Who Pays for AI? A $1.2 Trillion Hole in Wall Street's Forecasts

Market commentary, as of October 2, 2026. This is a theme piece, not a single-stock call. Opinion only, not investment advice.

The inconsistency

Last week Torsten Slok, Apollo's chief economist, pointed out something that should bother anyone who owns the S&P 500 right now.

• Sell-side analysts covering technology expect the sector's operating cash flow to rise by more than $1.2 trillion, to over $2.4 trillion by 2028. That is roughly double today's level.
• Analysts covering the rest of the S&P 500, the companies that would have to buy those AI and cloud services, forecast much smaller cash-flow gains.

Wall Street research is organized by sector. The tech analyst models explosive demand for AI. The analyst covering retailers, banks, insurers or manufacturers models those companies growing at a normal pace. Nobody's job is to check whether the two sets of numbers fit together. As Slok put it, both cannot be right.

So one of these has to give:

• The customers will grow much faster than their analysts expect, because AI makes them more productive and more profitable. If so, AI's users are mispriced today.
• Tech's forecasts are too high, because customers won't generate the cash to pay for all of this. If so, AI's builders are mispriced today.
• The money comes from somewhere outside both forecasts: debt, vendor financing, or other budgets being cut.

This post is about working out which of these is happening.

How big the bill is

The spending is unlike anything in recent market history:

• Amazon, Microsoft, Alphabet and Meta have guided to roughly $720-745 billion of capex in 2026. Individually: Amazon ~$200B, Microsoft ~$190B, Alphabet $180-190B and Meta ~$135B.
• Total global AI infrastructure spending is estimated at ~$750 billion this year and close to $1.1 trillion in 2027.
• UBS estimates that Amazon, Alphabet and Microsoft are spending about 102% of their cloud revenue on capex in 2026. Put differently, every dollar the cloud businesses bring in is going back into building more cloud. UBS projects about $4.1 trillion of hyperscaler capex from 2026 through 2028, more than three times what they spent over the previous six years combined.
• Free cash flow is already taking the hit. In Q1 2026, Amazon's free cash flow was negative $17 billion, and its trailing twelve-month FCF fell 95% to about $1 billion. Alphabet's quarterly FCF fell 46% year over year.

What comes back in

AI revenue is real and growing fast. But it's still far smaller than the spending.

• Anthropic's annualized revenue run rate reportedly went from about $9 billion at the end of 2025 to about $65 billion by July 2026. OpenAI is reported at close to $70 billion annualized in Q3. These are leaked run-rate figures, not audited revenue. OpenAI lost $38.5 billion on $13 billion of revenue in 2025.
• Total AI revenue across the industry is estimated at $150-200 billion a year.
• JPMorgan estimates that a $5 trillion AI buildout needs about $650 billion of revenue every year, permanently, just to earn a 10% return. That works out to about $35 a month from every iPhone user or $180 a month from every Netflix subscriber. Other estimates that include depreciation and a higher hurdle rate come out several times higher.

So the industry needs to grow its revenue roughly 3-4x just to hit a minimum return, while building out several more years of capacity. The growth is impressive. The gap is still very large.

Four places the money could come from

1. Corporate IT budgets (mostly money moving around, not new money)

Gartner now expects worldwide IT spending to grow 14.2% in 2026, to $6.37 trillion. The mix inside that number is what matters:

Almost all the growth is in buying AI hardware. IT services and communications are growing at about the pace of the overall economy. Gartner's own analyst says budgets are "being strained by inflation, supply shortages, rising hardware and memory costs, AI funding initiatives and shifting priorities." A CIO with a fixed budget who spends more on AI spends less on something else: consultants, legacy software seats, outsourced services. That moves revenue between tech companies. It doesn't create new cash flow for the sector, and it doesn't fill a $1.2 trillion gap.

2. Payroll (the bull case, and the one that matters)

This is the only source large enough to close the gap. U.S. employee compensation alone is well over $10 trillion a year. If AI lets companies do the same work with meaningfully fewer people, part of those savings can be paid to AI vendors, with the rest kept as higher margins.

The catch: if this is what happens, the customers' profits grow too. Each company replacing labor with AI should see its margins expand. That is exactly what the non-tech analysts are not forecasting. So the bull case for AI spending is, by its own logic, also a bull case for AI's customers, and probably a bigger one, because those companies are starting from depressed valuations.

The evidence so far is thin. In a recent survey of about 6,000 senior executives, roughly 90% reported no productivity gain from AI over the past three years. They expect only about 1.5% over the next three, or about 2.25% in the U.S. That could change quickly. But it isn't in the numbers yet.

3. The consumer (the weakest source)

JPMorgan's "$35 a month per iPhone user" comparison shows how much consumers would have to spend for consumer subscriptions to carry this. Consumer AI revenue is growing; OpenAI's Q3 consumer revenue reportedly beat its entire 2025 consumer total. But this is happening while consumer discretionary is at a 20-year low relative to the S&P 500, the 10-year Treasury yields more than 5%, the Fed is hiking and payrolls just came in at +29,000. A stretched consumer is unlikely to fund a trillion-dollar buildout.

4. Borrowing (where the money is actually coming from today)

When revenue doesn't cover spending, borrowing fills the gap. That is already happening:

• Hyperscalers are expected to issue about $420 billion of investment-grade debt in 2027. Morgan Stanley estimates more than half of hyperscaler spending will rely on outside capital.
• JPMorgan's funding plan for the buildout counts on ~$1.5 trillion from high-grade bonds, ~$200 billion from data-center securitizations and ~$150 billion from leveraged finance. That still leaves a ~$1.4 trillion gap, which it expects to be filled by private credit and government support.
• Vendor financing is growing. In August, Nvidia agreed to a $105 billion investment guarantee tied to an OpenAI data-center project, scaled back from roughly $250 billion discussed in July. Nvidia shares fell about 4.5% when the larger figure first came out. When a chip supplier finances its own customers' purchases, the resulting chip sales say less about real end demand.
• Credit markets are reacting. CoreWeave's borrowing cost has risen from about 5% to about 9%. Oracle's and Meta's data-center bonds trade in the low 90s, below their issue prices. Oracle's credit-default swaps are at a record high.

Borrowing can fund the build. It can't fund the return. Debt only buys time, and with the 10-year above 5%, that time is getting more expensive.

Where we come out

Add up the four sources. Corporate IT budgets mostly move money from one tech company to another. Consumers are stretched. Borrowing has to be repaid out of future revenue. The only source big enough to justify the forecasts is cost savings, mainly labor, at the customer. If that happens, the customers' earnings are too low today, not tech's.

That gives an asymmetry we like:

In both rows, the user side has better risk/reward. The builders need everything to go right. The users only need AI to help them a little. Because the market has crowded into the builders, the users are cheap. The top 10 stocks are about 38% of the S&P 500, while the median stock sits about 16% below its high.

Price is what you pay

The obvious pushback is that AI revenue is compounding faster than any software business in history. Anthropic went from a roughly $9 billion run rate at the end of 2025 to about $65 billion by July 2026. Why bet against that?

Our answer is that we aren't betting against the technology. We're questioning the price.

To be clear about what we are NOT saying: we are not calling a market top, and we are not predicting a repeat of 2000 or 2008. The economy is growing, earnings are strong and the leading companies are real businesses. Our point is narrower. The AI story is already priced in: today's prices assume years of future growth that hasn't happened yet. Buying now means paying in full for that growth up front, so even if it arrives, we wouldn't be rewarded for the risk. That's why we're not entering the trade, rather than because we expect a crash.

Two points explain why.

1. The law of large numbers

Going from $6 billion to $65 billion is a little over 10x. Going from $65 billion to $650 billion is also 10x, but it's a completely different task:

• The first jump added about $59 billion of revenue. The second needs about $585 billion, roughly ten times as many new dollars.
• $650 billion of revenue would be about twice what Microsoft brings in today. It's also about the same amount JPMorgan says the entire AI industry needs each year to earn a 10% return.
• Early growth comes from early adopters with the most obvious uses and the most money. Later growth has to come from smaller, more price-sensitive customers, and that is exactly where the cash-flow forecasts above are weakest.
• Competition and price cuts grow along with the market. The cost of a given level of AI capability has fallen quickly, so each new dollar of revenue takes more usage to earn.

2. How much of that growth is already priced in?

Anthropic is reportedly preparing to IPO at a valuation of more than $2 trillion, possibly before the end of the year, raising up to $100 billion. According to the leaked S-1 draft:

At $2 trillion, the stock would trade at about 43x annualized Q2 revenue and ~31x the $65 billion run rate. Working backward makes the point clearer. Suppose an investor wants a 10% annual return and assumes the company trades at a mature 25x earnings by 2032. It would then need about $140 billion of net income that year. At a 25% net margin, that is roughly $570 billion of revenue. In other words, most of the climb from $65 billion to $650 billion is already in the price. Investors buying at $2 trillion are paying today for an outcome that still has to happen.

The Amazon lesson

Amazon in 1999 was the right company. It became one of the most successful businesses ever. Yet its stock peaked in December 1999, fell about 94% by late 2001, and didn't get back to its 1999 high until late 2009. An investor who correctly picked the winner still waited a decade to break even.

Even if we buy the best company of the decade, we can lose money if we overpay. We think the risk/reward on AI's builders is unattractive at today's prices, whether or not AI turns out to be life-changing.

What breadth is telling us

The S&P 500 sits near a record high, but most stocks in it don't.

How to read this. The 200-day moving average is the most widely used dividing line between a stock in a long-term uptrend and one in a downtrend. A healthy bull market usually has 60-80% of stocks above it. When fewer than half are, and the index is still close to a record, a small group of very large stocks is holding the index up while most of the market is in its own correction. As of September 25, the share of stocks above their 200-day was in the bottom 7% of all readings taken this close to a record high. The last time breadth was this weak with the index near a record was April 2000. Two other readings point the same way: the number of stocks moving opposite to the index, and the lowest stock-to-stock correlation, are both at their most extreme since March 2000. In practice, the index has stopped reflecting what the typical stock is doing.

What has happened next, historically

In the short run (the next ~3 months), the record is mixed:

• In 13 earlier episodes of weak breadth near a record (1998-2024), breadth recovered without a major index decline 7 times. The index fell 5% or more within three months the other 5 times. Twelve-month returns afterward were roughly normal.
• The two closest matches, October 2018 (-18%) and December 2024 (-7%), both came with the 10-year yield at a one-year high, just like today (about 5.2%). Rates are the swing factor.

Over longer periods (years), the evidence favors the rest of the market over the leaders:

• Leaders tend to lag after reaching the top. Dimensional Fund Advisors found stocks typically do their best relative performance before entering the top 10, then underperform afterward. Intel beat the market by about 29% a year before joining the top 10, then lagged by about 6% a year for the following decade. Across 1971-2025, Bessembinder found that owning just the 10 largest stocks returned 10.7% a year, against 11.3% for the whole S&P 500. It worked in 2014-2025, but not in the four decades before that.
• Equal weight usually wins after a concentration peak. After the late-1990s peak, the equal-weight S&P 500 beat the cap-weighted index seven years in a row. Over 2000-2009 it returned about +65% while the cap-weighted index fell about 9%. After the Nifty Fifty peak, equal weight won 9 of 10 years (1974-1983).
• Smaller and cheaper stocks have led after every major peak. After the 2000 peak, the Russell 2000 Value index returned 183% from 2000 to 2006, versus 8% for the S&P 500. After the 1972 Nifty Fifty peak, small caps beat large caps by about 6.5 percentage points a year for a decade, but only after another painful year of underperformance first.
• The gap is historically wide right now. From 2023 to 2025, the cap-weighted S&P 500 beat equal weight by about 30 percentage points, the widest three-year gap since at least 1971 and wider than the late-1990s bubble. The cap-weighted index also trades at about a 30% valuation premium to equal weight, up from roughly even a decade ago.

Two caveats we take seriously:

• We're not using this to call a top. These are base rates, not a forecast, and we aren't predicting a 2000- or 2008-style decline. The leaders could keep doing fine in absolute terms. Our only claim is that, from here, the rest of the market offers better odds per dollar. History tells you what's likely, not when. Plenty of investors called a concentration peak in 2020. Concentration then rose another ~12 percentage points, and small caps kept lagging. Rotations like this can take years to start.
• Concentration today is partly earned. The top 10 make up ~41% of index weight and ~32% of its earnings. That's a far smaller gap than in 2000, when many of the leaders had little or no earnings behind their prices. This is closer to the Nifty Fifty than to the dot-com bubble: good companies at high prices, not bad companies at absurd ones.

Other data points worth weighing

• Plenty of new stock is coming to market. Equity issuance is running at about 3x the 2000 and 2007 peaks (closer to 2x adjusted for inflation). The Anthropic IPO alone could raise up to $100 billion. Heavy new supply has historically shown up near market tops.
• Investors are already heavily invested. Aggregate investor equity allocation is at its highest since 1999-2000. Stocks' rolling 10-year return lead over bonds is the widest since the 2000 peak. With so much money already in stocks, there's less new money left to push prices higher.
• IPOs are losing momentum. Average first-day IPO gains fell from 24% in June to under 1% by late September.
• Short-term flows. Pension funds were modeled to sell about $33 billion of U.S. stocks at quarter-end, in the 97th-98th percentile of estimates since 2000. That pressure should fade now that the quarter has turned.
• Seasonality helps us. Midterm-election years are historically weak into the vote and strong afterward. The fourth quarter of a midterm year has been one of the better stretches of the four-year presidential cycle.
• The economy isn't the problem. Atlanta Fed GDPNow is above 5%, ISM manufacturing has been in expansion 8 of the last 9 months, and Wall Street expects S&P 500 earnings to grow ~30% in Q3. This is a rotation and valuation story, not a recession call. That's also why we think the weak parts of the market have room to catch up instead of dragging the leaders down.

Where the value is: three dislocated sectors

We prefer the opposite setup, the one laid out in our dislocated-sectors framework: sectors where the bad news is already in the price, with plenty of upside if conditions merely stop getting worse. For reference, FactSet puts the S&P 500 at 19.0x forward earnings (25.7x trailing), and the index yields about 1.1%.

One note on that 19x: the index isn't expensive on paper because AI earnings are growing so fast. Wall Street expects S&P 500 earnings to grow 32% in 2026, with tech up 65% year over year in Q3. The multiple only looks reasonable if those forecasts hold, and the earlier sections of this post explain why we doubt they will.

1. Health care

• Trades at about 18x forward earnings, about a 20% discount to the S&P 500. By several measures that is its cheapest relative valuation in decades.
• Its 52-week market breadth is the weakest since the 2000 peak. Together with staples and bonds, it is one of the three most out-of-favor groups relative to the S&P since then.
• Why it's cheap: drug-pricing politics, patent cliffs, and money rotating into AI.
• Why it can work: aging populations, the GLP-1 wave, steady free cash flow and dividends. Health care is also one of the clearest places for AI to help: drug discovery, clinical trials, billing, claims and other administrative work.
• Examples: AbbVie at ~16x forward earnings; Novo Nordisk yielding ~3.9%.

2. Consumer staples, outside the expensive mega caps

• Staples are 4.5% of S&P 500 market cap, the lowest on record. Fund managers are a net 33% underweight, the most since January 2004. Combined, the consumer sectors are under 14% of the index, the smallest share since the 1990s.
• A caveat: the sector as a whole isn't cheap. Yardeni Research recommended underweighting it this summer because it trades near the market multiple with slow growth, propped up by expensive names like Walmart and Costco. The opportunity is in the packaged food, beverage and spirits companies that have been marked down hard:
• PepsiCo: ~18x forward earnings (vs. a 10-year median of 26x), 4.1% yield
• Hormel: ~12.9x forward earnings (vs. a 20-year average of 19x), 5.9% yield, the highest in its history
• Diageo: ~12.6x forward earnings (vs. a 25-year average of 18x), down ~60% from its 2022 high
• Why it's cheap: GLP-1 fears, consumers trading down, input costs, and dividend yields facing competition from a 5% 10-year Treasury.
• Why it can work: dividend yields of 4-6% that you're paid to wait on, and large cost bases (manufacturing, supply chain, marketing) where AI-driven efficiency would go straight to margins.

3. Consumer discretionary, outside Amazon and Tesla

• The headline sector P/E is actually the highest in the index at 22.6x. That's because more than 60% of the sector's weight is in three stocks, led by Amazon and Tesla. The rest of the sector tells a different story.
• Relative to the S&P 500, the sector hit a 20-year low this year, and its index weight is the lowest since the Global Financial Crisis.
• Why it's cheap: a Fed that is still hiking, a weak labor market (+29,000 payrolls today), and higher borrowing costs for households and for smaller, more leveraged companies.
• Why it can work: this is where the gap between the price and the worst case is widest. If oil and yields come down, these are the companies with the most to gain.
• Example: Disney at ~13.9x earnings (vs. a five-year average above 20x), with ~$10 billion of annual free cash flow.

Valuation figures are from late-summer and early-autumn 2026 reporting (FactSet, company data, published research) and will change with prices.

Our next thesis

That is what we'll look for in our next post: a company that benefits from AI and is undervalued because it sits in one of these distressed parts of the market. Specifically, we want:

• A large labor or operating cost base that AI can realistically shrink
• The scale, data and customers to actually deploy AI, not just talk about it
• No responsibility for funding the infrastructure. It buys AI as a service rather than building data centers
• A valuation that already reflects the bad news: a low multiple, a real dividend or free-cash-flow yield, and a balance sheet that can survive higher rates for longer

What could prove us wrong

• AI growth could defy the law of large numbers. If AI revenue keeps compounding at anything like this year's pace, the forecasts get closer to reality faster than skeptics expect, and today's multiples could look cheap in hindsight.
• The builders can afford a lot. Alphabet, Microsoft and Meta still produce large free cash flow from their core businesses. This isn't 1999-2001 telecom, when the companies doing the building had no other income. A capex slowdown would hurt suppliers more than it would hurt the hyperscalers themselves.
• Analysts covering the customers may simply be behind. Customer forecasts may be low because those analysts haven't modeled AI savings yet, not because the savings won't arrive. If so, the inconsistency resolves upward, which fits our thesis but not our timing.
• Value traps are real. Some of these stocks are cheap because their businesses are actually shrinking: GLP-1 demand changes, store brands taking share, drug-pricing reform. A low multiple on falling earnings isn't a bargain.
• The users could be cheap for other reasons. A Fed that keeps hiking, a weak consumer and $90 oil can keep "cheap" stocks cheap for a long time, whatever happens with AI.

Signals we're watching

• Hyperscaler capex guidance vs. cloud revenue growth on Q3 calls. Is the 102% ratio coming down?
• The Anthropic IPO. Watch the final S-1 numbers, the price and how it trades. A $2 trillion valuation on a ~$46 billion revenue run rate tests how much optimism public markets can absorb.
• Margins outside tech. Are non-tech companies citing AI in operating-margin gains, or only in their costs?
• AI credit spreads. CoreWeave, Oracle and data-center ABS. Credit tends to move before equity.
• More vendor-financed deals. Each new guarantee or equity-for-compute deal is a sign end demand isn't paying the full bill.
• Breadth. A sustained move back above 60% of stocks above their 200-day, with equal weight beating cap weight, would confirm the rotation is under way. A break below ~40% while the 10-year stays above 5% would point to the 2018-style outcome.

Sources

Apollo / Torsten Slok, "Tech's Trillion-Dollar Internal Inconsistency" (September 29, 2026); company 2026 capex guidance as compiled by industry trackers; I/O Fund Q1 2026 cash-flow analysis; UBS via 24/7 Wall St. (August 22, 2026); MIT Technology Review, "What must happen for AI's trillion-dollar gamble to pay off" (September 15, 2026); reported Anthropic and OpenAI run rates (August-September 2026, unaudited); JPMorgan data center and AI infrastructure financing analysis; Gartner IT spending forecast (July 27, 2026); Fortune and Axios on Nvidia-OpenAI financing (July-August 2026); Hedge Fund Tips Episode 363 recap (October 1, 2026) for AI credit data, sector positioning and individual stock multiples; Anthropic financials from the leaked S-1 draft as reported by Fortune and others (September 29, 2026); FactSet Earnings Insight (October 2, 2026); Schwab sector outlook (September 2026); Yardeni QuickTakes on consumer staples (2026); Investment House on health care valuation (June 2026); Motley Fool on PepsiCo (August 28, 2026); Amazon 1999-2009 price history; breadth data from Backtest (September 25, 2026) and S&P 200-day readings (September 30, 2026); Dimensional Fund Advisors via Forbes on top-10 stocks; Bessembinder, "Returns to 'Do-Nothing' Portfolios" via Novel Investor; Motley Fool on equal weight vs. cap weight (January 2026); GW&K, "After the Peak: What a Century of Market Concentration Teaches"; RBC Wealth Management, "The Great Narrowing"; market data as of October 1-2, 2026.