When it comes to a capital spending bubble, the trick isn't dancing when the music is playing, but knowing when it's time to sit down. As summer turns to fall, investors should keep on dancing to the AI beat.
With just four more months left in 2026, we understand the desire to step away. The S&P 500 has gained around 12% this year, but only slightly higher than it was at the beginning of June. The longer it remains stuck, the easier it is to focus on what could go wrong. And no issue looms larger than the massive amount of money companies are spending on artificial intelligence right now.
The scale of AI capital expenditures has been astounding. Since the start of 2024, roughly $500 billion has been spent on chips, $350 billion on power infrastructure, $200 billion on construction, and $100 billion on networking. That dwarfs the roughly $575 billion of capex from the entire S&P 500 in 2021, the year before OpenAI's ChatGPT launched.
As long as companies were paying for AI out of free cash flow, no one seemed to mind. But now, the cash-flow machines of old-Amazon.com, Alphabet, Meta Platforms, and Microsoft, among others-are going into hock to pay for their AI data centers, and the worries are growing. And for good reason. The amount they aim to spend is staggering-some $2 trillion over the next two years-and concerns are starting to show up in debt markets, where the cost to insure against a credit default has skyrocketed.
Wall Street is also fretting about the stocks. Shares of AI companies have dropped 20% from a June 52-week high. It's enough to make an investor ask if the end of the AI trade, if not the AI buildout, is upon us.
History suggests otherwise.
Capital spending bubbles almost always pop, but rarely when investors suspect they will. That has been the case for much of U.S. history, starting with the railroad buildout, through the dot-com boom, and beyond. While the AI capex explosion is almost certainly going to end badly, the chances of it ending badly now are low. Over America's 250-year history, the economy has absorbed spending on transformation technology equivalent to about 25% of economic output before things get really hairy, according to a Barron's analysis. Artificial intelligence isn't there yet, and won't be for years, suggesting that the AI trade-and the stock market rally-has more room to run.
"I think we're going to just keep going," says Melius Research's head of technology research, Ben Reitzes. "This will be bigger than anybody thinks, for longer than anybody thinks."
America has a long history of capex booms and busts. From the construction of the first railroads in the 1870s, through the electrification boom of the 1920s, to the dot-com and housing bubbles of the 1990s and 2000s, they have followed a similar pattern, according to Raymond James strategist Tavis McCourt. A new technology arrives, one that is met with enthusiasm by investors, pushing up valuations, and ultimately leading to equity and debt raises and market concentration issues before turning to bust. The booms don't just end on their own, however. McCourt notes that they require an outside shock to spur the decline, one that can be policy-driven, like a Fed interest-rate hike, or an event like the San Francisco earthquake of 1906.
Historically, these cycles have followed enough of a pattern to consider where artificial intelligence falls in the cycle. One could be dubbed the rule of 25, for the amount of total spending the U.S. has been able to digest during transformational booms as a percentage of the overall economy. At the start of the railroad boom in the 1860s, for instance, U.S. gross domestic product was about $10 billion a year, according to the National Bureau of Economic Research, while rail spending eventually totaled $2.5 billion before the 1873 panic arrived. Similarly, about $1.5 trillion was spent building internet infrastructure in the late 1990s, while the U.S. economy was only $6 trillion at the time. The same holds true for the industrial and electrification buildout of the 1920s.
Using that standard, it's possible to generate a top-down estimate of how much can be spent on AI before the economy is ready to collapse under its own weight. With U.S. GDP at roughly $30 trillion, the danger zone sits at about $7.5 trillion, or an additional $5 trillion to $6 trillion in domestic AI spending. Hyperscalers are projected to spend $3.7 trillion globally through 2029, and they aren't the only ones, with Oracle, SpaceX, Anthropic, and OpenAI, among others, building, too. At this rate, AI spending won't trip the rule of 25 until the early 2030s, six or seven years into the boom.
Even when it does, not every capex bubble ends up popping. For a capital spending boom to succeed, the outlays need to slow while profitability ramps up, Raymond James' McCourt notes, allowing the businesses to become self-financing. McCourt acknowledges the potential for the AI buildout to become sustainable, but notes that it's far more difficult without government backing or a relatively slow pace. "Historically, this smooth ramp to profitability without any hiccups in capital spending has never happened before at this scale, except for major government-funded capital spending booms," McCourt explains.
And that's where things get tricky. The hyperscalers-Meta Platforms, Microsoft, Amazon.com, and Alphabet-spent roughly $125 billion on new plants and equipment in 2021. In 2026, that number will be closer to $700 billion. In 2027, it will be $1 trillion, and about half of total capital spending for the S&P 500. That money is the flywheel driving the stock market, becoming revenue for just about everyone-from construction companies that build the data centers; utilities that provide the electricity to power them; electronics manufacturers who build the cables, connectors, and coolers to keep them running. And then there are the chip makers, whose products drive the computing power that makes AI work. Case in point: Nvidia, which Wednesday reported better than expected second-quarter earnings-and forecast a whopping 70% increase in revenue for fiscal 2028.
There are signs that the spending is becoming more profitable. Alphabet's cloud business grew 82% year over year in the second quarter. Anthropic's current annual revenue run-rate-Wall Street's favorite metric for AI start-ups-is currently estimated to be north of $65 billion, up from $9 billion at the end of 2025. That means Anthropic has added business roughly the size of 50-year-old Oracle in eight months. SpaceX, which subsumed Elon Musk's xAI in February, generated $3.2 billion in 2025 AI revenue, and will likely surpass $60 billion in 2027. Much of that revenue is coming from third parties, including Anthropic and Alphabet, which rent SpaceX's terrestrial data-center capacity for billions a month. It might also be the right kind of profitability: While the S&P 500 needs $2 to $3 in assets to generate $1 in sales, a gigawatt of computing power costing $40 billion to build can produce $40 billion in sales. That is an epically good business.
But funding $6 trillion for AI capex in a handful of years is asking a lot of U.S. capital markets. Yes, they handled Alphabet's $85 billion equity raise in June, and SpaceX set a record for a new public-company offering by raising $86 billion in its IPO and another $25 billion from debt markets just 10 days later. Others have had to get more creative. Nvidia will provide up to $100 billion in financing to OpenAI so it can use Nvidia graphics processing units, or GPUs, in up to 10 gigawatts of AI computing power. OpenAI, Oracle, and SoftBank launched the Stargate Project, an AI infrastructure funding project with commitments of some $500 billion.
Even-more opaque financing methods are popping up, as well. Meta Platforms created a special purpose vehicle called Beignet Investor to build five gigawatts of AI computing capacity in Richland Parish, La., with about $27 billion in debt and $3 billion in equity. Meta will own 20%, with Blue Owl Capital and other private-equity players owning 80%, but Meta will be the one renting the data center and paying the operating expenses, and it will be on the hook for the debt if it decides not to renew the lease. No wonder Beignet's debt is rated A+.
All told, funding "extraordinary growth" will require everything from public equity markets to private credit, says J.P. Morgan Securities analyst Tarek Hamid. He projects more than $5 trillion in AI data-center spending from 2026 to 2030, with $1.5 trillion funded from cash flow, $500 billion from new equity, $2.5 trillion from credit markets, and $1.4 trillion from alternative sources, like Meta's Louisiana deal, and governments.
With so much depending on financing-and untraditional financing at that-the bigger risk might be Fed rate hikes. Yields on long-term Treasuries are already near their highest levels since 2007, and the worry is that rate hikes could send yields higher across the board. That would make financing the AI buildout more expensive. "It stands to reason that sensitivity of AI buildout names to 10-year yields will increase meaningfully from here," writes 22V strategist Dennis DeBusschere. "We should expect AI buildout names to suffer, all things equal, if 10-year yields keep grinding higher."
Debt investors have taken notice. While bond spreads-the difference between the interest a company pays on its debt versus an equivalent Treasury-remain tight, credit default swaps, which act as a form of insurance, have surged higher. It now costs about twice as much to insure against an Alphabet default than it did at the beginning of the year, and it's a similar story for Meta and Amazon. Though the rates are well below concerning thresholds for investment-grade companies, the moves signal that markets recognize that Big Tech companies won't be the debt-free juggernauts they had been.
The balance sheets are still sound. The four hyperscalers still have essentially no net debt and are expected to generate roughly $1 trillion in earnings before interest, taxes, depreciation, and amortization, or Ebitda, in 2027.
Debt levels aren't a problem for 2028. For now, the focus should be on finding the companies that have the best chance of benefiting from the continued buildout. Choosing an outright winner is next to impossible. But there will be outright AI winners. They could be SpaceX, Tesla (or a SpaceX/Tesla combination), Alphabet, Anthropic, and OpenAI. The safest of the AI plays is clearly Alphabet. It has an existing business that is pretty good, with operating income in 2026 projected to be north of $170 billion. Shares trade for a not unreasonable 20 times earnings expected over the coming 12 months, about in line with the S&P 500.
But infrastructure plays might be even safer, especially as demand shifts. If the first iteration of the AI trade was simply about building the best chips and sticking them in CPUs, the second will focus on building "AI factories," according to Evercore analyst Amit Daryanani. Nvidia's Rubin is designed to do just that by working across compute, networking, and storage, while companies like Meta Platforms and Microsoft are creating tools to minimize the cost of using AI. "We think this broadens the AI infrastructure opportunity...toward memory, networking, power, cooling, and software," he writes.
Daryanani cites Amphenol and TE Connectivity as two stocks that could benefit from the trend toward more complex connectivity needs. Other popular stocks include Eaton, which provides power distribution and cooling solutions for data centers, among other products, and Vertiv Holdings, a leader in data-center hardware and services. Both are more expensive than they used to be-Eaton trades for about 28 times earnings, up from 25 times two years ago, while Vertiv changes hands at 32 times earnings, up from 27 times-though both are still well liked by analysts: 78% of analysts rate Eaton a Buy, while Vertiv is 85% Buy-rated.
The data-center backlash is another risk-but also an opportunity. While demand for AI is rising, the pushback is likely to slow the buildout of new data centers. That should boost shares of companies that already have projects approved, including SpaceX, CoreWeave, Nebius, and Fermi, according to 22V's Dauvin Peterson. CoreWeave, in particular, looks interesting. While the market remains focused on its debt, higher AI computing prices would boost earnings-and the stock.
The good news is that when the bubble pops-and it will pop-what was built will remain. As with the railroad, the internet, and electrification booms, it will take time for financial markets to recover. But the foundation will already be in place for the next expansion-and for the dance to begin yet again.
Write to Al Root at allen.root@dowjones.com
This content was created by Barron's, which is operated by Dow Jones & Co. Barron's is published independently from Dow Jones Newswires and The Wall Street Journal.
(END) Dow Jones Newswires
By Al Root
When it comes to a capital spending bubble, the trick isn't dancing when the music is playing, but knowing when it's time to sit down. As summer turns to fall, investors should keep on dancing to the AI beat.
With just four more months left in 2026, we understand the desire to step away. The S&P 500 has gained around 12% this year, but only slightly higher than it was at the beginning of June. The longer it remains stuck, the easier it is to focus on what could go wrong. And no issue looms larger than the massive amount of money companies are spending on artificial intelligence right now.
The scale of AI capital expenditures has been astounding. Since the start of 2024, roughly $500 billion has been spent on chips, $350 billion on power infrastructure, $200 billion on construction, and $100 billion on networking. That dwarfs the roughly $575 billion of capex from the entire S&P 500 in 2021, the year before OpenAI's ChatGPT launched.
As long as companies were paying for AI out of free cash flow, no one seemed to mind. But now, the cash-flow machines of old-Amazon.com, Alphabet, Meta Platforms, and Microsoft, among others-are going into hock to pay for their AI data centers, and the worries are growing. And for good reason. The amount they aim to spend is staggering-some $2 trillion over the next two years-and concerns are starting to show up in debt markets, where the cost to insure against a credit default has skyrocketed.
Wall Street is also fretting about the stocks. Shares of AI companies have dropped 20% from a June 52-week high. It's enough to make an investor ask if the end of the AI trade, if not the AI buildout, is upon us.
History suggests otherwise.
Capital spending bubbles almost always pop, but rarely when investors suspect they will. That has been the case for much of U.S. history, starting with the railroad buildout, through the dot-com boom, and beyond. While the AI capex explosion is almost certainly going to end badly, the chances of it ending badly now are low. Over America's 250-year history, the economy has absorbed spending on transformation technology equivalent to about 25% of economic output before things get really hairy, according to a Barron's analysis. Artificial intelligence isn't there yet, and won't be for years, suggesting that the AI trade-and the stock market rally-has more room to run.
"I think we're going to just keep going," says Melius Research's head of technology research, Ben Reitzes. "This will be bigger than anybody thinks, for longer than anybody thinks."
America has a long history of capex booms and busts. From the construction of the first railroads in the 1870s, through the electrification boom of the 1920s, to the dot-com and housing bubbles of the 1990s and 2000s, they have followed a similar pattern, according to Raymond James strategist Tavis McCourt. A new technology arrives, one that is met with enthusiasm by investors, pushing up valuations, and ultimately leading to equity and debt raises and market concentration issues before turning to bust. The booms don't just end on their own, however. McCourt notes that they require an outside shock to spur the decline, one that can be policy-driven, like a Fed interest-rate hike, or an event like the San Francisco earthquake of 1906.
Historically, these cycles have followed enough of a pattern to consider where artificial intelligence falls in the cycle. One could be dubbed the rule of 25, for the amount of total spending the U.S. has been able to digest during transformational booms as a percentage of the overall economy. At the start of the railroad boom in the 1860s, for instance, U.S. gross domestic product was about $10 billion a year, according to the National Bureau of Economic Research, while rail spending eventually totaled $2.5 billion before the 1873 panic arrived. Similarly, about $1.5 trillion was spent building internet infrastructure in the late 1990s, while the U.S. economy was only $6 trillion at the time. The same holds true for the industrial and electrification buildout of the 1920s.
Using that standard, it's possible to generate a top-down estimate of how much can be spent on AI before the economy is ready to collapse under its own weight. With U.S. GDP at roughly $30 trillion, the danger zone sits at about $7.5 trillion, or an additional $5 trillion to $6 trillion in domestic AI spending. Hyperscalers are projected to spend $3.7 trillion globally through 2029, and they aren't the only ones, with Oracle, SpaceX, Anthropic, and OpenAI, among others, building, too. At this rate, AI spending won't trip the rule of 25 until the early 2030s, six or seven years into the boom.
Even when it does, not every capex bubble ends up popping. For a capital spending boom to succeed, the outlays need to slow while profitability ramps up, Raymond James' McCourt notes, allowing the businesses to become self-financing. McCourt acknowledges the potential for the AI buildout to become sustainable, but notes that it's far more difficult without government backing or a relatively slow pace. "Historically, this smooth ramp to profitability without any hiccups in capital spending has never happened before at this scale, except for major government-funded capital spending booms," McCourt explains.
And that's where things get tricky. The hyperscalers-Meta Platforms, Microsoft, Amazon.com, and Alphabet-spent roughly $125 billion on new plants and equipment in 2021. In 2026, that number will be closer to $700 billion. In 2027, it will be $1 trillion, and about half of total capital spending for the S&P 500. That money is the flywheel driving the stock market, becoming revenue for just about everyone-from construction companies that build the data centers; utilities that provide the electricity to power them; electronics manufacturers who build the cables, connectors, and coolers to keep them running. And then there are the chip makers, whose products drive the computing power that makes AI work. Case in point: Nvidia, which Wednesday reported better than expected second-quarter earnings-and forecast a whopping 70% increase in revenue for fiscal 2028.
There are signs that the spending is becoming more profitable. Alphabet's cloud business grew 82% year over year in the second quarter. Anthropic's current annual revenue run-rate-Wall Street's favorite metric for AI start-ups-is currently estimated to be north of $65 billion, up from $9 billion at the end of 2025. That means Anthropic has added business roughly the size of 50-year-old Oracle in eight months. SpaceX, which subsumed Elon Musk's xAI in February, generated $3.2 billion in 2025 AI revenue, and will likely surpass $60 billion in 2027. Much of that revenue is coming from third parties, including Anthropic and Alphabet, which rent SpaceX's terrestrial data-center capacity for billions a month. It might also be the right kind of profitability: While the S&P 500 needs $2 to $3 in assets to generate $1 in sales, a gigawatt of computing power costing $40 billion to build can produce $40 billion in sales. That is an epically good business.
But funding $6 trillion for AI capex in a handful of years is asking a lot of U.S. capital markets. Yes, they handled Alphabet's $85 billion equity raise in June, and SpaceX set a record for a new public-company offering by raising $86 billion in its IPO and another $25 billion from debt markets just 10 days later. Others have had to get more creative. Nvidia will provide up to $100 billion in financing to OpenAI so it can use Nvidia graphics processing units, or GPUs, in up to 10 gigawatts of AI computing power. OpenAI, Oracle, and SoftBank launched the Stargate Project, an AI infrastructure funding project with commitments of some $500 billion.
Even-more opaque financing methods are popping up, as well. Meta Platforms created a special purpose vehicle called Beignet Investor to build five gigawatts of AI computing capacity in Richland Parish, La., with about $27 billion in debt and $3 billion in equity. Meta will own 20%, with Blue Owl Capital and other private-equity players owning 80%, but Meta will be the one renting the data center and paying the operating expenses, and it will be on the hook for the debt if it decides not to renew the lease. No wonder Beignet's debt is rated A+.
All told, funding "extraordinary growth" will require everything from public equity markets to private credit, says J.P. Morgan Securities analyst Tarek Hamid. He projects more than $5 trillion in AI data-center spending from 2026 to 2030, with $1.5 trillion funded from cash flow, $500 billion from new equity, $2.5 trillion from credit markets, and $1.4 trillion from alternative sources, like Meta's Louisiana deal, and governments.
With so much depending on financing-and untraditional financing at that-the bigger risk might be Fed rate hikes. Yields on long-term Treasuries are already near their highest levels since 2007, and the worry is that rate hikes could send yields higher across the board. That would make financing the AI buildout more expensive. "It stands to reason that sensitivity of AI buildout names to 10-year yields will increase meaningfully from here," writes 22V strategist Dennis DeBusschere. "We should expect AI buildout names to suffer, all things equal, if 10-year yields keep grinding higher."
Debt investors have taken notice. While bond spreads-the difference between the interest a company pays on its debt versus an equivalent Treasury-remain tight, credit default swaps, which act as a form of insurance, have surged higher. It now costs about twice as much to insure against an Alphabet default than it did at the beginning of the year, and it's a similar story for Meta and Amazon. Though the rates are well below concerning thresholds for investment-grade companies, the moves signal that markets recognize that Big Tech companies won't be the debt-free juggernauts they had been.
Comments