Two developments currently define the artificial intelligence market. Companies are generating revenue from AI products that many users already pay for. At the same time, investment in the sector is growing faster than current revenue. Investment plans imply that future subscription prices could rise from about $200 to$20,000, with a similar number of subscribers. The gap between these two trends raises a central question: if AI is useful, why are analysts discussing a bubble?

Some technology stocks have fallen in recent news coverage, including those of Microsoft, Google, and Anthropic. The analysis notes that Anthropic has not yet held an initial public offering, so its shares are not publicly traded. The broader point stands: a technology can be real while the prices investors pay for it are excessive.

Historical Precedents

The internet did not disappear when the dot-com bubble burst, but investors lost large sums. Similar patterns appeared earlier in the railway sector.

Railways in Britain, 1840s

• Investors recognized that railways were changing trade, travel, and land values. Steam locomotives shortened journeys and opened new markets.
• Investors concluded that almost any railway company would become profitable.
• In 1846, Parliament considered more than 700 railway projects.
• The construction boom was followed by a downturn and then a gradual recovery. The railways remained in operation.

Dot-com bubble, 1990s

• Investors bought companies with "dot-com" in their names.
• Some firms went public with only a presentation and no profit or clear path to profit.
• The market turned downward in March 2000. By October 2001, it had lost about two-thirds of its peak value.
• Telecom investments also collapsed. Excess fiber-optic cable had been laid in anticipation of rapid growth in internet traffic. Traffic grew more slowly than forecast. Operators cut investment, and some companies went bankrupt. Equipment became cheaper, and some of the fiber later became part of the expanding internet infrastructure.

The analysis identifies this as the key difference between a technology bubble and a wholly invented market. Useful assets and products can survive a bubble. However, the survival of assets does not protect buyers who paid too much, built more capacity than needed, or borrowed on the assumption of very high demand.

Assessing the Bubble

The answer depends on how the term is defined.

• If the claim is that AI is useless and generates no revenue, the evidence does not support it. Companies are earning money from AI.
• If the question is whether current investment will pay off and whether current expectations for every company are justified, the answer differs.

Cloud Revenue and Capital Spending

In October 2026, Alphabet reported the following results:

• Google Cloud revenue grew 103% for the quarter and exceeded $30 billion.

Amazon reported the following:

• AWS revenue grew 39% to nearly $46 billion for the quarter.

These figures reflect actual sales of cloud computing. Clients pay for servers, data storage, models, and access to them.

Expectations for spending are also large:

• Alphabet plans capital expenditures of about $180 billion in 2026.
• Amazon expects cash capital expenditures of about $220 billion.

These plans cover more than AI. Amazon's figures include its logistics network. Expanding computing infrastructure is one of the main motives, however.

For these investments to pay off, new capacity must operate for a long time at high utilization. Clients must also pay more for the service than the cost of development and deployment.

Product convenience and infrastructure profitability are connected, but they are not the same. A model that writes an email in three seconds does not imply that the provider earned $3 from it. The decisive question is who pays for the infrastructure.

The analysis concludes that AI is already a real market, but an investment bubble is forming around its expected future. The technology works, and some products are useful. Asset prices may already reflect profits that do not yet exist, and construction plans may exceed demand.

Market Participants

Three groups drive the market, and each earns returns in a different segment.

1. Public investors buy shares of chip, cloud, and software providers. Market prices reflect expected future profits in advance. If forecasts change, share prices can fall before the underlying businesses do.
2. Venture capital funds and large private investors fund model developers. They expect the leading developer to become a platform for office software, programming, search, and customer support. Their position is either a stake in future infrastructure or an expensive stake in a company that requires continuous new funding.
3. Cloud providers build data centers, sell computing capacity, and invest in companies that then rent that capacity. Microsoft, Amazon, and Google are both participants and suppliers in this process.

Interconnected Financial Flows

The relationships among these participants show how revenue and investment connect.

Microsoft and OpenAI

• Microsoft reported receiving $24.1 billion from commercial relations with OpenAI during the fiscal year.
• At year end, OpenAI owed Microsoft $6 billion.
• Microsoft reported investment commitments to OpenAI of $13 billion, of which nearly$12 billion had been funded.

Amazon, OpenAI, and Anthropic

• Amazon invested billions of dollars in OpenAI and Anthropic.
• Amazon then signed multi-year agreements under which the two companies use AWS and Amazon's chips.

In this structure, the investor, supplier, and client can be the same or closely linked parties. The cloud provider bills the model developer, the model developer uses the computing capacity, and the investment funds the infrastructure.

Such arrangements can accelerate growth. Investors provide capital to model developers, the developers rent servers from the investors, and cloud revenue increases. The central test remains whether independent clients enter the system and generate enough revenue to cover the total computing cost.

Infrastructure builders also participate. Developers prepare sites, equipment manufacturers expand production, and operators reserve electricity. If demand is very high, companies without sufficient capacity lose opportunities. If demand slows, companies that built too much face losses.

Where Demand Is Established

The most stable part of the market exists where clients make payments for defined services. Cloud platforms already sell:

• access to models
• data processing
• code generation
• internal document search
• support automation

Businesses were paying for many of these functions before the current investment wave. Cloud platforms, office software, search engines, and developer services all include AI features. Companies have clear reasons to buy them. A model can analyze a customer support request, locate a clause in a long contract, or produce a first draft of code.

Savings must appear in cost reductions and shorter timelines to justify the purchase.

Data centers also serve databases, enterprise applications, and video services, not only chatbots. Servers and networks are becoming standard equipment of the digital economy. Some capacity can be redirected to other tasks, although not every chip or building can be repurposed easily.

Financial results support the existence of customer demand. In the third quarter of 2026, AWS generated nearly $46 billion in revenue and more than$17 billion in operating profit. Google Cloud more than doubled its revenue over the year.

Uneven Profitability Across the Value Chain

Profitability differs across the chain:

• Chip suppliers can earn revenue immediately because they sell hardware.
• Cloud providers are paid as their services are used.
• Model developers must persuade clients to renew subscriptions and pay for computing.
• Enterprise buyers must demonstrate that employees complete more work or that the company has reduced costs.

At each step, the question of revenue becomes more pressing.

How a Correction Could Unfold

A bubble does not require a sudden collapse. It can contract through routine corporate decisions, such as frozen projects, renegotiated contracts, and reduced spending. The following sequence illustrates one possible path:

1. Companies find that users test AI products but pay less than expected.
2. Some services cost more to run than they generate, and others show no measurable economic benefit.
3. Users try products but do not use them weekly.
4. Model developers reduce training and computing purchases.
5. Cloud providers receive fewer orders and slow construction.
6. Equipment manufacturers lose orders.
7. Investors lower valuations of future profits.

Young companies that depend on investment and expensive server rentals are affected first. They must sell to larger firms, reduce staff, or close. Used accelerators and spare capacity then enter the market, and computing prices fall. Owners of expensive data centers lose, while companies that could not previously afford infrastructure gain access.

A more severe scenario occurs when construction is financed by debt and long-term obligations. Losses then extend beyond startup shareholders:

• Payments for idle capacity continue.
• Creditors seize collateral.
• Contractors lose orders.
• Regions fail to reach projected employment and tax revenue.

The impact depends on where the risk is concentrated. In the dot-com crash, technology stocks and industry investments suffered the largest losses, while the internet continued to develop. The real estate bubble posed a larger threat to the broader economy because falling prices coincided with large debts and financial instruments linked to mortgages. If creditors and large banks become deeply involved, a localized correction can spread.

What Would Remain

After a downturn, not all current companies or infrastructure would survive. The analysis identifies the following likely outcomes:

• Models that solve specific problems with acceptable unit economics would remain.
• Cloud platforms would remain, because computing demand extends well beyond AI.
• Data centers with adequate power and network connectivity would remain, even if some equipment became obsolete and was sold at a discount.

The most significant consequence may be less dramatic than current marketing suggests. AI would stop being sold as a tool that eliminates half of a company's staff. It would become a feature within familiar services. It would help draft documents, locate files, process inquiries, and prepare code.

If the bubble bursts, AI as a technology will not disappear. Companies, valuations, and plans that assumed unlimited future demand will disappear. Railways survived the railway mania. The internet survived the dot-com crash. The companies that convert models and data centers into services clients pay for repeatedly, rather than once out of curiosity, will remain in the market.