If the AI Bubble Bursts, the Price Tag Breaks First

The first thing to break: assumptions

In a true AI sell-off, the first test would not be whether models can write code, summarize documents or answer questions. The first test would be whether the market’s assumptions were too generous.

Investors would ask simpler, tougher questions:

• Are customers paying repeatedly, or just experimenting?
• Can AI services cover GPU, data-center and power costs and still produce attractive margins?
• Do today’s share prices and private valuations make sense against actual growth?
• Is infrastructure spending tied to confirmed demand?
• Is revenue coming from end customers, or is too much of it circulating inside the AI ecosystem itself ?

That is why an AI bubble bursting would most likely be a repricing event. Equity premiums, startup valuations, data-center economics and enterprise AI budgets would all be forced through a stricter filter.

Why the bubble talk has got louder

The concern is not that investment is too small. It is that it has grown so quickly.

Goldman Sachs says AI-driven data-center and infrastructure buildout is likely to total multiple trillions of dollars. It notes that Nvidia CEO Jensen Huang has said AI infrastructure spending could reach $3 trillion to $4 trillion by 2030, while Goldman analysts project hyperscaler capital expenditure alone at $1.4 trillion in 2025-2027 .

The 2026 numbers are already large. Goldman Sachs says Wall Street consensus for 2026 capital spending by hyperscaler AI companies has risen to $527 billion, up from $465 billion at the start of the third-quarter earnings season . A separate Goldman Sachs article says analysts expect the largest hyperscale cloud companies to spend more than half a trillion dollars on capital expenditures in 2026 .

Big spending is not, by itself, proof of a bubble. Goldman Sachs also notes that while the AI investment buildout is larger than past cycles in nominal dollar terms, it looks more benign when properly scaled . The key question is not simply how much is being spent. It is whether enough real demand and profit will appear to justify that spending.

1. Public markets: the AI premium comes under pressure first

Stock markets usually move before the real economy does. In a rising market, semiconductor companies, cloud platforms, data-center operators, power infrastructure providers and software firms can all be grouped under the same AI trade. In a correction, that broad label becomes less useful.

Goldman Sachs says investors are already becoming more selective about AI stocks . In a sharper repricing, markets would focus on company-by-company evidence: how much AI revenue is real, whether margins are improving, whether customers renew, and whether infrastructure costs can be absorbed without damaging profitability.

Market concentration raises the stakes. According to Goldman Sachs Research, the seven biggest technology companies account for more than 30% of the S&P 500’s market capitalization and roughly one quarter of the index’s earnings . If AI expectations are deeply embedded in those companies’ valuation multiples, an AI repricing could show up not only in speculative names but also in major market indexes.

2. Startups: demos matter less than repeat revenue

The available evidence does not quantify how many AI startups would face down rounds, mergers or closures in a downturn. But if public AI multiples fall, private-market fundraising logic would almost certainly become tougher.

Cresset says the AI sector shows signs associated with bubbles, including high valuations, heavy capital inflows and speculative behavior . Goldman Sachs also identifies the rise in valuations of AI-exposed companies as one reason bubble concerns have intensified .

In that environment, investors would be less impressed by a polished demo and more focused on repeat revenue, customer retention, cost structure, proprietary data, distribution and integration with existing business systems. The weakest companies would be those valued mainly because they carry an AI label, not because customers keep paying for a durable product.

A thin feature built on top of a general-purpose model may not be enough. In a correction, the question becomes: does usage grow without losses growing faster?

3. Data centers, GPUs and power: from build-first to demand-tested

The clearest real-world channel for an AI correction may be infrastructure. Data centers, chips and power contracts are where AI expectations become physical capital.

Goldman Sachs says AI data-center and infrastructure buildout could reach multiple trillions of dollars, with hyperscaler capex alone projected at $1.4 trillion over 2025-2027 . The 2026 consensus estimate for hyperscaler AI company capital spending is $527 billion .

The issue is not whether every data-center project is wasteful. The issue is whether the assumed demand arrives fast enough. If AI service revenue grows more slowly than expected, or if inference costs remain too high relative to what customers will pay, companies could move from “build first and wait for demand” to “build only against clearer demand.”

That would put more scrutiny on new data-center starts, GPU procurement schedules and power-supply plans. In a cooler market, investors would care less about headline capacity and more about utilization, long-term customer contracts, payback periods and reliable electricity access.

Cresset describes the next phase as a critical test of whether today’s AI infrastructure buildout becomes a platform for lasting innovation or one of the largest capital misallocations in market history .

4. Big Tech: not necessarily collapse, but lower multiples and slower spending

An AI correction does not automatically mean the largest technology companies fail. That is an important difference from the most extreme bubble scenarios.

Cresset argues that, despite signs of excess, strong profits, steady revenue growth and cash-funded infrastructure investment make a selective correction more likely than a systemic collapse . In other words, many of the companies spending heavily on AI have real businesses and significant cash flow.

But “survives” is not the same as “avoids a share-price shock.” Because the largest technology companies now represent such a large share of the S&P 500’s value and earnings, even a reduction in AI optimism could increase index volatility . The more realistic adjustment may be lower valuation multiples, slower growth in AI capital spending and a tougher ranking of which AI projects deserve funding.

5. The most fragile structure: a circular AI economy

Goldman Sachs points to the increasing circularity of the AI ecosystem as one reason bubble concerns are back . The worry is straightforward: is the money ultimately coming from outside customers with real needs, or is too much of it moving among AI companies, cloud providers, chip suppliers and investors inside the same ecosystem?

Some circularity is normal in a young technology market. Infrastructure providers, model developers and application companies often grow together. But in a downturn, markets become less forgiving. Revenue generated inside the ecosystem will be judged differently from revenue paid by end customers who use AI to solve an operational problem.

Enterprise AI will not stop — it will face a harder budget test

A bubble bursting would not necessarily stop companies from adopting AI. It would change the approval process.

The question would shift from “Do we have an AI strategy?” to “Does this AI system reduce costs, increase revenue or integrate into work that employees already do?” Projects with vague return on investment — such as showpiece chatbots, low-usage pilots or tools that never leave the experiment stage — would be more vulnerable.

Workflows with clearer measurement may fare better: customer-support automation, document processing, coding assistance, enterprise search and security operations. That is the practical version of the line Cresset draws between infrastructure that enables lasting innovation and spending that turns into capital misallocation .

What would hold up — and what would look exposed

The key test is not whether a company “does AI.” It is whether AI creates repeatable cash flow.

Bottom line: AI may survive while the AI price tag changes

If the AI bubble bursts, the central conclusion may not be that AI was useless. It may be that too many assets were priced as if AI growth, margins and infrastructure demand would all arrive quickly and smoothly.

Goldman Sachs says bubble concerns stem from valuations, massive investment and ecosystem circularity . The IMF warning reported by Al Jazeera shows why comparisons with the dot-com period have become harder to dismiss . At the same time, Goldman Sachs and Cresset do not present the current moment as an inevitable, system-wide collapse .

The most realistic scenario is a sorting process. Inflated expectations, weak business models and poorly justified infrastructure spending get marked down. AI tools that generate revenue, cut costs and keep customers paying remain standing — and may look more valuable once the hype clears.