Cheap Intelligence

When AI Eats Alpha

AI agents won't just automate investing. They may systematically eliminate the easy returns created by slow analysis, poor execution, emotional decisions, and market friction — just as a heavily indebted system becomes less able to tolerate mispricing.

Subscribe
MarketsMacroeconomicsAI
Listen to this post
AI Summary AI agents will commoditize sophisticated investment capabilities like continuous monitoring, analysis of thousands of securities, and optimal execution, transforming what once required expensive institutional infrastructure into a widely available software subscription. …
  • AI agents will commoditize sophisticated investment capabilities like continuous monitoring, analysis of thousands of securities, and optimal execution, transforming what once required expensive institutional infrastructure into a widely available software subscription.
  • As AI adoption spreads, the dispersion of returns attributable to now-commoditized skills will compress—not necessarily lowering market returns overall, but shrinking the premium that skilled investors earn for tasks machines can now perform.
  • This efficiency transformation is colliding with a financial system carrying elevated debt levels and compressed equity risk premiums, meaning AI-driven price discovery will arrive precisely when the system is least able to tolerate accurate pricing of mispriced assets.
  • When millions of AI agents independently analyze the same information and reach similar conclusions nearly simultaneously, markets may become both more efficient in normal times and more violent during transitions, as repricing that once took weeks happens in minutes.
  • The author assigns 80% probability that AI will materially compress alpha from common signals and slow analysis, but only 50-60% that the combination of AI efficiency and expensive capital creates a persistent compression regime, with just 25-35% odds this becomes the primary cause of a major crisis.

Most of what I have written about compression has been about valuation.7 The argument was not complicated. Once capital has a price again, debt gets refinanced at rates that no longer flatter it, growth stops being something a company can manufacture, and investors eventually lose the appetite for paying extraordinary prices for earnings that have not arrived yet.

Multiples compress.

There is a second compression behind that one, and I think it matters as much. It is the compression of return itself, and the thing driving it is AI — not because some model is about to find the perfect strategy, but for nearly the opposite reason.

Eventually everyone will have one.

The Market Has Always Paid for Other People's Mistakes

A considerable share of investment return has always existed because markets are imperfect.

  • Someone didn't read the filing, or misread the earnings call, or sold because they were frightened rather than because anything had changed.
  • Someone let a position run past the size they would ever have chosen on purpose, or never rebalanced, or executed badly on the way out.
  • Someone could not hold five thousand securities in their head at once.
  • Someone knew an asset was mispriced and had neither the capital, the technology, nor the mandate to do anything about it.

Professional investing has spent decades building increasingly sophisticated machinery to harvest those imperfections, and yet enormous pools of capital are still run by humans working under very human constraints. Attention runs out, analysis costs money, emotion gets in the way, institutions move slowly, and information takes time to travel. Every one of those deficiencies is worth something to whoever is operating better than the person on the other side of the trade, and that is the value AI threatens.

The great investment opportunity created by AI may turn out to be the destruction of investment opportunity itself.

Millions of Portfolio Managers That Never Sleep

The version most people picture is an investor asking a chatbot whether Nvidia looks expensive. That is not the version that changes anything. The one that does is an agent that never stops reading — every 10-Q and 10-K, every transcript and guidance change, every competitor announcement — while it watches rates, spreads, currencies and commodities move underneath the positions it holds. It knows what a given piece of news did the last few thousand times something like it happened. It knows what a position will cost to exit after tax, how much factor risk it is quietly carrying, and which holding has grown large enough to be a problem.

Then it executes. Continuously.

Today that capability belongs to hedge funds, quant shops, and the large asset managers. Tomorrow it is a software subscription. Eventually it is a commodity, and a market in which that capability is a commodity is a different market.

We Have Seen This Movie Before

High-frequency trading is the precedent for what happens when technology finds an inefficiency it can attack. Researchers at the Bank for International Settlements found that more than a fifth of trading volume in their sample occurred in latency-arbitrage races1 — contests over discrepancies often worth a fraction of a tick, fought with enormous technological investment. Once the opportunity was understood, competition collapsed into speed, and microseconds became the product.

AI runs the same process one level up the intellectual stack. The race is no longer about who executes first but about who understands first: who connects a new piece of information to everything else, decides what it matters for, and builds the right response. When millions of agents can do that almost instantaneously, the question stops being who understands first and becomes something more uncomfortable.

What happens when everybody understands at roughly the same time?

Alpha Compression

Alpha Has a Half-Life

Investment strategies share an inconvenient trait, which is that they stop working once enough capital discovers them. A mispricing exists, someone finds it, capital moves toward it and the mispricing shrinks; more participants pile in and the return falls further; and what once looked like extraordinary skill ends as a crowded trade producing ordinary returns.

That is reflexivity in its simplest form, and AI accelerates every step of it.

A July 2026 NBER working paper by Ralph Koijen and Bradford Levy confronts the problem head-on. As investors adopt AI, prices adjust, wearing away the very patterns those systems were trained to exploit.2 Their real-time benchmark still found that the best-optimized agentic systems, working from earnings-call transcripts, raised the explained variation in announcement-window returns from about 8% to nearly 20%.2

The systems are getting dramatically better at extracting the information embedded in markets, and that success is exactly what destroys the economic value of extracting it.

Intelligence Becomes a Commodity

For most of the history of markets, intelligence was the scarce input. Trained analysts, computing power, financial data, research departments, portfolio optimization and institutional risk management were all expensive, and their scarcity created rents. AI attacks nearly every one of those constraints at once. An individual investor may soon command the analytical capability that took an entire institutional research organization twenty years ago.

Democratizing that capability does not mean everyone earns extraordinary returns. It means extraordinary capability becomes ordinary, and markets do not pay extraordinary returns for ordinary capabilities. The AI-investing conversation keeps asking how much better AI will make an individual investor.

The more important question is what happens when it makes everyone better.

AI does not have to touch the market's return to do its damage. What it compresses first is the spread between investors — the part of the gap between a bad one and a good one that was really a gap in diligence. Picture three of them, earning 3, 7 and 12 percent. Hand each an agent.

  • The bad investor stops panicking, stops forgetting to rebalance, stops leaving cash idle, and lands somewhere near 6.5.
  • The average one was already close to the market and stays there.
  • The excellent one, whose 12 was mostly the reward for reading more carefully than everyone else, finds that everyone else now reads carefully too, and settles at 8.

The market's 7 is untouched. What has been taken out is most of the five points that used to be paid for doing well what machines now do for everyone — and the fight moves up a level, to the one point that is left.

Poor Execution Becomes Less Valuable to Exploit

Many investment mistakes are not analytical at all. They are execution failures:

  • Investors panic after declines and chase rallies.
  • They let positions grow dangerously large, leave cash idle, rebalance inconsistently, and ignore taxes.
  • They sell for emotional rather than portfolio reasons and fail to hedge risks they already recognize.
  • They misunderstand correlation, confuse volatility with permanent loss, and take risk without adequate compensation.

Every one of those behaviors is somebody else's opportunity.

Now put an agent between the human and the portfolio. The human can stay emotional; the portfolio no longer has to be. The human forgets; the agent does not. The human ignores four thousand securities; the agent watches all of them and compares them continuously. None of this makes anyone rational. It shrinks the amount of capital that irrational execution directly controls, and that is the distinction that matters.

Important

Markets do not need humans to become rational. They need machines standing between human intention and capital allocation, and that is what is being built.

Then It Collides With Debt

Under different circumstances, greater market efficiency would simply be good news — better price discovery, lower transaction costs, deeper liquidity, less wasted capital. Those benefits are real. But the transformation is arriving in a particular financial regime, and the regime is the point.

Federal Reserve's May 2026 Financial Stability Report
The Federal Reserve's May 2026 Financial Stability Report found the forward price-to-earnings ratio in the upper range of its historical distribution, with the equity premium well below its historical average.3 It also noted that riskier firms, particularly those relying on private credit, were having difficulty servicing their debt.3 In the New York Fed's accompanying survey of market contacts, half cited AI as a potential shock, up from 30% the prior fall, and what they flagged specifically was AI equity valuations and capital spending increasingly funded by debt.4

The debt calendar tells the same story from the other side. S&P Global now places the speculative-grade nonfinancial maturity peak in 2031, pushed out largely by bond issuance for AI and digital infrastructure, while debt rated B- and below rises to $268.8 billion in 2028, concentrated in the U.S.5

The machines being built to price capital more precisely are being financed, in part, by debt they will eventually price.

That is the collision. Systems capable of becoming extraordinarily efficient at determining the price of capital are arriving inside a financial system that has grown accustomed to capital being mispriced cheaply.

The Market May Become Less Forgiving

This is where return compression joins valuation compression. The old system contained enormous slack, and companies, investors and capital could all afford to be inefficient inside it:

  • Low rates made mistakes survivable.
  • Rising asset prices concealed the ones that weren't.
  • Expanding multiples produced returns even when the underlying business barely improved.

Human inefficiency, meanwhile, gave sophisticated investors a reliable source of outperformance.

Now remove those cushions one at a time. Capital has a price, and debt has to be refinanced at it. AI intensifies competition between companies, erodes information advantages, improves execution, accelerates price discovery and lets capital move faster. None of that has to mean a market that falls forever. It means a market that is increasingly unwilling to reward mediocrity.

Efficient Does Not Mean Stable

A market run largely by AI does not become calmer. It can become more efficient and more violent at the same time. In July the IMF's Tobias Adrian warned that AI is compressing the timelines of financial decisions, and that the flash crashes of the future may come less from coding errors than from many AI systems reacting in parallel to the same information.6

Most of the time such a market would feel remarkably efficient. Earnings are already modeled, macro data already anticipated, the portfolio implications already calculated, so mispricings close quickly and ordinary information moves prices less than it used to. Then something genuinely unexpected happens, and millions of systems independently reach nearly the same conclusion at nearly the same moment. Repricing that once took weeks happens in hours, and what took hours happens in minutes.

The market gets better at reaching equilibrium and more violent in the transition between equilibria.

Three Compressions

AI may ultimately produce three related compressions.8

Margin compression. Intelligence becomes cheap. Smaller organizations gain capabilities once reserved for large ones, competition intensifies, and the rents attached to organizational scale decline.

Valuation compression. Capital is no longer free. Refinancing matters, investors demand more compensation for duration and uncertainty, and extraordinary multiples become harder to sustain.

Return compression. Information, portfolio construction, and execution become commoditized. Easy alpha disappears, and the advantage once supplied by inattentive or poorly executing participants shrinks.

These forces reinforce one another: corporate margins ↓, valuation multiples ↓, excess returns ↓. None of that eliminates investment returns. It changes where they come from.

What Remains Scarce?

Compression becomes recursive. Information advantage compresses, then execution advantage, then portfolio-construction advantage, then model advantage. Eventually even the advantage of having AI compresses, because everyone has it.

When technology commoditizes something, value migrates toward whatever stays scarce. As intelligence, analysis, execution and portfolio optimization become abundant, each becomes less valuable in turn, and return moves toward what cannot easily be manufactured:

  • Truly proprietary information
  • Control of scarce physical assets
  • Unique distribution and regulatory privilege
  • Extreme patience and long-duration capital
  • The stomach to hold a position when every optimization model says trimming exposure is temporarily rational

Above all, return moves toward liquidity at the moment liquidity itself becomes scarce. The future advantage may belong less to whoever owns the smartest model than to whoever owns something no model can manufacture.

The Odds

Everything above is conditional. Here is what I actually believe, and how strongly.

Probability Claim
90% AI agents materially improve portfolio construction, research and execution
80% This materially compresses alpha from common signals, slow analysis and poor execution
70% Retail/institutional behavioral inefficiencies become meaningfully harder to monetize
75% Alpha increasingly migrates toward scarce information, capital, liquidity and unusual risk-bearing
65% Agent crowding makes markets faster and occasionally more unstable
60% This contributes meaningfully to lower excess returns across public markets
50–60% AI-driven efficiency + expensive capital/debt creates a persistent broader compression regime
25–35% This becomes the primary cause of a major valuation reset/crisis

That's why I'd put ~60% on the complete thesis but ~80% on the core mechanism. The uncertainty isn't whether AI makes markets dramatically more competitive. I think that's increasingly hard to argue against.

The uncertainty is where the displaced return goes. That question is the next article.9

The Final Irony

Investors have imagined AI as another tool for generating alpha, and initially it will be. Early adopters will hold an advantage, sophisticated systems will find opportunities humans miss, and agents will manage portfolios better than most individuals manage their own. But technological advantages diffuse, and what begins as alpha becomes infrastructure.

The spreadsheet did not permanently make spreadsheet users rich. Knowing how to search the web stopped being an edge. Algorithmic execution became part of the market itself. AI portfolio management will follow the same path: first an advantage, then widespread adoption, then compression. In the end, it becomes the minimum capability required just to participate.

When that happens, an uncomfortable fact surfaces. For decades, some portion of market returns came not from extraordinarily productive capital but from friction, imperfect information, and poor human execution. AI may remove much of that friction. It will do so at precisely the moment expensive capital and heavy debt loads are removing the cushions that let mispricing persist.

The next phase of compression won't be driven by interest rates alone. It will be driven by intelligence itself: a market brutally efficient at determining what assets are worth, at exactly the moment the financial system can least afford the answer.

Part of the series: Cheap Intelligence
  1. When AI Eats Alpha
  2. Where the Displaced Return Goes

Footnotes

  1. Aquilina, Budish & O'Neill, Quantifying the High-Frequency Trading 'Arms Race' — BIS Working Paper No. 955 (2021) — The precedent this piece builds on: more than a fifth of trading volume in the sample occurring in latency-arbitrage races, fought over discrepancies often worth a fraction of a tick. Once the opportunity was understood, competition collapsed into speed. https://www.bis.org/publ/work955.htm ↩
  2. Koijen & Levy, Assessing the Benefits of Optimized Agentic AI Systems for Asset Pricing — NBER Working Paper 35431 (July 2026) — The paper that states the self-defeating property directly — as investors adopt AI, prices adjust and wear away the patterns the systems were trained to exploit — while its real-time benchmark shows best-optimized agentic systems raising explained variation in announcement-window returns from about 8% to nearly 20% from earnings-call transcripts. Both halves of this section's argument come from the same source. https://www.nber.org/papers/w35431 ↩
  3. Financial Stability Report, May 2026 — Board of Governors of the Federal Reserve System — The regime the efficiency is arriving into: a forward price-to-earnings ratio in the upper range of its historical distribution with the equity premium well below average, and riskier firms — particularly those relying on private credit — having difficulty servicing their debt. https://www.federalreserve.gov/publications/files/financial-stability-report-20260508.pdf ↩
  4. Near-Term Risks to the Financial System — Federal Reserve, Financial Stability Report, May 2026 — The survey of market contacts: half cited AI as a potential shock, up from 30% the prior autumn, flagging AI equity valuations and capital spending increasingly funded by debt. https://www.federalreserve.gov/publications/2026-may-financial-stability-report-near-term-risks.htm ↩
  5. Credit Trends: Global Refinancing — Brisk Issuance Pushes Speculative-Grade Maturity Peak Out To 2031 — S&P Global Ratings, July 2026 — The specific study behind both figures. The speculative-grade nonfinancial peak moved to 2031 in a rapid shift — 2028 in January, 2029 in April, 2031 by July — and maturities of debt rated B- and lower rise to $268.8 billion in 2028, concentrated in the U.S. and in healthcare, high technology, and media and entertainment. https://www.spglobal.com/ratings/en/regulatory/article/credit-trends-global-refinancing-brisk-issuance-pushes-speculative-grade-maturity-peak-out-to-2031-s101697989 ↩
  6. Tobias Adrian, How Central Banks Can Contain Financial Stability Risks as AI Accelerates Change — IMF Blog, July 23, 2026 — The warning quoted here, in the IMF's own words: AI may improve liquidity and accelerate price discovery, and widespread adoption could amplify systemic risk under stress, with future flash crashes arising from many AI systems reacting in parallel to the same information rather than from coding errors. https://www.imf.org/en/blogs/articles/2026/07/23/how-central-banks-can-contain-financial-stability-risks-as-ai-accelerates-change ↩
  7. Valuation Gravity: Why 2022–2028 Could Mirror Historic Compression Cycles — The earlier writing this sentence refers to — The Compression Series entry where the valuation argument is made directly: a century of multiple-compression cycles, the 2022–2028 roadmap, the debt wall refinancing at rates that no longer exist, and the Dollar Trap. Where that piece compressed the multiple, this one compresses the return. ↩
  8. Compression Is the Story of This Decade — The Compression Series synthesis, which names four forces compressing American economic life at once and argues the conventional frames only see them one at a time. The three compressions set out here are the same argument narrowed to capital markets, with return compression as the term that piece did not yet have. ↩
  9. Where the Displaced Return Goes — Part two, which takes up exactly this question. Compression is a claim about something leaving an account, and money that leaves one account arrives in another or stops being spent. That piece argues the destination is four places at once — back to the counterparty who was quietly paying it, out to whoever sells the capability, upward to whoever owns what stays scarce, and a residual simply saved. ↩
Back to the Journal