The market bought a story. Two years into the most hyped technology cycle in a generation, that story is quietly breaking down in enterprise P&Ls across every major sector — and the multiple hasn't moved.
The story was elegant: AI adoption equals headcount reduction equals margin expansion. Every earnings call echoed it. Every analyst deck assumed it. Every company that announced an AI strategy got rewarded — not for results, but for the announcement. The market wasn't pricing AI as a technology. It was pricing AI as a cost narrative.
The problem is the narrative is wrong. And the tools to see that have existed for decades — if you're willing to apply them.
The Numbers Don't Match The Thesis
Microsoft, Uber, and a growing list of enterprises are discovering something their slides didn't model: AI adds a new cost layer on top of the existing one. Token-based pricing doesn't scale linearly — it scales with usage, and usage at enterprise scale is exponential. The more an organization deploys AI, the higher the bill. That's the inverse of the efficiency story that drove the trade.
The clearest admission came from inside Nvidia itself: a senior VP disclosed that his team's compute costs now exceed what they were paying in employee salaries. This isn't a temporary transition cost. It's the new cost structure. And it's coming from the company that benefits most when that bill goes up.
Jamie Dimon put it directly: "Efficiency gains from AI will likely be competed away and passed onto customers rather than permanently adding points to the bank's profit margin." That quote appeared in an earnings call. Almost nobody modeled it into their thesis.
The Prisoner's Dilemma Nobody Is Naming
Large enterprises are trapped. There is no clean exit.
They can't not adopt AI. Investor pressure, competitive optics, and board expectations have made adoption a mandate regardless of underlying economics. You announce an AI strategy because not announcing one is worse for the stock price than announcing one that doesn't work.
They can't easily cut the headcount the AI was supposed to replace. The friction is structural: legal liability, institutional knowledge, political capital, union agreements, severance costs. The playbook says AI will let organizations do more with fewer people. The reality is most companies are doing more with the same people — plus an AI bill.
Result: paying for both layers simultaneously. The old workforce and the new AI layer. Not as a temporary bridge — as a permanent feature of how large organizations actually work.
"Every player at the table is forced to adopt. Few can realize the savings the adoption was supposed to unlock. The infrastructure layer collects regardless. The trap is structural, not cyclical."
The Number Nobody Is Modeling
There is a principle in economics called Jevons Paradox: when the efficiency of using a resource improves, total consumption of that resource typically increases rather than decreases. Steam engines became more efficient — and coal consumption went up, not down, because cheaper operating costs expanded their deployment across every possible use case.
Token prices are falling fast. Usage is scaling faster. Agentic workflows — the next wave of enterprise AI deployment — consume exponentially more tokens than the prompt-response tools enterprises have been running. Each agentic task triggers 10–20 LLM calls. RAG architectures inflate context windows 3–5x. Always-on monitoring agents consume compute 24 hours a day.
The average enterprise AI budget has grown from $1.2 million per year in 2024 to $7 million in 2026 — a 5.8x increase in the same window that per-token costs fell by orders of magnitude. This is the number that breaks the efficiency thesis. And almost no one has built it into their models.
The Munger Diagnosis
Charlie Munger called it the Lollapalooza Effect: when multiple psychological forces stack in the same direction simultaneously, they don't add — they compound. The result is a narrative so self-reinforcing it can survive well past the point where the underlying math breaks down.
The AI efficiency thesis is a Lollapalooza event in markets. Count the forces:
Social proof: every CEO announcing an AI strategy on every earnings call. Authority bias: Nvidia, Microsoft, Goldman validating the narrative from stages and whitepapers. Commitment bias: once you've announced to investors, you can't walk it back without a stock price penalty. Availability bias: the demo wins are vivid and quotable; the runaway token bills are abstract and buried in quarterly filings. Recency bias: two years of AI-adjacent stocks rewarding the narrative. Incentive misalignment: management compensation tied to adoption announcements, not verified ROI.
Six psychological forces stacking simultaneously. None of them requiring the thesis to be true. All of them requiring the thesis to be maintained.
"This is not a market pricing AI correctly. This is a market pricing a bias stack. And the behavioral forensics tell you exactly how that ends: not gradually — all at once, when the Bayesian update forces through and the consensus can no longer explain away the evidence."
Who Wins Unconditionally
Before naming the trade, the unconditional position: the infrastructure layer.
Nvidia, cloud providers, and AI API companies are not making a bet on whether enterprise AI delivers efficiency gains. They collect on every token consumed, every agentic workflow spun up, every AI pilot that fails and gets restarted. Their economics sit upstream of the question everyone is arguing about. Usage scales regardless of ROI. Infrastructure wins regardless of outcome.
This is the picks-and-shovels position in a gold rush — with the added feature that the gold rush continues whether or not anyone finds gold. Every quarter that the efficiency thesis fails to deliver, enterprises redouble the investment trying to make it work. The infrastructure bill compounds.
The second winner: AI-native companies. Not the incumbents deploying AI tools — the challengers built from day one without the assumption that human headcount would scale with the business. Sequoia's 2025 portfolio analysis found that AI-native companies captured 78% of new enterprise contract value in categories where both AI-native and AI-augmented competitors existed. That gap isn't closing. It's widening, every quarter.
What AI-Native Actually Looks Like
Dealithic is one concrete example of what this architecture looks like in practice. The platform was built from day one without the assumption that analyst headcount would scale with deal volume. It handles deal intelligence, investor network management, IR, trading technology, and capital structure — the output of what would traditionally require a significant team — at a fraction of the cost structure. That decision isn't a feature of the product. It's a structural cost advantage that compounds every quarter an incumbent firm pays for both their AI stack and the human infrastructure underneath it.
The AI-native challenger was never paying for that human infrastructure at all. That's not a temporary head start. It's a gap that widens with scale.
The Only Question That Matters
Stop asking: is this company adopting AI?
Start asking: is this company AI-native, or AI-augmented?
The distinction is structural, not operational. An AI-native company was designed without the assumption that headcount scales with output. It reaches enterprise-level results with a radically different cost structure — and as Jevons dynamics compound, its cost base grows slower than its output. That gap doesn't close. It widens.
An AI-augmented company is a legacy organization that has added AI tooling on top of an existing headcount structure. It may be more productive per employee. But it is paying for both the productivity gain and the cost of the employees who generated it. The AI bill is additive, not substitutive. At scale, the economics of these two architectures compound in opposite directions. Incumbents can't replicate their way out of it — you can't build a native cost structure on top of a legacy one.
The Inversion
Munger's most useful move was inversion: tell me where I'll die so I won't go there.
Applied here: what would have to be true for the AI efficiency thesis to work?
You'd need enterprises to cut headcount proportionally and quickly. You'd need token costs to not exhibit Jevons dynamics at scale. You'd need AI-native competitors to delay market entry until incumbents realize their savings. You'd need the political and legal friction around large-scale layoffs to dissolve — simultaneously, across multiple jurisdictions.
None of those conditions are present. All four are moving in the wrong direction.
The market priced AI as near-certainty and applied zero margin of safety against Jevons. That's not an investment thesis — it's a bias stack in a trench coat. The infrastructure layer collects either way. The AI-native challengers compound their structural advantage every quarter. And every incumbent doubling down on AI strategy without a credible path to headcount reduction is widening the gap they cannot close.
The Bayesian update is coming. The question is whether you're positioned before it, or after. When a narrative this well-fortified finally corrects, it doesn't move gradually — it moves all at once. Six compounding biases on the way up means six compounding biases unwinding on the way down.
The analytical structure above draws from a behavioral forensics intelligence layer currently in R&D at Dealithic — a systematic framework for diagnosing decision failure patterns in financial contexts before they repeat.