Wall Street has a number for everything except honesty. This year it has settled on a curious consensus: the S&P 500 will close 2026 somewhere between 7,000 and 8,250, with the median Reuters poll of 47 strategists landing at 7,620 — barely above where the index sits today. The Dow is pegged near 52,500. Ed Yardeni, the most bullish voice on the Street, has raised his target to 8,250 on the strength of corporate earnings he says are “increasingly backed by hard numbers rather than market enthusiasm.” Bank of America and Stifel see 7,000 to 7,100. The spread between the most bullish and most bearish calls is the tightest in nearly a decade — and that, historically, is when markets get blindsided.
The reason for the consensus, and the reason for the unease beneath it, is the same word: AI.
Start with the scale of what is being built. Microsoft, Alphabet, Amazon, and Meta — joined increasingly by Oracle — are on track to spend somewhere between $600 billion and $700 billion on capital expenditure in 2026, nearly double what they spent the year before. Roughly three-quarters of that, more than $450 billion, is going directly into AI infrastructure: data centers, chips, power. Microsoft’s AI business alone has crossed a $37 billion annual run rate, up 123 percent year over year. Google Cloud’s backlog has nearly doubled to $467 billion. These are not small numbers, and they are not fictional ones.
READ: The illusion of infinite returns: Why the AI bubble mandates extreme caution (May 21, 2026)
But set them against the revenue of the companies actually consuming all that compute, and the picture gets uncomfortable. OpenAI ended 2025 with about $20 billion in annual recurring revenue. Anthropic’s run rate crossed $9 billion in January. Add them together and you get a sum that is roughly three percent of what the hyperscalers plan to spend building capacity for them this year alone. Somewhere between “build it and they will come” and “they have not come yet” lies the entire AI bubble debate.
What makes this round different from the dot-com crash, and worth taking seriously rather than dismissing, is the financing structure underneath it. Nvidia commits $100 billion to OpenAI. OpenAI uses a meaningful share of that capital to buy — what else — Nvidia chips. Microsoft and Nvidia jointly commit $10 billion to Anthropic, which in turn commits $30 billion to Azure cloud usage. Michael Burry, who built his reputation shorting subprime mortgages before anyone else saw the rot, is now shorting Nvidia and asking a pointed question on social media: “OpenAI is the linchpin here. Can anyone name their auditor?” His larger claim is sharper still — that true end demand is “ridiculously small” and that most customers are effectively funded by their own suppliers.
This is not accounting fraud. It is vendor financing, a practice with precedent — Uber subsidized its own driver and rider base in its early years to build market share. But there is a difference between subsidizing adoption in an already-proven industry and subsidizing demand for a technology whose return on investment is still being argued about in boardrooms. Even Sundar Pichai, not a man given to alarmism about his own company’s central bet, told the BBC there are “elements of irrationality” in the AI market. Sam Altman has said the same about investor sentiment generally. When the chief executives admit the exuberance, the rest of us are allowed to notice it too.
Layer onto this a financing shift that should worry anyone who lived through 2008: the hyperscalers’ capital spending, after dividends and buybacks, now exceeds their free cash flow. They are increasingly funding the AI buildout with debt rather than cash — more than $100 billion in bonds issued in 2026 alone — and investors are demanding record levels of credit-default-swap protection against the possibility that some of this paper goes bad. A boom funded by your own balance sheet is resilient. A boom increasingly funded by other people’s money is not.
Washington, for its part, is not regulating any of this — it is greasing it. In March, Trump brought seven hyperscalers to the White House to sign the Ratepayer Protection Pledge, committing them to fund their own power generation so that American households are not stuck with the electricity bill for AI data centers. The administration has dangled chip tariff exemptions in exchange for cooperation, intervened directly in the PJM power market to fast-track plant construction, and dismantled Biden-era AI safety reporting requirements in favor of what it calls a “Federal Preemption” standard. The message to the industry is unambiguous: build faster, and we will clear the obstacles. That is industrial policy in the service of a story Washington needs to be true — American AI supremacy — regardless of whether the underlying unit economics have caught up yet.
And here is the part that should concentrate any investor’s mind, retail or institutional: ten companies — Nvidia and Broadcom chief among them — now account for more than a third of the Nasdaq-100’s total weight, and over half of the broader Nasdaq Composite’s performance. The remaining ninety-plus names span software, healthcare, retail, and industrials, including a strong run this year from memory-chip makers like Western Digital, Seagate, and Micron riding the DRAM shortage. But when the index is this top-heavy, and the top is this exposed to a single, unresolved question — is AI revenue real demand or financed demand — a re-rating does not stay contained. It moves the whole board.
None of this means the AI buildout is fake, or that the technology will not eventually justify the spending. The compute shortage is real; the backlogs are real; the productivity gains, where they have shown up, are real. But “real” and “fully priced” are different claims, and right now Wall Street’s tightest-ever consensus is betting on both at once. That is usually the moment to ask the question nobody at the table wants to ask out loud.