Everything About AI
NextFin News - The four largest builders of AI infrastructure are on pace to spend roughly $730 billion on artificial-intelligence projects in 2026, yet the market is no longer rewarding the check-writers. A measure of how closely the biggest capital-expenditure spenders have moved with semiconductor stocks has collapsed to near zero, the weakest reading in about four and a half years, as investors demand proof that the spending will produce returns rather than just more spending.
The tension defines the current phase of the AI boom: the buildout is accelerating, not slowing. Alphabet and Meta Platforms both raised their 2026 capital-expenditure outlooks during recent earnings discussions, joining Microsoft and Amazon in a coordinated spending surge that would have been unthinkable two years ago. But the stocks of the companies writing the checks have stopped moving in lockstep with the chipmakers they enrich, and the gap between the two is now the clearest signal that investor patience has a price.
The Spending Keeps Rising While the Market Looks Away
The scale of the commitment is difficult to overstate. Alphabet is guiding to between $195 billion and $205 billion in 2026 capital expenditures, Meta to between $125 billion and $145 billion, Microsoft to about $190 billion, and Amazon to about $200 billion. Taken together, the four hyperscalers are investing at a run rate roughly 62 percent above the record $388 billion they spent in 2025, and Wall Street already expects the number to rise again in 2027. The money is flowing into data centers, networking gear, AI accelerators, custom silicon, and the power infrastructure required to run them.
Broader estimates put the picture even larger. Bank of America Global Research expects the AI hyperscalers — Microsoft, Alphabet, Amazon, Meta Platforms, and Oracle — to deploy around $795 billion in capital expenditures in 2026 and nearly $1.08 trillion in 2027. Goldman Sachs Research, using an augmented methodology that includes non-hyperscaler spending, forecasts global AI investment exceeding $1 trillion this year and argues that the commonly cited $794 billion hyperscaler figure likely understates the global total by about $200 billion. Charles Schwab's mid-year outlook puts the largest hyperscalers on track to spend nearly $800 billion this year, with capital expenditure approaching 75 percent of their cash flows — a ratio the firm describes as reminiscent of technology-company spending in the late 1990s.
Yet the market's response to each new upward revision has turned negative. When Alphabet lifted its 2026 spending projection by about $15 billion in its latest earnings report, its shares fell as much as 7.4 percent the following day. The reaction was not isolated to one company; shares of Google, Amazon, and Microsoft all sold off following their earnings calls, even as executives defended the spending as necessary to win the AI race.
The divergence between spenders and suppliers is the sharpest evidence that the trade has changed. Research compiled by Kevin Gordon of Charles Schwab shows the 30-day correlation between America's largest capital-expenditure spenders and the Philadelphia Semiconductor Index collapsed from +0.78 in April 2026 to near zero more recently, against an average of +0.60 since early 2022. That is close to the lowest reading in at least four and a half years. Since early June, semiconductor stocks have generally advanced while many hyperscaler shares have lagged — a decoupling that would have been hard to imagine when the two groups rose and fell together through 2023 and 2024.
"Markets are punishing the picks-and-shovels layer harder than the hyperscalers because it's the layer most exposed to a slowdown in the rate of capability improvement," said Erik Kratz, chief investment officer and co-head of wealth at Arena Private Wealth in Chicago.
That reading cuts against the surface narrative. The pressure is not evenly distributed: the suppliers of GPUs and accelerators face the greatest risk if the pace of capability gains slows, because their growth depends on the hyperscalers continuing to refresh hardware at the current frenetic rate. The hyperscalers, by contrast, own the customer relationships and the revenue streams — but they also own the depreciation.
The Returns Gap: $25 Billion Visible Against Hundreds of Billions Spent
The central question investors are asking is simple: who pays for all this, and when? Against the mountain of spending, the visible return remains thin. Bank of America analysts estimate that hyperscaler capital expenditure is already consuming about 94 percent of operating cash flow after dividends and buybacks. Meanwhile, AI services generate roughly $25 billion in direct revenue — approximately 4 percent of what is being poured into infrastructure.
That 4 percent figure is the arithmetic at the heart of the skepticism. Even if AI revenue doubles, it would still cover only a fraction of the annual capital bill. The gap forces a reckoning with the economics of the buildout: the companies funding it are betting that today's spending buys a durable competitive position and a stream of future cash flow that the current revenue run rate does not yet reflect.
Goldman Sachs strategists put the break-even point in stark terms: to earn an adequate return on their outlays, the hyperscalers would need annual AI revenue of about $300 billion in the coming years. Current AI revenue is a fraction of that target, which is why every earnings call now turns on the same question — how quickly can monetization catch up to capacity?
The strain is already visible in the cash-flow data. Microsoft generated $55.4 billion in operating cash flow in the second quarter of fiscal 2026 (April through June), invested $35.8 billion in AI data centers and semiconductors, and produced $19.6 billion in free cash flow. Alphabet, over the same period, recorded $39.1 billion in operating cash flow but failed to generate positive free cash flow after pouring $44.9 billion into capital expenditures — its first negative free-cash-flow quarter in roughly two decades. One company is funding its buildout from internal cash with room to spare; the other is already spending more than it takes in.
Barclays estimates that capital expenditure across the five largest hyperscalers is consuming about 90 percent of operating cash flow, with the alert threshold at 100 percent — a line the group is approaching. Bank of America's 94 percent figure suggests parts of the group are already there.
The correlation collapse between hyperscalers and chip stocks is more than a technical curiosity; it is a change in who the market believes bears the risk. For the past several years, the trade was simple: buy the companies building AI infrastructure and the chipmakers supplying them, and both sides moved higher together. That relationship has broken down because the two groups now face different risks.
The chipmakers' risk is volume. Their revenue depends on the hyperscalers continuing to buy GPUs and accelerators at the current pace. If the hyperscalers slow their refresh cycle — because utilization is lower than expected, because model efficiency improves, or because capital becomes more expensive — the semiconductor order book contracts quickly. That is why the picks-and-shovels layer is being punished hardest on fears of a slowdown in capability improvement.
The hyperscalers' risk is duration. They have locked in years of fixed obligations — leases, power contracts, purchase commitments — while the revenue from those assets is uncertain in both timing and magnitude. Their exposure is not to a single bad quarter but to a multi-year stretch in which depreciation and financing costs outrun AI revenue growth. That is a slower, more grinding form of risk, and it is why their stocks have not crashed even as the sentiment has soured.
This divergence also reshapes which companies benefit from the next leg of the buildout. Companies that can convert spending into durable, disclosed cash flow will be rewarded; companies that can only disclose larger capital budgets will not. The market is beginning to discriminate between the two, and that discrimination is likely to intensify through the coming earnings seasons.
The Hidden Balance Sheet: $1.65 Trillion in Off-Book Obligations
The on-balance-sheet capital expenditures tell only part of the story. A study of financial disclosures found that Alphabet, Microsoft, Amazon, Meta, and Oracle have accumulated approximately $1.65 trillion in off-balance-sheet obligations tied to AI infrastructure, data-center leases, GPU supply contracts, and special-purpose vehicles — a figure that has grown roughly eightfold since 2022 and now exceeds the $1.35 trillion these five companies report as on-balance-sheet debt.
A February 2026 Moody's Ratings report estimated $662 billion in data-center lease commitments signed but not yet commenced across the five hyperscalers. Under current lease-accounting rules, those obligations do not appear on the balance sheet until the facilities open. The gap between what has been promised and what is recorded is larger than the annual capital-expenditure budget of every hyperscaler combined.
Morgan Stanley analysis put total off-balance-sheet long-term commitments at six U.S. hyperscalers above $1.3 trillion, with purchase commitments of $640 billion and lease commitments of $675 billion — most not recorded on balance sheets. Meta's commitments equal 1.7 times its annual operating cash flow, while Oracle's exceed seven times. Unexercised lease commitments surged from $240 billion to $675 billion in a single year, with power-purchase agreements and third-party guarantees adding to the exposure.
The credit market has begun to price the risk. S&P downgraded Oracle's credit rating to BBB-, one notch above junk, on July 9, 2026, explicitly citing leverage from AI commitments including roughly $260 billion in future lease obligations. Oracle's off-balance-sheet AI debt stands at approximately $273.3 billion, up more than 2,900 percent in four years. The downgrade is a warning that the financing side of the AI boom is no longer invisible to lenders.
These obligations are not a scandal; they are a rule working as written. Lease accounting treats a commitment as a liability only when the asset is available for use. But the economic reality is that the companies have promised to pay regardless of whether the AI revenue arrives. If demand disappoints, the leases do not disappear — they convert into fixed costs that must be serviced from other parts of the business.
Dot-Com Echoes, With a Different Cast of Characters
The historical parallel investors fear most is the dot-com bubble, when companies spent heavily on internet infrastructure only to see poor initial returns and a crash in stock prices. Goldman Sachs strategists led by Ryan Hammond have framed the comparison directly: if hyperscaler capital expenditure reaches roughly $1.2 trillion in 2027 — about 50 percent above this year's level and above the $1.1 trillion Wall Street consensus — the cycle would claim a larger share of GDP than any technology investment cycle since the railroad buildout of the late 1800s. The bank sees spending rising again to about $1.4 trillion in 2028.
The railroad and telecom booms of prior decades both delivered genuine long-run productivity gains. Both also destroyed significant capital before the returns showed up. That is the uncomfortable lesson for AI investors: a technology can transform the economy and still punish the investors who financed its construction.
The dot-com comparison is useful but incomplete. The companies building AI infrastructure today are not unprofitable start-ups burning venture capital; they are some of the most cash-generative businesses in history, with diversified revenue streams in search, advertising, cloud computing, and enterprise software. Alphabet's core search franchise and Meta's advertising engine continue to produce tens of billions in quarterly profit that can fund the AI buildout even while AI itself is not yet paying its way.
But the mechanism of capital destruction in the railroad and telecom booms was not weak management or fraudulent accounting. It was overbuilding: too much capacity chasing too little demand, which drove down prices and left assets earning below their cost of capital for years. The question for AI is whether the current capacity buildout is ahead of demand by a quarter, a year, or a decade.
There is also a prisoner's-dilemma dynamic that keeps spending high even as returns disappoint. If one hyperscaler pulls back while the others keep building, the one that pulled back risks losing access to the best chips, the best sites, and the best talent — and ultimately losing the AI race. That logic makes continued spending individually rational even when it is collectively excessive.
"We always like discipline, but in this case, you're really needing to compete to come out the other end in a spot where you want to be," said Molly Pieroni, president at Yacktman Asset Management, which invests in Alphabet. "So we're not telling them that they need to pull back on capex and invest less."
The market's brief enthusiasm for monetization workarounds underscores how hungry investors are for any sign of returns. When reports surfaced that Meta was planning to lease out its spare cloud capacity to outside customers, its shares soared as much as 21 percent in a matter of days. The reaction to a plan to rent out unused capacity — rather than to a breakthrough in AI revenue — shows how starved the market is for evidence that the infrastructure will pay for itself.
Cyclical Doubt or a Structural Shift in How AI Is Valued?
The critical judgment for investors is whether the current skepticism is cyclical — a temporary loss of confidence that will reverse once the next wave of AI revenue arrives — or structural, a permanent repricing of how the market values AI spending. The evidence points to a cyclical confidence shock layered on top of a structural change in the burden of proof.
The cyclical case is strong. History shows that during infrastructure buildouts, returns lag spending for years before the payoff arrives. The railroad and telecom networks took a decade or more to generate their full economic value. If AI follows the same pattern, today's negative sentiment is simply the trough of confidence that precedes the revenue curve catching up to the capital curve. The S&P 500 has more than doubled since the current bull market began in October 2022, which is what a cyclical confidence trough looks like inside a longer uptrend.
But the structural change is real and should not be dismissed as noise. The market has shifted from valuing AI spending as an unambiguous positive — where every dollar of capex signaled growth optionality — to valuing it as a claim on future cash flow that must be underwritten. That is a regime change in how capital is allocated to AI, and it will not revert on its own. Management teams can no longer assume that announcing a higher capital budget will lift their stock. They must now show utilization rates, revenue per dollar of capex, and a path to returns on invested capital. That burden of proof is permanent.
The two forces can coexist without contradiction: the sentiment cycle will likely mean-revert as AI revenue grows, but the valuation framework that produced the current skepticism is here to stay. Investors who treat this as purely cyclical risk being caught off guard by the new discipline the market demands; investors who treat it as purely structural risk missing the next leg of a genuine infrastructure supercycle.
The Counter-Thesis: Pulling Back Would Be the Bigger Mistake
The strongest argument against the skeptics is that slowing spending now would be the costlier error. The AI race is a winner-take-most competition for model quality, distribution, and enterprise relationships. A hyperscaler that cuts capital expenditure today may win short-term approval from investors focused on free cash flow, but it risks losing access to the scarce inputs — advanced chips, data-center sites, and power capacity — that determine who leads the next generation of models. In that scenario, today's stock-price gain would be purchased at the price of tomorrow's competitive position.
Pieroni's point captures this logic: discipline matters, but not when it means losing the race. The hyperscalers' diversified cash engines — search, advertising, cloud, enterprise software — give them the capacity to fund the buildout through the lean years, and the competitive cost of falling behind is likely higher than the financial cost of overbuilding.
Yet this counter-thesis has a limit, and it is quantifiable. The argument holds only if AI revenue grows fast enough to absorb the capacity being built. If direct AI revenue remains below 10 percent of infrastructure spending for an extended period — say, through 2027 — while capital expenditure continues to consume more than 100 percent of operating cash flow, then the "compete to win" logic breaks down. At that point, overbuilding becomes capital destruction, and the dot-com parallel stops being a cautionary tale and starts being a template.
The specific falsifying signal is this: if hyperscaler capital expenditure exceeds operating cash flow for four consecutive quarters while AI revenue-to-capex stays below 10 percent, the thesis that this spending is a rational investment in a durable moat is wrong, and the market's skepticism is justified. That combination would mean the companies are funding today's capacity with tomorrow's claims on cash, with no visible revenue to service them.
What to Watch: The Signals That Will Decide the Next Leg
Three categories of data will determine whether investor faith is restored or further eroded.
First, monetization metrics. Investors need disclosed figures on AI revenue per dollar of capital expenditure, data-center utilization rates, and the payback period on AI assets. Meta's plan to lease spare cloud capacity is one early example of the kind of monetization the market wants to see at scale. Without these disclosures, every earnings call will renew the same skepticism.
Second, the financing side. Watch for further credit-rating actions on hyperscalers with the largest off-balance-sheet obligations, spreads on their bonds relative to peers, and any signs that lenders are demanding tighter covenants on data-center leases. Oracle's downgrade is the first move; it will not be the last if leverage continues to climb.
Third, the correlation between spenders and suppliers. If the 30-day correlation between hyperscalers and the semiconductor index reverts toward its historical average of +0.60, it would signal that the market once again believes the two groups share a common fate. If it stays near zero, the decoupling is structural, and stock selection between the two groups will matter more than sector exposure.
Scenarios for the next 12 months split cleanly. In the base case, hyperscaler capital expenditure grows more slowly in 2027 than in 2026, AI revenue doubles but remains below 10 percent of spending, and stocks trade range-bound as the market waits for clearer returns. In the upside case, utilization rates surprise to the upside, a hyperscaler discloses a clear path to 20 percent returns on AI invested capital, and the correlation with chip stocks recovers — lifting the whole group. In the downside case, capital expenditure breaches 100 percent of operating cash flow for the group while AI revenue stalls, credit spreads widen, and the dot-com comparison moves from analyst notes to headline risk.
The short-term read is that volatility will remain elevated through earnings seasons, with each upward capex revision punished until monetization data improves. The medium-term read is that the companies with the strongest internal cash generation — Microsoft and Alphabet — are best positioned to fund the buildout without dilution or distress. The long-term read is that the AI infrastructure buildout is a genuine structural shift in the global capital stock, comparable to railroads and telecom, and that it will eventually generate real returns — but only for the investors who can endure a multi-year stretch in which the capital comes first and the returns come later.
The AI buildout is no longer a story about how much companies are willing to spend. It is a story about how much they can earn on what they have already spent — and the market has decided that question matters more than the next capital budget.
Explore more exclusive insights at nextfin.ai.
