The US economy is growing more rapidly than other G7 nations, primarily fueled by massive investments in AI models, data centres, and related chips and technologies.
The US composite PMI, an indicator of economic activity, increased from 56 in August to 58.4, marking the strongest rise in private-sector growth since July 2021 and the fourth straight month of accelerating expansion. This uptrend was led by the service sector, which experienced its sharpest output growth in over five years, while manufacturing also picked up pace. New orders surged at the fastest rate since April 2022, and manufacturing employment saw its strongest expansion since February 2021.
Back in June, I commented that AI was just ‘one big trade for the US economy’, but by September this seems to understate its scale. The AI expansion is on track to become the largest economic investment in US history, far surpassing previous monumental projects like 19th-century railroads, the 20th-century highway network, and the 21st-century internet.
Experts predict that the combined capital expenditure of five major hyperscalers—Alphabet, Amazon.com, Meta Platforms, Microsoft, and Oracle—will reach $4.2 trillion between now and 2029, according to FactSet. Spending on data centres alone is set to exceed investment in canals, railroads, and the electrical grid combined, with forecasts by the Brookings Institution estimating $10.3 trillion in outlays from 2025 to 2032. This equates to an extraordinary average of 3.6% of GDP annually, reflecting an unprecedented reliance on a single sector for US economic growth.
Up through July, private investments in data centres have totaled $37 billion, with many facilities not yet operational.
By contrast, private construction spending on residential and commercial properties—including houses, apartments, and shopping centers—was approximately $46 billion lower than the previous year in the first seven months of 2026.
AI-related investments have generated about 750,000 new positions since 2023, based on LinkedIn data. These roles command impressive wages, with the median annual salary for AI positions at around $180,000, compared to $80,000 across all job listings.
Most notably, this AI boom has significantly bolstered stock market wealth. By the second quarter of 2026, US stock and mutual fund assets reached $63 trillion, almost double their value at the end of 2022, according to Federal Reserve data. However, the bulk of these gains have benefited the wealthy, since ordinary workers typically hold little stock or bonds.
International investors are increasingly allocating funds into US assets, now holding an all-time high of $39 trillion in stocks and bonds, rising steadily since 2022. This inflow supports the US dollar’s strength and boosts equity valuations. Geopolitical conflicts in Ukraine and Iran have driven foreign investors to seek safety and growth opportunities in the US market.
Meanwhile, the swelling demand for data centre components such as memory chips is raising prices for tech goods. Import costs for computers, peripherals (like hard drives), and semiconductors were about 20% higher in August compared to last year. These increased input costs contribute to the rising prices of consumer electronics—such as iPhones and gaming devices—and add to overall inflationary pressures.
However, there is a critical challenge: the gap between hyperscalers’ spending and available cash flow is widening rapidly. In 2027, the capital expenditures of Amazon, Meta, Microsoft, and Alphabet are expected to surpass $1 trillion for the first time, while their collective free cash flow—funds generated from existing business profits—is projected to drop under $100 billion. This contrasts with a year ago when free cash flow was about $200 billion against $300 billion in capex. AI spending is accelerating just as the cash to cover it is dwindling.
As this disparity grows, hyperscalers increasingly depend on debt and equity markets to fund their AI endeavors.
The concern is that if AI investments don’t yield adequate profits, stock markets may experience sharp declines as investors exit. Currently, US stock market valuations are highly inflated relative to earnings. The CAPE ratio, which measures price-to-earnings over time, is above levels seen before the 2008 financial crisis and approaches those from the 2000 dot-com bubble.
Will profitability emerge? According to research from Fathom Consulting, for the multitrillion-dollar AI surge to become profitable, AI-related revenue from these tech giants must grow by $600–800 billion within the next two years. Conversely, Panmure Liberum’s analysis suggests that current capital expenditure and revenue forecasts through 2030 imply a negative internal rate of return on invested capital for Alphabet, Meta, Microsoft, and Oracle.
Therefore, hyperscalers face a choice: drastically scale back AI investments to levels that ensure reasonable returns on existing capital, or hope for a tremendous profit surge driven by future demand. Cutting spending would signal to investors that AI’s promise is failing, likely triggering a sell-off and market crash. Consequently, companies feel compelled to continue increasing their expenditures.
Meanwhile, the price companies pay for AI computing usage, tracked by the LLM Token Expenditure Index, has plummeted to $0.97—the lowest since the index’s late 2025 inception and over 50% below its summer peak. This drop is due to cheaper models, competitive open-source AI from China, and reduced inference costs, making AI use increasingly affordable. While beneficial for adoption, this trend depresses revenue growth for AI labs, complicating efforts to cover infrastructure expenses.
AI companies like OpenAi and Anthropic maintain they will soon become profitable, enabling hyperscalers to reap substantial returns. However, many such claims rely on questionable profit projections. AI-related revenues increasingly consist of “other income”—such as contracts with fellow AI firms—which now account for 54% of pretax profits, flattening true earnings growth.
In fact, AI firms are staying afloat through a process known as ‘circular financing,’ where one company lends money to another, allowing the borrower to report profits. Sona Asset Management has analyzed the AI ecosystem, revealing a complex web of interdependencies where each player relies heavily on others to deliver results.
Sona’s findings also show AI companies’ revenues are closely linked to capital expenditure decisions made by one or two other major firms, indicating a structurally fragile system on the brink of collapse.
A central question remains: will AI significantly enhance US labor productivity enough to foster decades of economic growth? Anthropic plans to sell $100 billion in shares this November, valuing the company at $2 trillion. To justify this, it released a report claiming that if AI truly takes off, the US GDP could increase by 32% by 2030—implying annual growth rates up to 15%, far exceeding the existing maximum of about 2.5%.
This prediction appears wildly exaggerated, presuming AI boosts productivity massively as every American enterprise adopts AI tools while redundantly dismissing millions of workers.
Historically, automation advanced at roughly 2% of tasks annually over two centuries without pushing growth much beyond 2%. Previous ‘general-purpose technologies’ required decades to spread widely after becoming viable; for example, electrification took about 40 years to noticeably improve factory productivity. Similarly, Comin and Mestieri’s research reports average adoption lags of roughly 45 years globally, with recent technologies still requiring 7–18 years.
AI adoption seems to be happening faster than with PCs or the internet, but adoption is merely the initial stage. Productivity often initially declines while companies invest in complementary assets, a phenomenon termed the “productivity J-curve” by Brynjolfsson, Rock, and Syverson. Even some of the most optimistic insiders acknowledge this; for instance, Sam Altman recently admitted “I think I was wrong about a few things, but one of them, in terms of the speed, one of them is the economy just has so much inertia. […] we’ve all been too ambitious on timelines […] Society and the economy will adapt more slowly.”
Additionally, most jobs involve physical tasks, while AI advancements mainly boost cognitive abilities. Only about one-third of economic output relies on computer-based work (Epoch AI’s remote-work study). The remaining two-thirds of economic activity happens in sectors like mining, construction, food service, healthcare, and elder care. Automating these physical tasks by 2035 would require billions of robots capable of performing diverse manual work, which must be designed, built, installed, and maintained at an enormous scale within a decade. Although progress is occurring, robotics development lags behind software, and versatile, cost-effective robots capable of replacing human labor are not yet available.
To date, clear evidence of broad productivity gains is lacking; total factor productivity—a metric capturing effects of new technologies—is currently below trend.
AI use is rapidly increasing, especially among service-sector companies. In the past two years, AI adoption rose from 25% to 61% in services, and from 16% to 51% among manufacturers. However, only 17% of service employees and 7% of manufacturing workers actively use AI in their roles.
As noted by the Federal Reserve Bank of New York: “The next phase is crucial, as currently AI mainly accelerates writing, summarizing, coding, or analysis. In the future, AI agents and specialized tools might manage entire workflows. At that point, productivity and employment impacts could reach unprecedented levels. AI usage is soaring while corporate investment remains moderate, and the cost of AI capabilities continues to decline. This dynamic favors wide AI adoption and productivity gains but poses a more complex challenge for profitability across the AI ecosystem.” In essence, productivity may rise in the future, but this could come at the cost of diminishing returns for all participants, a classic capitalist dilemma.
Moreover, inherent flaws exist within AI Large Language Models. Studies by Oxford and Cambridge researchers two years ago demonstrated that LLMs trained solely on AI-generated content suffer from irreversible degradation. Each successive model learns only from the previous model’s output, and by the ninth generation, coherence collapses; for example, inquiries about medieval church towers yield nonsensical answers like lists of jackrabbits. This phenomenon, dubbed ‘model collapse,’ means that machine-generated data increasingly deteriorates subsequent AI quality.
Warning signs of a looming crash continue to grow. Rising interest rates coupled with falling bond prices could trigger a crisis. Alternatively, failed IPOs of companies like Anthropic or OpenAI, or stiff competition from ‘open weight’ LLM models originating in China and elsewhere, could erode profits. The AI boom represents the largest gamble in US economic history and carries considerable risk.
Original article: thenextrecession.wordpress.com
About the Author
