The same basic chart, which depicts the deployment of AI technologies, a reduction in staff, and a decline in cost per output, is currently being displayed in conference rooms throughout America’s biggest corporations. The presenter gives a nod. The execs give a nod. Everyone agrees that in 2026, efficiency will look like this.
Goldman Sachs would want a word after handling the macroeconomic data with the same level of patience that large investment banks often use to dampen the enthusiasm of their own clients. Without the diplomatic pretense, the conclusion is that approximately $450 billion has been spent on AI infrastructure and hardware, with virtually no contribution to real economic growth. Not disappointing. Not humble. Insignificant.
The natural tendency is to write this off as the kind of unconventional viewpoint that garners attention because it goes against the wave of AI optimism that has pervaded every investment pitch, keynote address, and earnings call over the last three years. However, the Goldman argument holds up when examined closely, and the explanation—rather than the headline figure—is what makes it most intriguing. The sophisticated computer hardware that powers American AI operations is mostly imported.
The majority of the capital that goes out of the domestic economy when a business purchases a cluster of Nvidia GPUs is recorded by official macroeconomic accounting as a drag rather than a contribution. The funds were used. Other people received the economic credit. Even though the quarterly earnings reports don’t accurately reflect that math, the GDP statistic does.
The Goldman analysis’s J-curve argument is arguably the most historically sound, and it should cause naysayers to reconsider before pronouncing the entire wave of AI investment to be a waste. Not because electricity wasn’t useful, but rather because businesses needed time to rethink their workflows around what electricity made possible rather than simply bolting motors onto already-existing steam-powered machines, the electrification of American industry didn’t result in significant productivity gains for about thirty years after the technology became commercially available.
Similar trends were seen with the internet: massive investment in the 1990s, dramatic failure of many early bets, and then more than ten years before the real productivity shift showed up in economic data. AI might be in that same uncomfortable place between expenditure and rewards. It’s also feasible that the payout takes longer than anyone who is currently modeling it anticipates.
Due to its direct observability, the corporate behavior critique is the section of the Goldman analysis that causes the greatest immediate distress. Across industries, there is a clear trend of employing AI to cut headcount rather than increase worker competence. This trend is frequently declared with confidence, seeing cost reduction as a strategy rather than as the absence of one. A business can cut expenses by hiring seven workers instead of 10 when it uses AI to do a task.
Its ability to create new products, break into untapped markets, or provide the kind of innovative production that genuinely stimulates the economy has not necessarily improved. A different strategy was used by the companies experiencing real productivity gains, such as the 30% increases in software engineering and customer support deployments. They kept the employees, redesigned the process, and let the AI handle repetitive tasks so that the knowledgeable staff could concentrate on issues the AI was unable to resolve.
There is a sense that the business sector is making a repeat of the mistake made with previous technologies, focusing on the simpler statistic (cost per person) rather than the more difficult one (value created per dollar of expenditure), as this dynamic is being observed across industries. The infrastructure is genuine.
The ability is genuine. The GDP data already provides an answer to the question of whether the organizations implementing it are using it in ways that will eventually aggregate into economic growth, and the current response is not promising. It’s still unclear if the investment thesis necessitates a more thorough reconsideration of what AI is truly being asked to perform, or if the J-curve logic holds and everything is resolved in ten years.
