The Profits That May Never Arrive
What If AI’s Productivity Gains Aren’t Enough to Justify Its PricePhoto by Igor Omilaev on UnsplashWhat if AI actually delivers real productivity gains, but never enough to justify what labs need to charge to stay solvent?For a while now, everyone’s been arguing over AI’s economic impact. Is it…
What If AI’s Productivity Gains Aren’t Enough to Justify Its PricePhoto by Igor Omilaev on UnsplashWhat if AI actually delivers real productivity gains, but never enough to justify what labs need to charge to stay solvent?For a while now, everyone’s been arguing over AI’s economic impact. Is it real? How big is it? What does it do to GDP? Virtually every economist, consultancy, and central bank has tossed out a number. But the sheer gap between these forecasts tells you something is deeply wrong with the question.On one end sits Daron Acemoglu, who ran the numbers in 2024 and projected a modest total factor productivity boost of 0.5 to 0.7 percent over ten years, «disappointing relative to the promises,» as he put it. On the other end, Goldman Sachs threw out a $7 trillion figure, McKinsey floated up to $25 trillion (traditional AI and generative AI), and some tech optimists even predicted GDP would double in a decade if AGI arrives on schedule. Forecasts that differ only slightly, right?Why does this matter so much right now? Because ever since AI companies started asking for staggering sums of capital to keep building, this stopped being an academic debate and became a question of survival.The frontier labs can only pay back their investors if they generate massive revenue. They’ll only get that revenue if companies adopt AI en masse. And companies will only adopt AI en masse if they see immediate, undeniable savings or productivity gains that boost their own bottom line.In short: the gains for businesses must be big enough to justify the high prices the labs need to charge in order to pay back their investors. Until then, the whole sector is a giant gamble.See Nvidia, for example: it invests in the very companies buying its chips, backs mega-deals using future chip supply as collateral, and essentially lends customers the money to buy its products. Critics call it «circular financing», a setup where the same dollar gets counted as revenue at multiple stops. Jensen Huang calls the charge «ridiculous,» insisting end-demand is genuine. But Wall Street remains unconvinced.The most sophisticated defense of AI’s slow financial payback comes from Erik Brynjolfsson. His «Productivity J-Curve» theory explains that transformative technologies always look underwhelming early on. Companies pour money into intangible setup costs that don’t show up in standard GDP data. Productivity looks flat right until everything clicks – and then it suddenly skyrockets.By early 2026, amid growing market impatience, Brynjolfsson claimed the curve was finally turning upward, attributing a 2.7% US productivity jump in 2025 also to AI. Maybe he’s right. Or maybe his model simply found what it was built to look for.But set credibility aside. What if a huge chunk of the value AI creates simply stays diffuse? Not delayed, as the J-curve seems to teach, but just impossible for the creators to capture as profit?There are two main reasons for this.First, as Brynjolfsson himself documented with free digital goods like Wikipedia or digital maps, some tech creates immense consumer surplus that never shows up in GDP or corporate profit margins. The creators simply lack the pricing power to charge for the full value they provide.Second, and much more importantly, AI might boost efficiency at an individual company, but never enough for that company to pay a high fee to the AI provider. A fee high enough to cover the labs’ need to recover their investments.This is not impossible at all. On the contrary, it is a mechanism we already see at work in many situations involving highly beneficial things that are managed and paid for ( in one way or another, at least partially) by the state.Take transportation and electricity networks. Their contribution to productivity is enormous and beyond dispute; railroads, power grids, and telecommunications lines built the industrial economy. Yet no private infrastructure network has ever recovered its fixed construction costs by charging users a price set by market competition.Left entirely to market forces, the infrastructure would still be useful, but no price that kept it affordable to users would easily cover the capital that built it. So instead, society found a way to socialize the cost: rate-regulated utilities, nationalized networks, or regulated monopolies granted in exchange for capped prices.Why might AI be similar? Because the same general utility and cost structure applies. Building frontier models and the data centers that run them requires enormous fixed investment, while serving one more user costs almost nothing.Put that structure into a market where several labs are racing each other on price and capability at once, and the price gets bid down toward marginal cost, near zero, long before anyone recovers the capital that went into building it. If this dynamic applies to AI, frontier labs may never earn enough revenue to recoup their massive infrastructure investments.Or take education. The benefits of an educated workforce to the economy as a whole can exceed the returns that individual firms can capture (and are willing to pay for). As a consequence, private payments may be insufficient to sustain its provision.This is one economic rationale for extensive public funding: firms gain access to educated workers without bearing the full cost of their education. AI could follow a similar pattern too, as long as, say, the advantages for the individual firm tend to erode quickly because of the collective raising of the bar driven by the very spread of the instrument, or competition tends to quickly make the solutions adopted by firms outdated.The debate then completely shifts. It’s no longer about whether the state should claim a share of tech giants’ future profits, a favorite topic among policy think tanks. It becomes a question of whether the state will have to subsidize these systems. If AI becomes essential infrastructure for society, keeping it running and accessible becomes a public interest.Meanwhile, everyone is fighting for a share of future profits that may never materialise. Unlike in Europe, private US firms aren’t required to publish open balance sheets. Right now, these labs look less like profit engines and more like cash-burning furnaces, but no one could ever know for certain, because public scrutiny of their accounts is impossible.Maybe it’s time to ask a different question: Given how systemically important these companies have become, shouldn’t they be required to open their books to the public?Until we know how much these labs are really burning (or not) and at what rate, we can’t even seriously discuss whether they should be the ones subsidizing the state in a dystopian post-AI society, or whether the state will end up subsidizing them.SourcesAcemoglu, D. (2024). The Simple Macroeconomics of AI. NBER Working Paper No. 32487.Briggs, J., & Kodnani, D. (2023). Generative AI could raise global GDP by 7%. Goldman Sachs Research.McKinsey Global Institute (2023). The economic potential of generative AI: The next productivity frontier.Brynjolfsson, E., Rock, D., & Syverson, C. (2021). The Productivity J-Curve: How Intangibles Complement General Purpose Technologies. American Economic Journal: Macroeconomics, 13(1), 333 – 372.Brynjolfsson, E. (2026, 15 febbraio). The AI productivity take-off is finally visible. Financial Times.This story is published under the Generative AI publication. Connect with us on LinkedIn and follow Zeniteq to stay in the loop with the latest AI stories. Let’s shape the future of AI together!The Profits That May Never Arrive was originally published in Generative AI on Medium, where people are continuing the conversation by highlighting and responding to this story.Source: Generative AI Pub — Published — Category: Image AI