warren Buffett has argued that excessive diversification protects investors who lack knowledge. Both ideas can be true. Concentration is powerful when the edge is real, the valuation is tolerable and the holder can survive years of looking wrong. Diversification is powerful when confidence is merely well dressed.

The market, inconveniently, does not issue certificates telling investors which condition applies.

The bill attached to the story

The AI boom can continue because supplier earnings are still rising, strategic competition keeps budgets elevated, and the companies financing the buildout remain unusually strong. It can also become more fragile without AI “failing”. A useful technology, a growing end market and a poor investment return are perfectly compatible.

Cash has already completed the first leg of the journey, from hyperscalers to semiconductor and infrastructure suppliers. The second leg, from infrastructure to customer productivity, is under way but uneven. The final leg must bring cash back to the original spenders at a rate above their cost of capital, before technological obsolescence eats too much of the asset base.

Irrational exuberance is therefore neither a verdict nor a timing signal. It is a question with a bill attached. The cycle will not turn when a calendar announces that the bubble is four years old. It will turn when tomorrow’s cash flow no longer arrives quickly enough, lasts long enough, or belongs to the companies investors thought they had bought.

Sources:

Greenspan speech, FRED, Yahoo Finance, OpenAI, Reuters, Bank for International Settlements, Micron investor materials, Goldman Sachs, Syz Group, Hendrik Bessembinder (“One Hundred Years in the U.S. Stock Markets”, 2026; Hendrik Bessembinder, “Which U.S. Stocks Generated the Highest Long-Term Returns?”, 2024)