NextFin News - The United States economy expanded in the first half of 2025 for one reason above all others: artificial intelligence. Investment in information-processing equipment and software accounted for just 4% of gross domestic product yet drove 92% of GDP growth over that period, according to Harvard economist Jason Furman's reading of the data. Remove that spending and the economy was essentially flat, growing at an annualized 0.1%. Strip away the AI boom and America's expansion nearly vanishes. That is the wobbly foundation beneath one of the most concentrated bull markets in modern history, and it is why the question circulating through C-suites and investment committees has shifted from "is there an AI bubble?" to "what happens to everything else when the builders stop pouring concrete?"
The Scale of the Buildout
The numbers are difficult to overstate. The five largest hyperscalers — Amazon, Alphabet, Meta, Microsoft, and Oracle — are on track to spend roughly $600 billion to $725 billion on capital expenditure in 2026, nearly three times the approximately $238 billion deployed in 2024. CreditSights estimates the top five will commit about $602 billion in 2026, up 36% from roughly $443 billion in 2025, with approximately 75% of that — around $450 billion — directed at AI infrastructure. Company guidance points even higher: Amazon at $200 billion, Microsoft at $190 billion, Alphabet at $180 billion to $190 billion, and Meta at $125 billion to $145 billion. After the mid-2026 earnings round, tallies of the big four alone ran close to $725 billion.
The Bank for International Settlements put the two-year tab at more than $1 trillion for 2025 and 2026 combined, and noted that these commitments are outpacing the earnings and free cash flow of the firms making them, forcing some to issue debt to finance the buildout. Capital intensity has reached levels that would have been unthinkable a decade ago: 57% of revenue at Oracle and 45% at Microsoft, by CreditSights' measure. Oracle illustrates the strain: it raised $43 billion of debt and $5 billion of equity during fiscal 2026, including a record $25 billion bond sale, while free cash flow for the year was negative $23.7 billion against $32.0 billion of operating cash flow.
At the center of this spending stands Nvidia. In the fiscal first quarter ended April 26, 2026, the chipmaker reported record revenue of $81.6 billion, up 85% from a year earlier, with data center revenue of $75.2 billion — up 92% and equal to 92% of total revenue. Gross margins held near 75%. The market has rewarded that dominance: Nvidia's market capitalization stood at $5.38 trillion as of September 18, 2026, making it the single most valuable company on earth and worth about 7% of the entire S&P 500 on its own.
The S&P 500 stood at 7,650.5 that day, while the 10-year Treasury yield hovered near 5.01% — a rate that matters because it is the discount rate against which every future dollar of AI-generated cash flow must be measured.
The Economy Is Running on a Single Engine
The mechanism of the dependency is straightforward accounting, which is precisely what makes it dangerous. Gross domestic product is the sum of consumption, investment, government spending, and net exports. When business investment in AI equipment and software surges, GDP rises mechanically. The St. Louis Federal Reserve attributes 39% of third-quarter 2025 GDP growth to AI-related investment; the BIS estimates AI investment contributed roughly one full percentage point to real GDP growth in 2025. By some readings, including Furman's, the figure for the first half of the year was closer to 92%.
Investment is also the most volatile component of GDP. It is the accelerator pedal, not the engine block. A sector that represents 4% of the economy but supplies more than 90% of its growth is not a diversified expansion — it is a concentrated bet. When the investment cycle normalizes, growth normalizes with it, and the arithmetic works in reverse.
This is where the cyclical-versus-structural distinction must be drawn cleanly, because conflating the two produces the wrong conclusion. Artificial intelligence itself is a structural shift: a general-purpose technology that will reorganize production, labor, and productivity over decades. The evidence for that is in the technology's diffusion, not in this year's spending. But the pace of the capital expenditure is cyclical. It is a boom, and booms overshoot. The fiber-optic networks laid during the dot-com era became the backbone of the modern internet — a structural legacy — but the companies that overbuilt them in 1999 were mostly wiped out, and the economy endured a severe recession in 2001. The technology survived. The valuations did not.
The current cycle carries the same shape: a durable technological revolution financed by a transient spending frenzy. The risk to markets and the economy sits entirely in the second half of that sentence.
Why the Spending Keeps Climbing Even as Returns Lag
If the returns on AI investment remain unproven, why does the spending accelerate? The answer is not optimism alone. It is a prisoner's dilemma.
Each hyperscaler must keep building because if it does not, it loses the race for AI cloud share — and cloud infrastructure is the moat around the entire business model. The BIS observed that this investment race "may be partly driven by the perception that only a small number of" firms can win it. In game-theoretic terms, the dominant strategy for every player is to spend, regardless of whether the collective outcome is rational. Individual rationality produces collective overcapacity.
There is a second, more uncomfortable possibility that critics raise: some of the revenue flowing back to the hyperscalers is circular. Venture-funded AI startups rent compute capacity from the same cloud providers whose investors also fund the startups. When the same pool of capital finances both the shovel seller and the hole digger, the revenue that appears to justify the next round of capex is partly a reflection of the capex itself. It is not proof of end-user demand.
The revenue side has not kept pace with the spending. Hyperscaler capital expenditure is growing at 36% to 75% year over year depending on the measure — Goldman Sachs recorded 75% third-quarter growth — while AI-related cloud revenue, though expanding rapidly, is growing more slowly. That gap is the profitability question at the heart of the bubble debate: every dollar of capex must eventually earn its keep, and the longer the lag between spending and monetization, the more fragile the financial engineering becomes.
The financing structure, however, is not the fragility most investors fear. Unlike 2008, AI capex is funded primarily through hyperscaler free cash flow, corporate debt, and direct equity issuance. Private credit firms such as KKR and Apollo operate entirely outside the banking system. This distinction matters for the transmission of risk: losses would fall on institutional investors and high-net-worth individuals rather than on banks whose balance sheets amplify shocks across the economy.
The Market Has Priced Perfection Into a Handful of Stocks
While the economic dependency is concentrated, the market dependency is even more so. The so-called Magnificent Seven — Nvidia, Microsoft, Meta, Alphabet, Amazon, Apple, and Tesla — account for more than a third of the S&P 500's total market capitalization. Since the launch of ChatGPT in November 2022, the group has appreciated more than 260%, while the average S&P 500 stock gained just 35%.
Nvidia alone, up more than 1,000% since that launch, commands approximately 7% of the index. The concentration is at levels reminiscent of the dot-com peak, but with a crucial difference: today's leaders are enormously profitable, generating hundreds of billions in free cash flow. That profitability is what makes the concentration defensible — and what makes a reversal so dangerous, because the earnings expectations embedded in these prices assume the capex supercycle continues uninterrupted.
Here is the second-order chain that most of the market is not pricing. A capex slowdown does not merely hurt equipment vendors. It forces downward revisions to hyperscaler earnings, which compresses the valuations of the largest index constituents, which triggers a broad index drawdown, which then hits household wealth. Equities now represent nearly 30% of household wealth, up 10 percentage points from a decade ago. The 2022 and 2025 bear markets produced only modest damage to the real economy, but those were valuation-led corrections with the capex cycle still accelerating. A correction that coincides with an actual deceleration in capital spending would transmit through the GDP identity directly, not just through sentiment.
The 10-year Treasury yield near 5% compounds the problem. Every deferred dollar of AI revenue is worth less when discounted at 5% than at 3%, and the highest-multiple stocks in the index are the most sensitive to that discount rate. The AI trade and the rates trade are not separate risks. They are the same risk viewed from two angles.
Notably, the deceleration is already expected. Goldman Sachs analysts project hyperscaler capex growth slowing from 75% in the third quarter of 2025 to 49% in the fourth quarter and to 25% by the end of 2026. The question is not whether growth slows — the consensus says it will. The question is whether a 25% growth rate is still enough to sustain Nvidia's data center trajectory and the index weights built on top of it.
The Counter-Thesis: This Is Not 2008, and Bubbles Can Be Generative
The strongest argument against the alarm runs through a September 2026 analysis by Boston Consulting Group economists Philipp Carlsson-Szlezak and Paul Swartz, who contend that the question itself is misframed.
There's a new parlor game circulating global C-suites: Is there an AI capex bubble — and if so, when will it pop?
Asking the right question matters, they argue, and timing a bubble is impossible; the more productive frame is to size the exposure accurately and trace the channels through which a bust would actually transmit. Their case rests on three channels. First, an abrupt stop to AI investment would produce a direct drag of roughly 1% of GDP — meaningful, but not recession-defining on its own. Second, the wealth effect from an equity sell-off carries weight, yet two post-Covid bear markets demonstrated limited real-economy spillover. Third, and most important, the credit channel is contained: to be structurally scarring, they argue, a bursting bubble must leave crippling losses in the banking system, and AI capex financing is not likely to weigh on bank balance sheets the way housing did in the middle of the 2000s.
They go further, invoking the dot-com era not as a cautionary tale but as a structural lesson. Amazon lost nearly 95% of its value and Global Crossing went bankrupt, but the fiber-optic cables those investments funded became critical infrastructure for two decades of growth. Think of a bubble, in their framing, as solving a collective-action problem: if systemic risks are contained, it may be more sensible to cheer a bubble on than to fear it.
Nvidia's chief executive has made the bullish version of this argument in his own words. Speaking in February, Jensen Huang called the industry's surging capital expenditures "justified, appropriate and sustainable," adding:
To the extent that people continue to pay for the AI and the AI companies are able to generate a profit from that, they're going to keep on doubling, doubling, doubling, doubling.
That argument is correct on systemic risk and wrong on market risk. The absence of a banking crisis does not prevent a severe equity correction; 2000 proved that. The Nasdaq lost 78% of its value without a single systemic bank failure. The BCG framing also understates the concentration of today's index: in 2000, the damage was confined largely to the technology sector; today, passive ownership means a drawdown in seven stocks is a drawdown in every index fund, every target-date portfolio, and every household that owns the market.
The falsifying signal for the bear case is specific and observable: if hyperscaler capital expenditure growth stays above approximately 30% year over year through 2027 and AI cloud revenue growth keeps pace — meaning revenue per dollar of capex does not deteriorate — then the capex is being absorbed by demand, the bubble thesis fails, and the buildout is being vindicated rather than overbuilt. Conversely, if capex growth slows while revenue growth slows faster, the gap widens and the correction case strengthens.
What to Watch
The outlook splits by time horizon, and the three horizons point in different directions.
In the short term, sentiment and liquidity dominate, and both are vulnerable. The next round of hyperscaler earnings guidance will set the tone: any downward revision to 2026 or 2027 capex plans would be read as the first crack. Nvidia's data center growth rate — 92% in the most recent quarter — is the single most watched number; a deceleration below roughly 50% would force a broad re-rating of the equipment supply chain.
In the medium term, fundamentals decide. The base case is that capex growth decelerates but remains positive through 2027, the market absorbs a 10% to 20% correction in AI leaders, and the economy slows without tipping into recession. The downside case is that capex growth turns negative while AI revenue disappoints, producing a 30% or larger drawdown in the Magnificent Seven and tipping an economy that is running on a single engine into contraction. The upside case is that AI revenue catches up with the infrastructure, utilization rates prove out, and the buildout is vindicated the way the fiber buildout eventually was.
In the long term, the structural call stands apart from the cyclical one. Artificial intelligence is real, and the infrastructure being built will be used. The question is not whether the technology delivers — it is whether the companies that paid the most for the shovels survive the consolidation that follows every boom.
Investors should watch four signals: hyperscaler capex guidance at each earnings cycle; the ratio of AI cloud revenue growth to capex growth; Nvidia's data center revenue trajectory; and the 10-year Treasury yield, which sets the price of every future AI dollar.
The AI boom is not a house of cards because the technology is fake. It is a house of cards because the economy has been asked to balance on a single, cyclical pillar while the market pretends that pillar is the whole building.
Explore more exclusive insights at nextfin.ai.
