Markets love a growth story that starts with a paradox. AI is sold as abundance, but its first-order effect is contraction. Efficiency always destroys more capacity than it preserves. When a major asset manager warns the number of public companies could fall by a fifth as AI rolls through the economy, that is not sensationalism. It is a base case. The last time corporate America embraced a general-purpose technology with religious zeal, we got fewer, larger, more leveraged firms and a long tail of stranded incumbents. This time the cycle is faster, the capex is bigger, and the winner-take-most dynamics are sharper.
AI is a scalpel, and scalpels cut. The early evidence is not subtle. Companies plowing money into automation are also cutting headcount. The near-term margin boost is real, but so is the attrition among firms that cannot keep up on cost or product velocity. In complex systems, efficiency is a form of fragility when it reduces redundancy. The economy becomes more brittle as duplicated roles and processes vanish. Like a forest after a controlled burn, the ecosystem looks cleaner right up until a drought hits.
History backs the sequence. Electrification, numerical control, ERP, and offshoring did not produce broad-based firm proliferation. They produced consolidation, scale economies, and a cull of middling players. The adoption curve is a J-curve: pain first, then gains. Investors are anchoring on the later slope and ignoring the front-loaded shakeout. Economists with a long view argue AI is real and still early, which is likely correct. That makes the contraction phase more, not less, probable. Early innings concentrate power, cash flow, and talent in a few dugouts. The scoreboard looks great if you own the winners. It looks silent if you own the rest.
In market ecology, AI acts like an invasive species. It does not create a level playing field; it steepens the slope. Capital, data, and distribution advantage compound in a power-law. We already see it in the valuation gravity around the infrastructure layer. Investors extrapolate current unit economics and assume ubiquity will expand the total addressable market for everyone. More often, ubiquity compresses price and funnels value to the bottlenecks: compute, power, and proprietary data.
This concentration is not just an academic point. It is an index construction risk. The headline looks bullish as a handful of platforms lift the cap-weighted averages. Under the surface, breadth narrows, and the correlation structure changes. When liquidity rotates, the exit doors are small for the bottom 400 names. If the number of listed firms falls another 20 percent through failures, take-privates, and roll-ups, the index becomes a barbell with a heavier weight on one end. You can call that diversification, but it behaves like a single bet on a few balance sheets, grid connections, and supply chains. The illusion of safety increases systemic fragility.
Game theory clarifies the corporate behavior. The payoff matrix creates a prisoner’s dilemma. If your rivals spend on AI, you must spend to avoid falling behind, even if the expected project returns are mediocre. The rational outcome is industry-wide overinvestment, margin compression, and price wars in categories that AI commoditizes. We have seen this film: the railroad boom, the fiber optic glut, the cloud infrastructure land grab. Suppliers prospered; many customers earned subpar returns until capacity was rationalized by time, bankruptcy, or mergers.
AI also changes cost structure. More fixed costs in compute and model integration mean higher operating leverage. That looks marvelous in an upswing and catastrophic in a downswing. Investors often misprice that convexity because the good periods are more visible than the bad. Add regulatory and compliance overhead to the equation. Model governance, data provenance, and liability insurance do not scale as cleanly as inference calls. Most management teams are modeling the upside with Monte Carlo optimism and the downside with single-point estimates. That is the wrong tail to neglect.
Under the hood, the system relies on a few chokepoints. Advanced chips come from a narrow vendor set. High-bandwidth memory is tight. Data centers require power that many grids cannot supply on the timelines promised in slide decks. A utility interconnect delayed by 18 months does not care about your product roadmap. In engineering, load plus resonance breaks bridges. Here the load is capex; the resonance is a synchronized corporate rush into the same architectures, vendors, and talent pools. One supply shock or policy pivot and the oscillation climbs.
There is also a softer fragility. When companies right-size away tacit knowledge, they swap seasoned operators for automated workflows and prompt libraries. It works until it does not. Model drift, subtle hallucinations in edge cases, and legal ambiguity around training data are slow variables that become fast variables under stress. Public sentiment has picked up on this gap. Skeptics argue the current generation of large models is derivative rather than creative, better at remixing than inventing. Even if that is only partly true, it means moats built on content, back-office process, or basic analytics will erode. Revenue per employee can rise while revenue per customer falls, and multiples do not like that arithmetic.
Investors and executives should also expect M&A to accelerate shrinkage. If AI makes integration and cost takeout more predictable, private equity and strategics have a stronger incentive to buy, rationalize, and delist. The public roster thins. The remaining companies are larger, more efficient, and more exposed to shared risks. That is not a crisis scenario. It is simply the mechanical outcome of incentives and technology diffusion.
The right frame is not optimism versus pessimism. It is fragility versus antifragility. Who gains from volatility and who needs a constant tailwind to survive. Businesses with variable cost structures, pricing power, and real options tend to benefit from a deflationary technology shock. Those with high fixed costs and competition on open models tend to bleed. Picks-and-shovels logic still applies, but not the lazy version. The bottlenecks are energy, power equipment, grid software, and the verification and security layers that make AI safe enough to scale. There is also optionality in firms that convert AI into lower working capital and better cycle times without betting the balance sheet on owning the model stack.
For portfolio construction, assume the investable universe shrinks and the dispersion of outcomes widens. Index concentration increases headline risk and liquidity risk. Narrative cycles will be violent. The base rate says many companies will adopt AI and few will monetize it at scale. That is the same pattern we saw in previous general-purpose technology waves. The market will eventually re-rate productivity winners, but not before it culls the long tail. A 20 percent drop in public company count is not an outlier. It is the midpoint of a power-law transition.
The paradox holds. AI will create abundance, but first it will remove redundancy. That is how systems get leaner, and for a time, more brittle. Investors should prefer balance sheets over blue-sky, cash flow over promises, and optionality over purity. The discipline is simple: invert the story. Ask not what AI can add, but what it will take away. Then measure whether the business still stands when the scaffolding is gone.
