Generated by Codex with GPT 5.6 Sol High

Artificial intelligence is attracting capital on the scale of a mature industry while still being measured like an early-stage startup. Private model developers publish incompatible revenue run rates. Public technology companies make vast commitments outside their conventional debt. Investors struggle to tell whether AI investment is earning an adequate return. Even national statistics have trouble recording the technology’s costs and benefits.

The danger is not merely untidy accounting. When markets cannot compare like with like, a rumour about revenue or a fashionable adjusted metric can jolt technology shares. The AI boom may ultimately justify today’s spending, but unreliable numbers make it harder to tell productive investment from financial excess.

Private-company revenue is a moving target

The immediate confusion concerns OpenAI and Anthropic. In the space of a week, reports put OpenAI’s revenue near both \$50bn and \$70bn. Neither figure necessarily describes sales recognised under ordinary accounting rules. Both companies rely on versions of annual recurring revenue, or ARR, which typically annualise a recent month and assume that customers keep spending at the same pace.

That convention is already optimistic for fast-changing usage businesses. Comparability becomes worse when companies define the underlying revenue differently. Anthropic includes model sales made through cloud providers, whereas OpenAI reportedly subtracts its partners’ share. Anthropic later told investors that it had generated \$17bn in third-quarter revenue and was profitable on an adjusted basis, but that disclosure did little to clarify how it related to the \$65bn ARR reported in July.

Valuation methods are becoming equally creative. One investment firm proposes valuing model labs on earnings before interest, tax, depreciation, amortisation and the enormous cost of training their models. Removing the expense most central to the business can make a capital-intensive lab look healthy without showing whether its economics are sustainable. Until the companies list and publish standardised accounts, investors are pricing them with figures that may share a label but not a definition.

Public accounts reveal only part of the bet

Listed technology giants provide audited statements, yet their AI exposure is still hard to see. Amazon, Google, Meta and Microsoft have made roughly \$2.4trn in off-balance-sheet commitments, more than they have borrowed through bond markets.

Their AI income is no easier to evaluate. Microsoft has described an AI business with a \$37bn annual revenue run rate; Amazon has cited a \$25bn annual recurring revenue run rate. Both figures remain black boxes that reveal too little to calculate a dependable return on capital.

That return is the central question. Goldman Sachs researchers estimate that if cloud companies are to earn 10-20% on recent data-centre investment and AI developers are to retain a 30% margin, the industry needs about \$1trn in annual revenue. Customers would have to extract still more than that in productivity gains to make the chain worthwhile. Many bosses, however, lack a clear way to measure the value produced by their AI spending.

The boom can inflate profits before proving demand

AI investment also distorts aggregate corporate earnings. Spending by the technology giants immediately becomes revenue and profit for chipmakers, equipment suppliers and construction firms. For the buyers, the same outlay passes through income statements only gradually as depreciation. The result is an early surge in suppliers’ profits without an equally immediate charge against the companies funding the build-out.

Mark-to-market gains on stakes in private AI labs add another layer. In the second quarter, unadjusted earnings for the S&P 500 rose by 50% from a year earlier. Excluding the revaluation of private investments held by Alphabet and Amazon reduced that growth rate to 32%. The lower number is still strong, but the gap shows how much headline profit can depend on valuations inside the same investment cycle.

This does not prove that the AI boom is hollow. It means that current profits combine operating performance, infrastructure transfers and paper gains that should not be treated as interchangeable. Investors need to know which component they are buying.

Economic statistics are losing clarity too

The measurement problem extends beyond company accounts. Rising memory-chip prices have caused statisticians to overstate inflation for some software products. A report by Epoch AI suggests that American GDP growth has been understated by 0.3 percentage points over the past year because Nvidia’s cross-border sales are recorded imperfectly.

At the same time, the institutions meant to improve the data are weakening. American statistical agencies have lost funding and capability, while the post-Enron accounting watchdog has had its budget cut. Strong inspection scores for the largest audit firms therefore offer less reassurance than they would under a better-resourced system.

The article’s warning is not that every AI number is false. It is that definitions, incentives and institutional capacity have failed to keep pace with the sums at risk. The industry needs standard revenue measures, fuller disclosure of infrastructure obligations and credible ways to connect AI spending with customer value. Without them, investors are not simply betting on an uncertain technology. They are betting without a reliable scoreboard.