Wall Street's AI Bet: What Happens If the Machines Don't Deliver?

There is a number in circulation that goes roughly like this: sixty to seventy per cent of American stock trading is done by algorithms. It appears in market commentary, in brokerage explainers, in the background of a hundred think-pieces about machines taking over Wall Street.

We went looking for where it comes from. We could not find it.

What we found instead were commercial trackers citing each other, blog posts citing the trackers, and one 2020 staff report from the Securities and Exchange Commission whose text could not be extracted from its file. No regulator publishes a current figure for the share of U.S. equity volume that is machine-driven. The honest position, after a day of retrieval, is that the number people repeat is an industry estimate of unknown provenance being repeated as though it were a statistic.

That matters, because the number carries an argument inside it. If most trading is algorithmic, the implication goes, then most trading is AI, and the market is a machine that could seize up. Each step of that needs testing separately, and only the first survives contact with the evidence.

Three different things are being called “AI trading”

The Federal Reserve drew the distinction more carefully than most market commentary does, in a box published with its November 2025 Financial Stability Report. Algorithmic trading, it wrote, is “automated, computer-driven trading based on predefined trading strategies” — and such systems have been used for decades for market making, order execution and arbitrage. Generative AI is a later layer: it adds “self-learning based on historical experience, adaptation based on current market conditions, and analysis of unstructured data, such as text.”

Then the Fed said the thing that gets left out of the machine-takeover story: AI’s sophistication currently costs speed. “The greater model complexity and the use of additional information by AI currently come at the cost of reduced speed, and thus the suitability of the latest AI models for trading decisions depends on the application.”

Its overall assessment was that “the majority of AI applications in trading today seem to be building upon established practices in machine learning and sophisticated data analysis techniques, rather than representing a significant departure from existing methods” — providing “efficiency gains, without a fundamental change in the trading process itself, at least for now.”

That is a regulator saying the machines are real, are not new, and are not the ones making the decisions.

So: algorithms execute a great deal. AI decides far less. And investors pay enormous sums for companies expected to profit from AI, which is a third thing entirely and the one where the money actually is.

The direction of travel is still one way

None of that is complacency. The same 2024 IMF analysis that found generative AI “used in only limited ways by actual investors” also set out how fast the ground is shifting beneath that finding.

In patent applications related to algorithmic trading, the IMF reported, the share of AI content “has risen from 19 percent in 2017 to over 50 percent each year since 2020.” AI-driven exchange-traded funds, it found, behave differently from their predecessors: “While a typical actively managed equity ETF turns over its holdings much less than once a year, AI-driven ETFs do so about once a month.” And price discovery itself has quickened: “since 2017 and the introduction of LLMs, the movement of US equity prices 15 seconds after the release of the Fed minutes seem to be more consistently in the direction of the longer-lasting movement seen after 15 minutes, in contrast to the apparently uncorrelated movements in the pre-LLM period.”

Read that sentence carefully, because the hedge is inside it. The IMF writes that prices seem to move that way. That is an observation about a pattern in the data, not a measurement of who is trading.

The stress evidence is real, if modest in scale so far. “Several AI-driven ETFs saw increased turnover during the March 2020 market turmoil,” the same analysis found, “indicating potential for increased herd-like selling during times of stress.”

But the Fed also recorded the counter-case, and it deserves equal space: AI may reduce correlated trading by giving participants richer information and more varied signals, and algorithmic traders have strong incentives to differ from one another, because non-collusion pays when others collude. On the historical record, it concluded, correlated trading “has not necessarily been detrimental to market quality.”

The exposure moved while everyone watched the price

Here is where the story changes shape.

For most of the past two years, the warning about AI and markets was a warning about valuation. The Fed’s November 2025 report said the forward price-to-earnings ratio of S&P 500 firms “remained well above its historical median,” with the equity premium “well below” its median.

By May 2026, with data through 23 April, its tone had not changed much: the ratio of equity prices to earnings for S&P 500 companies “stayed in the upper range of its historical distribution,” and an estimate of the equity premium “remained well below its historical average.”

Then, on 13 October 2026, the IMF published a different picture. In its Global Financial Stability Report, the fund wrote that AI firms had “delivered exceptional earnings growth” and that “valuation pressure measures such as forward price-to-earnings ratios have eased in several equity markets.”

Earnings caught up. Multiples compressed. Anyone still describing this market as trading at dot-com multiples is working from data that has been superseded.

What has not eased is the borrowing.

The IMF’s figures, drawn from its own analysis of the AI ecosystem:

• Hyperscalers’ share of global investment-grade bond issuance has risen from less than 1 per cent to 10 per cent since 2024, because “free cash flows are rapidly declining as hyperscalers try to sustain large AI investment needs.”
• Hedge funds’ gross equity positions plus equity margin loans have grown from 10 per cent to about 15 per cent of advanced-economy GDP, “reflecting an increase in levered positions in AI-related equities.”
• The share of single stocks inside leveraged exchange-traded funds has risen from 16 per cent to 28 per cent, “as retail investors have used these funds to gain levered exposures to AI stocks.”

And underneath all of it, a structure with no clean precedent: fourteen publicly disclosed transactions since 2023, in which hyperscalers put roughly $300 billion of equity into financially weaker AI firms, supplemented by $650 billion of debt, in exchange for $1.4 trillion in future revenue commitments through services or chip orders.

The IMF’s description of why this is dangerous is worth quoting precisely, because it is the whole mechanism in one sentence: equity and debt investments in those firms “depreciate in value precisely when future revenue streams tied to them are impaired.”

This is what the fund means when it names, as the first of three salient near-term risks, “an AI-driven market correction” — amplified, it says, “by rising leverage and concentrated exposures.”

What a disappointment would actually look like

The IMF’s trigger is not a dramatic failure. It is something quieter: “growing evidence that AI capabilities can be delivered at lower investment levels … or that potential economic gains may not be as large as expected.”

Model that through, and its own World Economic Outlook scenario has an unwinding of AI investment reducing global GDP, investment and consumption by as much as 0.4, 1.3 and 0.6 per cent respectively by 2028. Investment falls hardest. Consumption falls least, because it falls through wealth effects rather than through jobs.

That is a correction story, not a collapse story. The distinction matters and the IMF draws it explicitly: these risks “by themselves do not necessarily threaten financial stability but could become systemic — impairments to financial intermediation and credit provision, with serious negative economic effects — if they are amplified by underlying vulnerabilities.”

On whether they would be: the fund’s stress tests found “the banking system would remain adequately capitalized in a bond yield spike and AI-led correction scenario,” though “a significant share of vulnerable institutions would emerge,” and some banks’ growing entanglement with non-bank lenders “could prove a source of fragility.” Household debt, it notes, “remains relatively low in aggregate.”

A correction is plausible. A 2008 is not what the evidence describes.

What rhymes with 1999, and what does not

The comparison is unavoidable and it is usually made badly.

What genuinely rhymes: a general-purpose technology whose productivity payoff is disputed; an infrastructure build-out financed increasingly with debt; and a market where a handful of firms carry the story. The IMF’s modern version of “the internet will change everything, eventually” is its suggestion that AI capabilities may arrive at lower investment cost than the current build-out assumes — which is the polite way of saying the data centres may be overbuilt.

What does not rhyme, and what the dot-com analogy usually omits: the leading AI companies are, in the IMF’s words, “strong” in profitability and have delivered “exceptional earnings growth.” Their 1999 counterparts largely had neither revenue nor a path to it. And the novel risk here — the circular structure of vendor investment, customer commitments and debt — has no equivalent at this scale in the earlier cycle.

The dot-com comparison is a useful instinct and a poor spreadsheet. Anyone who tells you the two cycles are the same has not looked at the earnings.

The uncomfortable conclusion

The premise of this investigation was that Wall Street has become dangerously dependent on AI. The evidence supports a modified version of that, and the modification is the story.

Dependence on AI as infrastructure is real but is not, on the regulators’ own account, currently the acute risk. The sixty-per-cent figure is unverifiable and the AI component within it is smaller than the framing implies.

Dependence on AI as an investment thesis is real and large — and it has been moving out of equity prices and into credit, leverage and circular commitments, exactly where a disappointment would be amplified rather than absorbed.

That is a harder risk to see than a soaring multiple, and a harder one to unwind.

What this report does not establish is a date, a trigger or a magnitude. Nobody should read it as a forecast. It establishes that the measurements most often used to describe AI’s role in markets are weaker than their use implies, and that the exposures regulators now flag in detail sit somewhere other than the number everyone quotes.