The headlines are screaming that the AI boom is finally hitting a wall. Tech stocks have sold off. Microsoft trimmed its capital expenditure forecast. Jefferies is warning that the AI investment cycle may have peaked. After three years of spending every dollar as if it were a down payment on the future, the market has started asking an uncomfortable question: can the revenue ever catch up with the spending?
But the honest answer is more complicated than the “bubble bursting” narrative suggests. AI spending is not collapsing. It is decelerating—and that distinction matters enormously for how we understand what comes next. The four largest hyperscalers are still on track to spend roughly $700 billion** on capital expenditure in 2026, and consensus estimates put their combined spending at over **$1.04 trillion in 2027. That is not a slowdown in absolute terms. It is a slowdown in the rate of acceleration—a mid-cycle correction, not a bust.
The reality check is real. But it is a check on expectations, not on the technology itself.
What “Slowdown” Actually Means
The confusion stems from conflating two different things: the level of spending and the pace of growth.
On the level, the numbers are still staggering. The four major hyperscalers—Microsoft, Amazon, Alphabet, and Meta—increased their combined capital expenditure by 108% year-on-year to $202 billion in the third quarter of 2026 alone**. For the full year, they are expected to spend close to **$800 billion, and BofA Global Research projects nearly $1.08 trillion in 2027**. Morgan Stanley expects hyperscaler data-center capex to peak at roughly **$1.5 trillion in 2027 before growth slows to about 12% in 2028.
On the pace, the signals are more nuanced. Microsoft reduced its capex forecast for the year to $170 billion from $175 billion. Alphabet held its guidance at $195 billion to $205 billion rather than raising it, as it had done in previous quarters. Both companies reported capital expenditures that fell short of analysts’ estimates. Jefferies described this as evidence that “the AI capex cycle may have peaked, or at least commenced a mid-cycle correction”.
In other words: the spending is still enormous, but the upward revisions—the quarterly ritual of “we’re going to spend even more than we said last time”—have stopped. That is the reality check.
The Four Forces Driving the Deceleration
The slowdown is not the result of a single factor. It is the convergence of four pressures that have been building for months.
1. The return-on-investment question has become impossible to ignore. The most cited data point is a MIT NANDA study from July 2025 which found that 95% of organisations saw no measurable return on their generative AI investments. The study has its critics—measuring ROI for AI is notoriously difficult, and the methodology has been questioned—but it has become a shorthand for a broader anxiety: the spending is outrunning the evidence that it is working. As Business Times put it, “data centre construction is following an exponential curve, with each year’s plans dwarfing the last. But AI adoption is not exponential”.
2. Financing costs are rising. The AI buildout has increasingly been funded by debt, and the credit markets are starting to push back. Bond yield spreads for major hyperscalers have widened, and credit default swap spreads for Oracle and CoreWeave have jumped sharply. CoreWeave, the largest listed specialist AI cloud provider, has seen its borrowing costs rise, and its stock has lost more than 38% in a month. Jefferies warned that “financial markets were increasingly signalling reluctance to fund AI-related investments”.
3. Political and community resistance is mounting. The 2026 US midterm elections delivered a clear message: voters do not want massive, energy-hungry data centers built in their neighborhoods. AI was on the ballot across multiple states, with measures seeking to delay or restrict data center developments, and voters “stuck up two fingers at Big Tech almost every time”. JPMorgan estimated that 37.5 gigawatts of planned 2028 capacity could face pauses, and another 15 gigawatts could face onerous conditions—paying grid costs, ending tax incentives, imposing water limits, and banning non-disclosure agreements during permitting.
4. Supply-chain bottlenecks are real. Microsoft’s CFO said that third-party providers of AI chips had delays that will push out some available capacity until the first half of 2027. Alphabet’s finance chief said outside computing power “came online only at the very end of Q3”. The constraint is no longer just chips—it is power, land, permitting, and the physical capacity to build at the pace the industry planned.
The Market Reaction: Punishing the Picks and Shovels
The stock market’s response has been revealing. It is not the hyperscalers that have borne the brunt of the selling. It is the companies that supply them.
The Philadelphia Semiconductor Index was down 14% in the month to October 23. Engineering and construction firms tied to the AI buildout—Quanta Services, Emcor, Vertiv, Comfort Systems—were “hammered especially hard”, as was GE Vernova, the turbine maker. The neoclouds—smaller data center operators that borrowed heavily to build capacity—stumbled on the prospect of construction delays while they were still paying interest on their debt.
The logic is straightforward: if the rate of AI infrastructure spending is slowing, the suppliers who benefited from the acceleration are the most exposed. The hyperscalers, by contrast, have businesses that generate cash from advertising, cloud computing, and software. They do not need to become AI companies overnight.
The Counter-Argument: A Pause That Refreshes
Not everyone sees the slowdown as bad news. In fact, some of the most respected voices on Wall Street see it as bullish.
Ed Yardeni, the veteran strategist, has argued that “a slowing down in the AI build-out is actually a bullish story”. His reasoning is simple: the hyperscalers have been spending so aggressively that their free cash flow has turned negative. Amazon generated $168.3 billion in operating cash flow** over the past year but reported **negative $42.4 billion in free cash flow as AI investment accelerated. If they slow the pace of construction, “their free cash flow will be less negative, maybe even turn positive”.
The Financial Times Lex column made a similar point: “Delays may frustrate, but they also make it less likely that trillions of dollars are invested in equipment at risk of becoming obsolete or depreciating faster than its financiers expected”. A slower buildout reduces the risk of stranded assets and overcapacity.
There is also a macroeconomic argument. Evercore’s Krishna Guha wrote that slowing the AI capex mania could ease “overheating risk in the next year or two and plausibly extend the AI investment cycle”—helping to keep the economy in balance without requiring further rate hikes. Substack writer Derek Thompson put it more bluntly: “Anti-AI politics is bad for AI stocks in the short run but very good for preventing a bubble in the medium to long run”.
Is This the “Big One”?
The honest answer is: probably not, but the risk is real.
The hallmarks of an unsustainable bubble are present. The Financial Times notes “high levels of increasingly creatively structured debt, questionable levels of true revenue potential, and head-scratching examples of circular financing”. The Bank for International Settlements has pointed out that nearly 29% of AI firms’ investment deals are in other AI firms, and more than half of incoming investment comes from other AI firms—a structure that could turn into a “circular firing squad” in a crisis.
The IMF has warned that a steep drop in AI investment could cut global GDP by 1.1% by 2028, including a 1.8% drop in US output. AI-related investment contributed 1.39 percentage points—or 67%—of US real GDP growth in the four quarters to the third quarter of 2026. If that spending engine stalls, the macroeconomic consequences would be significant.
But a bubble bursting requires a trigger, and the most likely trigger—a sudden, disorderly collapse in spending—has not materialised. What is happening instead is a controlled deceleration: a recognition that the pace of the last two years was unsustainable, and a correction toward a more sustainable trajectory. Polymarket traders put the odds of an AI bubble bursting by December 2026 at just 8%, rising to 23% by June 2027.
AI spending is slowing down.
But the slowdown is not a collapse. It is a reality check—a moment when the market, the financiers, and the body politic have collectively said: prove that this works before we keep spending at this pace.
The technology is real. The demand for AI compute is real. The revenue is growing—Google Cloud revenue jumped 103% to $30.7 billion**, and AWS grew **39% to $46 billion. But the gap between spending and revenue is still wide, and the willingness to bridge that gap with debt and optimism has limits.
The next phase of the AI boom will be defined not by how much is spent, but by how efficiently it is spent. The companies that can demonstrate returns, manage their cash flow, and navigate the political and physical constraints on data center construction will be the ones that emerge strongest. The ones that cannot will be the ones that the market punishes.
The AI boom is not over. It is growing up. And growing up is always a reality check.
