The Marginal Cost of Intelligence

Tue, 15 Sep 2026 00:00:00 GMT

I was at a hotel bar in Dallas with a vendor account executive who was trying to sell me a four-hundred-thousand-dollar AI deployment. He had three slides ready, and they were all about the model. The model’s benchmarks. The model’s context window. The model’s fine-tuning roadmap. I asked him what the four hundred thousand dollars was actually for. He paused, drank his water, looked at his deck, and said: “Eighty thousand is the model and tooling. The rest is services.”

That was the sentence I had come for. The model was twenty percent of the price. The other eighty percent was labor. We sat at the bar for another forty minutes while I asked him to walk me through what that labor actually was, and he could not. The deck did not have a slide for it. The proposal did not have a line item for it. The vendor’s website did not describe it. Eighty percent of what he was selling me had no marketing material. It was a folk craft, sold inside a fixed-price contract.

That conversation was when I stopped thinking about AI procurement the way I had been trained to think about software procurement. The thing he was selling me was not a model. The model was the cheap part, the part the vendor put on the slides because procurement knew how to evaluate it. The thing he was actually selling was a bundle of integration labor, governance work, and change management whose price had little to do with the work itself, and whose cost to me would depend on something neither of us yet understood: not what I was buying, but where I sat in the economy.

The marginal cost of intelligence depends on who is buying. This is the strangest thing about the AI economy, and procurement has not yet noticed.

For a decade, procurement learned a useful lesson from SaaS. The cost of software was decomposable into known parts: license, support, integration, training. The license dominated the bundle. Vendors competed on license price. As scale increased, volume discounts arrived. The seventeenth seat cost less than the third in real terms. Marginal cost approached average cost, and average cost approached zero. The whole apparatus of enterprise software procurement was built on that curve.

AI does not have that curve. AI has four layers, and they behave nothing like the SaaS stack. The model layer (the capability itself) is approaching commodity pricing; you can rent a frontier-class model for the cost of a sandwich per million tokens. The integration layer (the labor of making the capability work inside a specific environment) is bespoke handwork that resists standardization. The governance layer (the compliance, audit, security, and policy work) is regulated overhead whose cost is set by the buyer’s industry, not the buyer’s size. The change-management layer (the labor of getting humans to actually use the system) is sociological work whose cost scales with the buyer’s politics.

When you sign an AI contract, you are buying a bundle of all four. The vendor prices the bundle as a single number. You pay the single number. You think you are buying a model. You are not. You are buying integration labor from the vendor’s services team, governance work the vendor’s compliance staff will perform alongside your audit committee, and change-management labor your own people will absorb whether the contract mentions it or not. The model is the smallest layer in the bundle. The model is also the only layer the marketing talks about.

This is the foundational error. The bundle is not a unit. It is a stack of differently-priced labor, and the labor coefficient on each layer depends on who is buying.

Call this the friction tax.

The friction tax is the cost imposed by your own organizational position before any work is done. An enterprise pays a higher friction tax than a mid-market company because it has more systems to connect, more departments to coordinate, more audits to pass, and more vested interests to placate. The vendor bakes the friction tax into the price, because the vendor must staff the project to your scale. The buyer assumes the price reflects the value of the work. The price reflects the friction of doing the work at the buyer’s position.

The friction tax was invisible in the SaaS era because it was small. The marginal cost of installing one more user onto a SaaS product was negligible regardless of who you were. With AI, the marginal cost of deploying one more useful capability into a real workflow is dominated by integration labor, and integration labor is dominated by the friction tax. A capability that costs ten dollars in raw model spend can cost ten thousand dollars to deploy at the right buyer, and ten times that at the wrong one. The friction tax is the multiplier between the model cost and the deployment cost. It is also the most variable quantity in the AI economy.

The strangest consequence is that scale has inverted.

In every previous wave of business software, scale was an advantage for the buyer. The larger you were, the better the deal, the lower the unit cost, the more leverage at renewal. With AI, scale is an advantage for the vendor and a disadvantage for the buyer. The enterprise pays more in absolute terms (which the vendor wants) and pays more per unit of useful deployed capability (which the procurement department does not see). The mid-market pays less in absolute terms and substantially less per unit of useful capability, but does not buy, because it benchmarks against the enterprise price and concludes AI is unaffordable. Both sides of this misprice are wrong.

In the SaaS era, scale rewarded the buyer. In the AI era, scale punishes them. Procurement has not internalized this because doing so would require unlearning twenty years of received wisdom about leverage.

Twenty years ago Paul Graham wrote an essay called The Power of the Marginal. He was arguing about who creates new things, and his answer was that “great new things often come from the margins, and yet the people who discover them are looked down on by everyone, including themselves.” He then enumerated the structural disadvantages of being an insider: “the selection of the wrong kind of people, the excessive scope, the inability to take risks, the need to seem serious, the weight of expectations, the power of vested interests, the undiscerning audience, and perhaps most dangerous, the tendency of such work to become a duty rather than a pleasure.”

Graham was writing about creators of new technology, and his marginal player was a lone hacker with a laptop. The cost of being on the margin, for that player, was nothing. The cost of being inside the funded center was significant. The marginal creator won by default whenever those costs diverged.

Twenty years later the substrate has changed but the dynamic has not. Replace “creator” with “buyer.” Replace “lone hacker” with “operator of a five-hundred-person regulated services business.” Now read Graham’s list again. The selection of the wrong kind of people: enterprise AI is procured by staff whose career incentives reward caution, not deployment. The excessive scope: every enterprise AI project starts with a steering committee and a phased rollout. The inability to take risks: a public AI failure is asymmetric, because a small win is invisible and a small failure is a board-level event. The need to seem serious: the contract bakes in the cost of being seen to take AI seriously, which is mostly governance theater. The weight of expectations: the deployment is supposed to deliver enterprise-level outcomes from day one. The power of vested interests: every workflow being augmented has a budget owner who will be measured on the outcome. The undiscerning audience: the deployment is judged by executives who do not use it. The tendency of work to become duty: by the time the system is live, the people who would have championed it have moved on.

Graham’s catalog of insider disadvantages is a description of what creators face inside a funded center. Read with one substitution it is a description of what buyers face inside a funded center. Same dynamic, new substrate.

There is one thing Graham could not have seen in 2006. His marginal creator could work alone. The marginal deployer of AI cannot. The infrastructure required to deploy modern intelligence—the data plumbing, the integration scaffolding, the governance layer, the entity resolution underneath—has to be built by someone, and a five-person startup cannot build it alone the way a five-person startup could build a web application in 2006. The marginal player today is not the lone hacker. The marginal player today is the operator who has enough scale to fund a real integration but not so much scale that the friction tax destroys the unit economics. The sweet spot is narrower than Graham’s sweet spot, and it is structurally different. Graham’s marginal creator lived at zero. Today’s marginal deployer lives in the middle.

This is where the marginalist’s argument has to update Graham’s. The marginal creator could win alone. The marginal deployer needs scale, but not too much scale. The new marginal player is somewhere between the lone hacker and the enterprise, and that somewhere is the mid-market.

The economic shape of the opportunity follows from this. Between commodity SaaS pricing (which is too thin to fund the integration labor a real deployment needs) and enterprise AI pricing (which is too heavy because it amortizes labor the mid-market does not require), there is a missing pricing tier. The size of the gap is most of the economy. There are roughly two hundred thousand US companies with annual revenues between five million and five hundred million dollars. Almost none of them are well-served by current vendor pricing. The enterprise tier is overpriced for their position; the commodity tier is under-supported. The most efficient available buyers of AI labor in the economy, the buyers with the lowest friction tax, are not buying.

Three things follow if you accept the argument.

First, if you are an operator, stop pricing AI against vendor quotes. Price it against your own friction tax. If your integration surface is small, your friction tax is low, and the dollar you spend on AI buys you more useful intelligence than the same dollar at a company ten times your size. Deploy aggressively while the misprice persists, because it will not persist forever.

Second, if you are a vendor, stop building mid-market editions of enterprise products. The mid-market is not a smaller version of the enterprise. It is a structurally different buyer with a different cost shape. The pricing model that wins this tier is more like a managed service with light governance and heavy integration than like a self-serve product with discount licensing. The integrator who captures it will be more profitable than the platform vendor.

Third, if you are an investor, the largest available margin in applied AI over the next five years lives in the integration layer for the mid-market buyer. Not in the frontier model layer. Not in the consumer layer. In the awkward space between commodity SaaS pricing and enterprise AI pricing, where two hundred thousand US companies are waiting to be priced correctly. That space is most of the economy, and it is currently mispriced in both directions.

The deepest claim of the marginalist is this. In the SaaS era, the marginal cost of one more user was nearly zero, and the price reflected it. In the AI era, the marginal cost of one more deployed capability is dominated by labor, and the price does not reflect it. The friction tax is the multiplier between what AI looks like on the marketing page and what it costs in your environment. The friction tax is paid by every buyer. It falls hardest on the buyers who think they have the most leverage.

The next dollar of AI spend at a five-hundred-person mid-market company buys more useful intelligence, in real terms, than the next dollar at a fifty-thousand-person enterprise. Procurement does not believe this. The pricing pages of every major vendor say the opposite. Procurement and the pricing pages are both wrong, in a way the market has not yet repriced.

The marginalist’s bet is on the awkward middle. It is the most efficient buyer of intelligence in the economy, and it does not yet know it.

The Marginalist

The Marginal Cost of Intelligence