What Gets Obscured
Every wave of technological change manufactures the same illusion: that the underlying laws have been transcended. But underlying laws are never transcended, only obscured. The internet spent twenty years perfecting the attention economy, perfecting it so thoroughly that an entire generation of practitioners took it for commerce itself — acquisition, retention, conversion, scale first and monetize later. But the attention economy is not the whole of commerce. It is one particular expression of commerce under one particular cost structure. Outside its glare, the more fundamental things — what the thing you sell costs to make, why consumers pay, how supply and demand constrain growth — never went anywhere. People simply stopped looking at them.
AI arrived, and the same illusion appeared again: this time the claim is that AI will redefine products, redefine interaction, redefine everything. The AI product teams at every large company are performing the same motions — drive DAU, work retention, run the funnel, scale first and monetize later. This methodology rests on an unstated premise: that AI products and internet products share a cost structure.
“Scale first, monetize later” is not universal commercial wisdom. It is a highly specific strategy, parasitic on a highly specific cost structure. The essence of that cost structure is this: software’s marginal cost approaches zero. Code written once runs a hundred million times; serving one more user generates almost no incremental cost, so you can run at a loss to grow users and, once scale arrives, every additional unit of attention is pure profit. The KPI language of DAU, retention rate and conversion funnel was tailored for the zero-marginal-cost model — it measures the efficiency of aggregating attention, and aggregating attention is worth doing precisely because each additional unit costs almost nothing.
AI’s cost structure runs the other way. Every user call generates real inference cost; every token consumes compute. This is not the cost structure of software, it is the cost structure of manufacturing — serve one more customer and COGS rises, concretely. Someone will point out that token costs are falling fast, which is true, but however low they fall they remain a per-request variable cost and can never become “write once, run infinitely.” This is not a difference of degree, it is a structural difference: software’s marginal cost curve tends to zero, AI’s tends to a positive asymptote. That structural difference is why “scale first, monetize later” does not work arithmetically for AI. The bigger you scale, the more you burn — and unlike the internet, where marginal cost was zero to begin with, so that past a certain scale fixed costs amortize away, average cost approaches zero, and scale itself is the cure. AI’s marginal cost is continuous and rigid. Scale is not a cure.
AI’s commercial position is closer to electricity. Nobody consumes electricity as a finished product; people consume the things electricity drives — light, machinery, computation. Power companies sell electricity to factories and enterprises, and factories sell products to end users. Electricity’s business model was metered from the first day and priced against the value it creates, and no logic of “give the power away free first and work out monetization later” ever existed (determined by cost structure, not by industry convention). AI’s natural position may be the same: embedded in an enterprise’s production process, amplifying output directly, charging against the value created.
Some also hold that AI, beyond supplying productivity, is a consumer good. Hard to agree. Consumer goods satisfy desires: entertainment, sociality, aesthetics, identity. Productivity tools satisfy needs: finish a task faster, more accurately, more cheaply. AI’s core capability is processing and generation, which is a productive act. When you are in the consumer-goods business, embedding AI is never the core act; the core is the least sexy traditional model there is — create demand and sell a product. Even the usages that look most like consumption — chatting with an AI, asking it for a joke — offer an experience that is thin next to real consumer goods: against TikTok on entertainment, against WeChat on sociality, against games on immersion, AI does not compete on a single consumer dimension. The one dimension on which it crushes traditional tools is exactly productivity: one person plus AI does the work of three. That is not a consumer product’s value proposition. It is a production tool’s.
Technology changes. The discipline of cost structure does not. The constraints of supply and demand do not. The underlying logic of why a consumer pays does not. Every wave of new technology promises, with a smooth tongue, that this time is different; and after every wave, the ones left standing are those who obeyed 1+1=2.