How much productivity is needed to justify the AI trade?

There is a puzzle at the heart of the American economy right now. The economy is growing at a respectable pace—just below 2%—yet consumer sentiment does not reflect it. 

A handful of mega-cap firms’ investments, based on a single technology, have been driving growth narrowly—rather than creating a broad-based expansion that would lift a large segment of society. Underpinning these investments has been an assumption: that AI will lift productivity across the wide swaths of society and the economy—restaurants to retail to healthcare—eventually lifting GDP growth substantially. So far, AI has not done so, although it has greatly elevated the S&P 500.

The chart below captures the divergence in a single picture.

The dark blue line is the stock market, which continues to trade near all-time highs at valuations that are historically elevated. The light blue line tracks consumer sentiment, which has been mostly declining for three years. Stocks and sentiment have decoupled in recent years, after moving roughly in tandem for much of history (Exhibit 1), and that gap needs an explanation.

The answer lies in today’s artificial intelligence (AI)-driven economy. Booming data center construction and tech-related investment is boosting overall growth, but that growth is felt narrowly across the economy. Here, we aim to explain how the AI economy works and what needs to happen for the boom to continue. 

An economy whose engine is quietly and rapidly changing

Beneath the disconnect sits an economy still growing but whose drivers have shifted materially over the last 12–18 months.

Consider the American household first. It has been battered by repeated inflationary shocks in recent years. Successive bouts of supply-side inflation have left Americans’ incomes trailing prices, and households have responded by drawing down their savings (Exhibit 2).

This drawdown is not sustainable: Households cannot indefinitely spend more than they earn, and once the savings cushion is exhausted, consumer spending is likely to pull back, unless income growth begins to outpace inflation again. The Iran conflict and the associated spike in oil prices are tied to much of the recent decline in real household income growth. If that conflict abates and global oil prices decline and stay lower, household purchasing power would recover to some degree.¹

If consumers are not carrying the economy, what is? Corporate investment. But not all corporations: High-tech capital spending is running at cycle-high growth rates while low-tech spending is contracting in real terms (Exhibit 3). Spending in the housing, commercial real estate and transportation equipment sectors is declining, reflected in the low-tech spending line in the chart. The topline capex figure today looks strong nonetheless because a large volume of spending on data centers and the components that go in them, undertaken to train and run AI models, is offsetting weakness elsewhere.

In short, on close inspection, the investment boom of the mid-2020s is a bet placed by a handful of mega-cap firms on a single technology. This capex is a narrow growth driver rather than a broad-based expansion, consistent with how central the AI trade has been to stock market returns overall in recent years.²

Will this capex accelerate growth? All eyes should be on productivity

We think the question of how AI capex will impact growth warrants the most attention. Much of corporate earnings expectations, and the willingness to fund data centers as though demand for machine intelligence were unlimited, rests on a single premise: that AI will lift productivity—not at software firms alone but across the entire economy.

Software companies have indeed experienced impressive productivity boosts from AI. Some task-specific productivity gains exceeded 30%.³ But the American economy is not a software company. To move national productivity meaningfully, AI has to deliver gains in hotels and hospitals, in restaurants and retail stores, and across the broad base of the real economy. So far, the evidence isn’t there.

To be sure, there was a notable uptick in economy-wide productivity during the pandemic, a byproduct of remote work and a reshuffling of the labor market.⁴ That improvement, though, has since faded, and recent data show productivity growth returning toward its subdued pre-pandemic trend (Exhibit 5).

The economy’s return to a sluggish productivity trend matters because the current AI capex cycle is enormous by historical standards: Estimates put AI-related investment at above 2% of U.S. GDP—a scale of concentrated, technology-driven spending rarely seen outside major historical infrastructure build-outs. AI capex underpins the entire case for public and private market optimism. If the touted macro productivity gains do not materialize, the justification for high valuations across the AI ecosystem weakens.

Furthermore, if valuations were to decline, that would likely weigh on the real economy through the wealth effect operating in reverse.

How much does productivity need to accelerate to justify the AI trade?

Productivity needs to accelerate to justify the AI trade, but how much is a very hard question to answer precisely. However, a simple framework helps to clarify the math. We looked at the single most important AI company, Nvidia, and economic projections from the Congressional Budget Office (CBO).

Nvidia is the leader of the “AI trade,” with a current market capitalization of about $5 trillion—the highest valued company in the world. If we make a conservative assumption and say that Nvidia can grow its market cap by 5% per year for the next 10 years, that would bring Nvidia’s 2036 market cap to $8.55 trillion.

How much earnings would Nvidia have at that point? If we assume that in 10 years (in a steady-state equilibrium) Nvidia would have a 20x P/E ratio, that would imply about $430 billion in 2036 profits. (In 2026, Nvidia is expected to have a net income of about $210 billion.)

Now the CBO’s baseline projections come in: It forecasts total U.S. corporate profits in 2036 to be $5.5 trillion. That implies, given our analysis, that Nvidia’s share of total U.S. corporate profits in 10 years would be just under 8%.

That outcome would be a historical anomaly. In post-World War II history, the largest share of total profits earned by any one company was ExxonMobil’s 5%–6% share in 2008—which ultimately proved unsustainable because it was due to a bubble in the oil market.

We are left with a binary outlook: Either the CBO is wrong and total profits in 2036 end up (considerably) larger than what its baseline outlook is assuming, or growth in AI spending, and thus Nvidia’s outlook, slows more than expected.

How does this relate to productivity? The current CBO baseline assumes 1.75% labor productivity growth over the next 10 years. This allows us to back in to how much Nvidia’s share of total profits would fall under different, and higher, productivity scenarios (Exhibit 6):

So, to arrive at a more realistic Nvidia share of profits in 2036 (say, under 5%), we’d need to see productivity growth accelerate to above 5%. That is a high hurdle and has not, in fact, been achieved in post-World War II history.

There is one caveat we can make to this analysis: It might be too strict, or overly pessimistic, to assume that Nvidia will only achieve a normalized P/E of 20x in 2036. P/Es are structurally higher today in general, given more wealth relative to GDP in the global economy. So we can relax this and assume a potentially reasonable 30x steady-state Nvidia P/E in 2036. Exhibit 7 shows how that would change the math:

Now the hurdle rate falls to about 3% labor productivity, which seems more achievable. That would also be on par with the acceleration in productivity in the economy in the late 1990s and early 2000s during the internet revolution.

What is the read on recent productivity trends?

Is 3% productivity achievable? In the last three quarters, U.S. labor productivity has averaged about 1.3%, which is consistent with the low productivity backdrop of the 2010s rather than an AI-induced acceleration.⁵

There are also signs that the technology is struggling to scale, in practice, partly because AI tokens⁶ have quickly become expensive, prompting companies to ration consumption (or turn to cheaper Chinese AI models) rather than expand it. Even early proponents of AI are becoming more cautious. Chamath Palihapitiya, a notable tech investor and enthusiast, recently opined on CNBC in July that “You're starting to see a little bit of the wheels come off” when it comes to converting AI deployments into profits.

Additionally, growing anecdotal evidence suggests some companies that instituted AI layoffs expecting the technology to fully replace labor have been forced to make a U-turn and rehire after realizing the flaws in the technology and/or when they became aware they hadn’t fully understood the human element embedded in modern production.⁷

One last indicator: A 2025–2026 large-scale survey of business executives in the U.S., UK, Germany and Australia found that 89% of those surveyed reported no measurable impact on their business’s labor productivity from AI deployments in recent years.⁸

In sum, the economy appears to be still in (or stuck in) AI's experimental/ideation phase, rather than entering its productivity phase.

How long will markets tolerate a productivity revolution that has not yet appeared?

How booms actually end

History is fairly consistent on how booms come to conclusions: Transformational technologies routinely end in boom-and-bust cycles. The railway boom reshaped a continent but still ended in financial panic and depression. The Suez Canal revolutionized global shipping yet left its backers in a debt crisis. These technologies delivered enormous long-run benefits to society, but the financial rewards to those who built them out often proved insufficient.

Crucially, super-cycles rarely end because the technology stops working—the internet remained useful long after the dot-com bust. They end when the economics of building out the infrastructure stop adding up, through one of three channels: supply overwhelms demand, demand slows unexpectedly, or financing dries up.

Busts through each of these channels have occurred before, and each has left signs worth watching.

• Oversupply is the lesson of the late 1990s, when capacity was built far ahead of demand and much of it sat idle.
• Slowing demand is the lesson of the early 2020s, when spending kept accelerating even as revenue growth cooled, forcing a painful valuation reset in 2022.
• A financing squeeze is the most abrupt of the three, though the current AI cycle appears comparatively well-funded, for now.

The common thread is that these turning points are nearly impossible to identify in real time: At the peak, earnings look strong, construction pipelines are full and management commentary is confident. The inflection arrives quietly when investors stop asking how fast capacity is growing and start asking what return it is generating.

That brings us back to the two diverging lines we began with. The stock market is pricing in the expectation that this technology will transform the entire economy, and quickly. The consumer is living in an economy that has not yet been transformed. One of those lines will eventually move toward the other. Which one—and over what time frame—is the central question for the years ahead, and the outcome will depend entirely on whether, and when, the promised productivity gains finally materialize.