When Intelligence Becomes Infrastructure

Working Thesis 01. This paper distinguishes current evidence, inference, and forecast.

The premise

Most writing about AI and business asks the wrong question. It asks whether a company will adopt AI, how fast, and how much productivity it will gain. That question is real, but it does not decide durable advantage — and its usual answer is wrong in an instructive way.

"Everyone is adopting it" depends entirely on who you count. Among large, self-selecting respondents, McKinsey found 88% of organizations regularly using AI in at least one function. Among all U.S. firms, the Census Bureau found 18% — about 32% weighted by employment — and most adopters used it in three or fewer functions. Adoption is spreading quickly, but it is uneven and shallow. It is a race in progress, not a settled fact, and being in the race is not the same as winning it.

Start from a narrower premise than the headlines. What is collapsing in cost is not "intelligence" in general but one specific thing: competent cognitive output — draft text, routine code, summaries, images, bounded analysis. Frontier capability, hard judgment, and reliable autonomy may stay scarce for a long time. The claim is only that for tasks a capable model can already do, doing them stops being a source of advantage, because access to comparable capability becomes broadly available.

When a capability becomes cheap and broadly accessible, access alone becomes less differentiating, and value moves toward the complements that remain scarce. The comparison to electricity is a cliché — "AI is the new electricity" — and, taken loosely, a misleading one: infrastructure owners can be immensely profitable when scale, networks, or regulation stay scarce. The sturdier historical lesson is quieter. General-purpose technologies deliver their gains only after the organizations using them are rebuilt around them; electrification's productivity payoff lagged its adoption by decades, and the OECD makes the same case for generative AI today. The question is not only who supplies it, but who owns the scarcity around it.

So the strategic question is not, "Will you use AI?" It is: what becomes more valuable as competent cognitive output becomes cheap?

This document advances three working hypotheses:

1. AI compresses advantages built on routine cognitive production; who captures the value depends on the scarce complements surrounding that production.
2. In software, value shifts away from writing code and toward the things code sits on — problem definition, proprietary context, state, integration, verification, distribution, and accountability.
3. Durable companies pair products with positions that are hard to reproduce: trusted relationships, exclusive data, network participation, regulatory authority, and real-world execution.

I. What cheaper cognition does to a business

The input that collapses

Cheaper cognition is not a metaphor; it is a measured trend. The cost of reaching a fixed level of capability — say, GPT-3.5-level performance — fell roughly 280-fold between late 2022 and late 2024, and across the performance thresholds Epoch AI studied, inference prices fell at rates ranging from 9× to 900× a year, with a median near 50×. That is the cost of a fixed capability, not of the frontier, which has not fallen the same way. But for the band of competent, routine output, the direction is unambiguous, and three pressures tend to follow.

Deflation in the commoditized layer. Any business whose value was "we can produce competent output you cannot produce yourself" is under pressure, because competent output is getting close to free. That describes a large share of agencies, middlemen, basic analysis, routine content, tier-one support, and boilerplate software. In some bounded tasks the floor rises faster than the ceiling; elsewhere the effects remain uneven. Being merely competent, on its own, stops being sellable.

Abundance instead of scarcity. Cheaper cognition does not shrink these markets; it expands them — the Jevons paradox, the observation first made about coal that using a resource more cheaply tends to increase its total consumption. More gets written, run, made, and automated. But abundance destroys the value of any single unit, and businesses get pushed off "paid per artifact" toward "paid for outcomes, guarantees, or access."

Value migration. The profit does not vanish; it moves to whatever is now the bottleneck. This is the logic Joel Spolsky called commoditizing your complement: when one thing gets cheap, demand shifts to what it pairs with. When routine thinking is cheap, the scarce complements become proprietary and real-time context; distribution and the trust to be chosen; the authority to be accountable; and the ability to act in the physical, regulated world.

Why adoption alone wins nothing lasting

There is a real asymmetry in how this technology helps: it tends to raise the floor faster than the ceiling. The cleanest evidence is a customer-support study where AI lifted average productivity 14%, and 34% for the least experienced workers, with little effect on the best. But this is a tendency, not a law, and it is bounded. On tasks outside AI's "jagged frontier," a study of consultants found AI users 19 percentage points less likely to reach the right answer; in one field experiment, experienced open-source developers were about 19% slower using early-2025 tools than without them — while believing they had been faster. Where AI compresses the gap between competent and excellent, it erodes advantages built on being more competent than the next company. Where it degrades, it rewards knowing its edges. Either way, skill at doing the routine work is a weaker moat than it was.

That compression sets up the trap. Because rivals increasingly gain access to comparable baseline capabilities, AI is a rising tide that lifts a market's boats together — competed away as a source of advantage about as fast as it is adopted. The data already shows the gap: adoption is broad, yet in the same surveys only a minority of companies report a measurable effect on the bottom line. This is the Red Queen dynamic among organizations — running to stay in place. AI is a multiplier; what matters is what it multiplies. Applied to a commodity, it yields cheaper commodities. Applied to a proprietary position, it compounds the position.

The company changes shape

The rest of this section is forecast more than finding. Where one person plus AI can do what once took a team, the minimum viable company shrinks toward one and leverage per person rises. Expect more tiny, high-margin firms, and inside larger ones, fewer but more senior people.

That hides an emerging risk worth naming. The traditional path — a junior does the grunt work, learns by doing it, becomes senior — runs straight through the work AI now absorbs. Early evidence is consistent with pressure on that first rung: workers aged 22–25 in the most AI-exposed occupations have seen a roughly 16% relative decline in employment since generative AI's arrival, via reduced hiring rather than lower pay, and Big Tech new-graduate hiring is down more than 50% since 2019. Whether this compounds into a senior-talent shortage a decade out is a forecast, not a finding — the studies are associational and confounded by the post-2021 correction. But if the training rung is being pulled up, durable organizations will have to rebuild deliberately what used to happen for free.

II. What happens to software

Fifteen years ago the argument was that software was eating the world. The stranger turn now underway is that software is being eaten by the models that write it. For forty years its economics rested on one scarcity: writing correct software was hard, slow, and expensive, so whoever did it could charge for the artifact — licenses, then seats, then subscriptions — for years. That scarcity is thinning, and things downstream of it move.

Code production becomes less scarce

Be precise about the claim: code production is getting cheaper, which is not the same as working software being free. In late 2024, more than a quarter of new code at Google was AI-generated, then reviewed and accepted by engineers; by early 2025 the company put the figure past 30%, though firms define "AI-generated" differently. Microsoft's CEO put it at 20–30% of the code in its repositories. Yet the same METR experiment that found experienced developers slower with AI is a caution against triumphalism: generating plausible code is cheap, and getting correct, maintainable systems is still work. The defensible statement is that the cost of producing software is falling fast — not that the artifact is already worthless.

That is enough to matter. "SaaS" in its purest form — we wrote the code, we rent you a seat — was justified by writing the code being the hard, payable part. As that premise erodes, so does the pretense underneath it: for most software businesses, the code alone was rarely the moat. It was switching costs, data gravity, network effects, distribution, and brand — assets software companies accumulated while pretending the code was the thing. AI forces each company to answer honestly what it was actually selling.

Where this likely goes

The rest of this section is inference, not observation — plausible extrapolation of the same logic, offered as hypotheses to test rather than facts to bank.

Software may drift from capital equipment toward something more like content: abundant, contextual, and short-lived, assembled for a particular user, task, and context, then discarded. If so, the "product" abstraction weakens, because you cannot sell features that are free to generate.

The interface may invert. For decades humans operated software; increasingly, software acts on behalf of humans — agents — and agents become the users of other software. Interface polish loses power where the user is a program that wants clean state, permissions, and machine-readable context rather than a dashboard. In its place a new scarce good appears: being legible and trustworthy to agents. If agents choose tools, read data, and execute transactions, the businesses that structure their data and capabilities for machine consumption — verifiable, real-time, permissioned — are the ones that get selected. Whether agents mature into reliable buyers on this timeline is exactly the kind of claim this thesis means to keep testing.

The strongest inference is where the premium moves. As AI lets code be produced faster than humans can review it, the verification burden grows: generated changes can be plausible, functional, and still wrong in ways that are expensive to discover later. Value moves to the boring, hard, trust-laden parts: verification, provenance, evaluation, security, compliance, uptime, integration, and data quality. "Anyone can generate it" makes "someone certifies it works, and is accountable when it doesn't" worth paying for.

The compressed version: when code is cheap to write, the value is in everything that isn't the writing — the state, the data, the integrations, the trust, the accountability, the distribution, none of which a model conjures from its weights.

III. What makes a business durable

Parts I and II converge on one test. For any asset a business holds, ask:

Does abundant, cheap cognitive output make this asset worth more, or worth less?

If more, it is a candidate for durable advantage. If less, it is a melting asset no matter how healthy today's income statement looks. Most incumbents hold a mix, and the work of the decade is to tell which is which and migrate value onto the compounding assets before the deflation arrives. It is early enough for that to be a choice; it will not stay one.

The melting advantages

These lose value as competent output gets cheap. Relying on them is the default failure mode, because they felt strongest in the old world.

• "We are more competent, faster, or cheaper at producing X." Competence in routine production is the thing being commoditized.
• Proprietary methodology or know-how. If it can be fully described, a model can learn it. Secret sauce that lives in documents and heads is exposed.
• Information asymmetry — being the expert who knows what the customer does not. Models democratize expertise and dissolve the asymmetry many advisory and professional-services businesses were built on.
• Aggregating public information. The classic aggregator play gets disintermediated when a model can gather and synthesize the same public sources directly.
• Interface polish as the primary differentiator. Valuable while a human is the user; shrinking if agents take over.

The compounding advantages

These hold or gain value — and, tellingly, most were never about doing the work. They are about position, relationship, and trust, which is why AI leaves them standing.

• Proprietary, real-time context. The strongest candidate, because cheap cognition raises demand for the one thing it cannot synthesize: data off the public internet, continuously refreshed, tied to real-world behavior. But data is a moat only under conditions — it must be exclusive or costly to reproduce, legally usable, causally tied to better outcomes, and continuously renewed. Public or purchasable data is not a moat; it can be copied.
• Distribution and trusted relationships. When production is free, being the default a customer — or a customer's agent — reaches for is the bottleneck. Attention and trust do not commoditize.
• Network effects and marketplaces. Value that comes from other participants cannot be generated from a model's weights; it is assembled in the world, one relationship at a time.
• Accountability. Someone must stand behind the outcome and carry the liability. That role grows scarcer as output becomes abundant and unaccountable by default.
• Physical and regulated execution. Atoms, logistics, licenses, and trusted action in high-stakes settings sit past the API boundary; they are protected, and the operations around them get cheaper — the good kind of AI exposure.
• Taste and judgment. When execution is cheap, direction is everything. Knowing which of infinite cheap outputs is the right one does not commoditize, because it is the act of choosing, not producing.

Positions, not products

The unifying idea: a durable business owns a position, not a product. A product is a snapshot, and AI makes snapshots cheap and disposable. A position is a place in the system — a system of record, a network, a trusted rail, a working data loop — that keeps generating value as the products on top of it churn.

The strongest version is the data flywheel: usage generates proprietary data, the data improves the product, the better product wins more usage. But a flywheel is a moat only when each turn is genuinely exclusive and hard to reproduce — otherwise it is just a diagram. Its power is conditional on the data qualifiers above, and most claimed flywheels fail one of them.

Working conclusion

AI will not split companies into adopters and non-adopters. It will split tasks by where the technology actually lowers cost, and firms by whether they can convert those cost reductions into positions that are hard to reproduce.

Some advantages built on routine production will erode. Others — exclusive context, trusted distribution, integration, network position, accountability, real-world execution — may become more valuable. None is automatically a moat; each has to stay useful, scarce, and hard to copy under stronger AI. For software, the likely transition is not from valuable code to worthless code, but from code as the dominant bottleneck to a wider set of them: deciding what should exist, supplying the context, holding the state, verifying the behavior, integrating the systems, and standing behind the result.

That is the research program this thesis implies, and the one we intend to pursue: measure where the cost of cognition is actually falling, watch where value moves in response, and find which complements stay scarce long enough to hold an advantage.

Sources & notes

The argument is reasoned; the figures below anchor its load-bearing claims. Where a number is contested or provisional, we say so — the caveats tend to strengthen the case.

Adoption and cost.

1. U.S. Census Bureau — Bonney et al., The Microstructure of AI Diffusion (CES-WP-26-25, 2026): 18% of U.S. firms use AI in a business function (≈32% employment-weighted); 57% of adopters use it in three or fewer functions. The all-firms counterweight to McKinsey.
2. McKinsey & Company, The State of AI (2025): 88% of (large, self-selected) organizations use AI, yet only a minority report a measurable bottom-line effect — the clearest evidence for the Red Queen point.
3. Stanford HAI, 2025 AI Index: cost to reach GPT-3.5-level performance fell ~280× (Nov 2022–Oct 2024). This is cost-to-fixed-capability, not the frontier.
4. Epoch AI, LLM inference price trends: inference prices fell 9×–900× a year across benchmarks (median ~50×), "rapid but unequal"; the fastest drops may not persist.

Productivity — and its limits.

5. Brynjolfsson, Li & Raymond, Generative AI at Work (NBER w31161, 2023; QJE 2025): +14% average, +34% for novices, in customer support. The "raises the floor" evidence.
6. Dell'Acqua et al., Navigating the Jagged Technological Frontier (HBS 24-013, 2023): large gains inside AI's frontier; 19 points less likely to be correct on a task outside it.
7. METR, Early-2025 AI and experienced open-source developers (2025): experienced devs ~19% slower with AI in mature repos. METR has since flagged the result as historical and is redesigning its method; it does not claim AI now speeds developers up.

Software.

8. Sundar Pichai, Alphabet Q3 2024 earnings call: more than a quarter of new code is AI-generated, "then reviewed and accepted by engineers"; put past 30% in Q1 2025. Definitions of "AI-generated" vary by company.
9. Satya Nadella at LlamaCon (Apr 2025), via TechCrunch: 20–30% of Microsoft's code (an approximate, spoken figure).
10. Marc Andreessen, Why Software Is Eating the World (WSJ, 2011).

Labor.

11. Brynjolfsson, Chandar & Chen, Canaries in the Coal Mine? (Stanford, 2025): relative employment decline for ages 22–25 in the most AI-exposed jobs — ~13% in the August 2025 draft, revised to 16% in the later version — from ADP data. Associational, not causal.
12. SignalFire, State of Tech Talent 2025: Big Tech new-graduate hiring down more than 50% since 2019; the share of top CS graduates landing at the largest firms down more than half since 2022. Proprietary data; AI is one cause among several.

Concepts.

13. Jevons paradox — W. S. Jevons, The Coal Question (1865): improving the efficiency of a resource's use tends to raise, not lower, its total consumption.
14. Commoditize your complement — Joel Spolsky (2002).
15. Red Queen — Leigh Van Valen (1973); applied to organizations by William Barnett (2008).
16. General-purpose technology and the productivity lag — Paul David, The Dynamo and the Computer (1990); OECD, Is Generative AI a General-Purpose Technology? (AI Papers No. 40, 2025); "AI is the new electricity", Andrew Ng.