The uncomfortable question about AI is not whether it works, but who pays the bill

Let's start by separating the wheat from the chaff. Will Lockett's text, published on his Substack, is an avowedly critical opinion piece, not a neutral financial analysis, and its most substantial part sits behind a paywall. The figures it handles come from the article itself and could not be independently verified. With those caveats up front, the thesis deserves to be discussed because it points to a real tension in the sector.

The central figure it cites is striking: according to the author, OpenAI is said to have brought in $5.7 billion in the first quarter of 2026 and posted, against that, a non-GAAP loss of $6.9 billion; that is, $1.22 lost for every dollar billed. He adds that non-GAAP metrics tend to paint a rosier picture than standard accounting ones, and mentions accumulated losses of $38.5 billion over the course of 2025. These figures should be treated as the author's claims pending confirmation, not as established facts.

Lockett extends his criticism to Anthropic, whose image of being close to profitability he attributes —citing analyst Ed Zitron— to a market position he considers unsustainable and to a favorable reading of EBITDA. Here he touches on a legitimate technical point: EBITDA excludes depreciation and interest, and in a business that devours investment in data centers, that exclusion can mask a negative free cash flow. It is a reasonable methodological warning, regardless of whether one shares the verdict.

Where the article enters speculative territory is in forecasting a possible bankruptcy of OpenAI in 2026 or 2027 and in speaking of a "bubble" sustained by circular investment and the absence of regulation. These are the author's hypotheses, not certainties, and they must be presented as such. Predicting the collapse of a company that continues to raise capital on a massive scale is a risky exercise that recent history has refuted more than once.

What is valuable, stripped of the apocalyptic tone, is the underlying question: the cost of training and running inference on frontier models is real, enormous and well documented, and current API and subscription prices do not cover it comfortably. That gap exists. The open question —and here informed optimism has its place— is whether the efficiency curve of computing, the improvement of the models and the emergence of high-value use cases will close it before the patience of capital runs out. We do not know yet. But distinguishing the difficult economics of a young sector from the prophecy of its downfall is exactly the kind of nuance these headlines tend to trample.