An AI Bubble? Five Tests to Separate Adoption, Revenue, and Expectations

On March 31, 2026, OpenAI announced that it had closed its funding round with $122 billion in committed capital at an $852 billion post-money valuation. One month earlier, on February 27, it had announced $110 billion of new investment at a $730 billion pre-money valuation, with additional investors still expected. Those figures are not incompatible prices: $730 billion before the money plus $122 billion committed produces $852 billion after the money.

The scale feeds the question of an artificial-intelligence bubble, but a large number cannot answer it. Nor can the existence of real users or revenue. An innovation may transform the economy while attracting capital at a price that requires unreachable returns. The useful skill is translating financial headlines into five accounts: valuation, revenue, margin, investment, and concentration of risk.

First account: what the valuation means

OpenAI's preliminary announcement placed the round at a $730 billion pre-money value. Pre-money is the negotiated value of the company before new capital is added. The closing announcement raised committed capital to $122 billion and gave an $852 billion post-money figure. Post-money incorporates the new investment.

A private valuation is not cash in the bank, revenue, or a price at which anyone can sell the whole company. It comes from a transaction involving a fraction of the company and terms the public announcement does not disclose. Different classes may carry different rights, preferences, or schedules. Without the financing documents, the figure is a useful negotiation reference, not a complete map of economic value.

Committed capital must also be separated from capital already funded. OpenAI uses “committed capital”; its release does not publish the timing of every installment or all conditions. Calling the $122 billion revenue would count the wrong thing. It is investor financing that the company must convert into capacity, products, and a return.

Second account: revenue with the denominator visible

OpenAI said in March that it was generating $2 billion in revenue per month. Mechanically annualizing that month produces $24 billion; dividing $852 billion by $24 billion yields an approximate ratio of 35.5. The calculation exposes the expectation embedded in the price, but it is not an audited multiple: one month may not represent a year, “revenue” does not necessarily mean recurring revenue, and the valuation includes the new capital.

The same release lists 900 million weekly active users, 50 million consumer subscribers, and nine million paying business users. These are company-reported operating metrics, not public financial statements with notes and an audit. A weekly user, a subscriber, and a business seat are not interchangeable units. Estimating monetization requires knowing which group pays, how much, for how long, and at what service cost.

The claim that OpenAI “remains unprofitable” does not appear in those announcements. The company discloses a revenue figure, but not net income, cash flow, or a cost breakdown. Missing information proves neither profit nor loss. With the primary sources located here, margin and cash consumption cannot be calculated.

Third account: margin before celebrating demand

A company can multiply sales while destroying value if every additional unit costs too much. Generative AI must pay for inference, training, labor, distribution, storage, and depreciation or rent on data centers. The relevant measure is not only how many tokens are served, but the margin left after producing a useful answer.

Public accounts from adjacent companies show the tension without revealing OpenAI's economics. In its third fiscal quarter of 2026, Microsoft reported $82.9 billion in revenue, $38.4 billion in operating income, and $31.8 billion in net income. It also said its AI business had surpassed a $37 billion annual revenue run rate, up 123% year over year.

Annual revenue run rate is not annual revenue already recognized; it projects a recent pace. Microsoft does not disclose a standalone profit for its AI business in that release. On the earnings call, it added that Microsoft Cloud gross margin declined year over year because of continued AI infrastructure investment, partially offset by efficiency. Growth and margin pressure can coexist.

Fourth account: investment, useful life, and utilization

Building capacity before it is needed can create an advantage or a surplus. Microsoft reported $31.9 billion of capital expenditure in the quarter; about two thirds covered short-lived assets, primarily GPUs and CPUs, while the remainder covered long-lived assets. The distinction matters because a chip that ages quickly must earn back its cost faster than a building used for years.

The test follows four variables by cohort: capital placed in service, utilization rate, revenue generated, and margin after energy and operations. A data-center reservation may look like demand while it is still construction. A multiyear contract can improve visibility, but analysts must ask how much is cancellable, when it is recognized, and which customer concentrates the obligation.

The supplier side confirms that spending exists. NVIDIA closed fiscal 2026 with $215.9 billion in revenue, up 65%, including $193.7 billion from Data Center, up 68%. Its full-year GAAP gross margin was 71.1%. Those are infrastructure-provider sales and margin, not a measurement of the return every customer will earn.

During an investment rush, a toolmaker can prosper even when some buyers never recover their spending. NVIDIA's revenue is evidence of demand for chips and capacity, but it cannot validate every laboratory, application, or data-center valuation built on top of them.

Fifth account: who pays whom and who bears the risk

OpenAI's round combines overlapping roles. Its announcement attributes $50 billion to Amazon, $30 billion to NVIDIA, and $30 billion to SoftBank; it also describes infrastructure partnerships with Amazon and NVIDIA. NVIDIA contributes capital while supplying systems. Amazon contributes capital, distribution, and compute. The network may align incentives and accelerate deployments.

It also requires restraint. A supplier investing in a customer does not prove that every purchase is artificial or that money automatically returns to the investor. Establishing “circular financing” would require contracts, cash flows, purchase obligations, and accounting treatment that the releases do not disclose. The overlap in roles is verified; its precise economic effect is not.

A useful map draws capital, orders, delivered capacity, and end-customer demand as separate flows. It then asks how much revenue comes from independent customers, what percentage depends on a few buyers, and what happens when one data-center project slips. Risk does not vanish because large companies share it; it may remain concentrated in the same chain.

What the dot-com comparison can teach

In May 2000, the Federal Reserve Bank of San Francisco noted that technology companies represented 36% of market capitalization but less than 10% of assets, employment, and sales among the public companies it studied. The analysis did not say the Internet was useless. It asked what earnings growth would be needed to justify unusually high prices.

That is the valid analogy. A financial correction does not require the technology to fail. It may occur because profit arrives late, accrues to fewer companies, or falls short of the price paid. Important differences remain: OpenAI is private, infrastructure is producing present sales, and major providers have profitable businesses. Comparison requires aligned periods, populations, and metrics, not a historical multiple pasted onto current investment.

A dashboard that can be repeated

Update five rows every quarter: valuation and new capital; recognized revenue rather than only a run rate; margin and cash flow; capacity placed in service and utilization; customer and supplier concentration. Beside every figure, record the date, source, unit, and whether it is measured, annualized, or promised.

A warning sign is valuation outrunning revenue for several periods while margin does not improve. Another is capacity construction without disclosed utilization or non-cancellable commitments. A favorable signal is customer cohorts renewing, cost per task declining, and operating cash funding a growing share of investment. No row alone decides “bubble.”

The transferable skill is not asking whether AI is real, but what future cash flow today's price requires and what evidence supports it. OpenAI's $852 billion value expresses an enormous expectation; its reported revenue, Microsoft's growth, and NVIDIA's sales demonstrate real economic activity. Between them sits the decisive calculation: how much sustainable value the customer receives per dollar of capital before the assets age.