Issue #283

Big Tech's Capex Is Half of Wall Street's Profit Growth

Goldman's $1.2 trillion capex forecast hides a bigger number: how much of it is propping up S&P 500 earnings.

Business

Oswarld (Kwangseob Ahn)· Writer & editor of INLEVEL9

Add as a Google preferred source
Big Tech's Capex Is Half of Wall Street's Profit Growth

The English draft accurately matches the Korean source with correct numbers, headings, footnotes, and links. No Hangul remains. No corrections needed.

A Number You Should See Before the $1.2 Trillion

On September 25th, Goldman Sachs put out a forecast for how much America’s five hyperscalers1 will spend on capital expenditures2 next year: $1.2 trillion. That’s nearly 50% more than this year’s roughly $800 billion, and higher than Wall Street’s consensus of $1.1 trillion. For 2028, Goldman sees the figure climbing to $1.4 trillion. Most headlines led with that number.

But read that alongside a report Goldman’s U.S. equity strategy team published two days earlier, and a different number stands out. The team estimates that nearly half of this year’s growth in S&P 500 earnings per share (EPS)3 came from AI-related spending. And if hyperscaler investment next year comes in $250 billion above or below forecast, S&P 500 earnings growth moves roughly 6 percentage points in the same direction.

It’s worth pausing on just how unusual this year’s U.S. corporate earnings have been. S&P 500 EPS rose 51% year-over-year in Q2, and 26% over the trailing four quarters. The 30-year average is 7%. Nearly half of that extraordinary growth is being carried by the spending plans of a handful of companies. At this point, where next year’s investment lands—somewhere between Wall Street’s consensus and Goldman’s forecast—can shift U.S. corporate earnings growth by several percentage points on its own.

Today, I want to trace how that link works—and why Goldman thinks AI investment’s contribution to earnings will shrink even as spending keeps rising.

The Same Dollar Gets Booked at Different Speeds on Two Ledgers

When a hyperscaler spends $1 on a data center, that dollar instantly becomes someone else’s revenue: the chipmaker building GPUs, the companies making servers and network gear, the industrial firms selling power equipment, the utilities supplying the electricity. This is the channel Goldman points to when it says AI investment is flowing into these sectors’ profits. On the seller’s side, revenue and profit are booked the year the goods change hands.

The buyer’s ledger moves differently. For the hyperscaler, that same dollar isn’t this year’s expense—it’s an asset. Servers and equipment get depreciated4 over several years, chipping away at the cost bit by bit. I covered why the choice of useful life matters so much in Where Did the Railroad King’s Dividends Really Come From?; here, let’s just focus on the timing gap.

While spending is climbing steeply, this creates a window in the economy’s aggregate ledger where profits look outsized—the seller books full revenue at once, while the buyer books only a fraction of the cost. But once the pace of spending growth slows, depreciation on all the equipment bought over the past few years keeps piling up. Newly added revenue shrinks, even as the costs of equipment purchased in prior years keep accumulating year after year.

This is exactly the shape of Goldman’s forecast. Investment keeps rising—about $800 billion this year, $1.2 trillion next year, $1.4 trillion by 2028. But the growth rate itself falls, from 94% this year to around 50% next year, and lower still by 2028. According to reporting on the underlying report, Goldman calculates that while next year’s capex adds roughly 11 percentage points to S&P 500 earnings growth, hyperscaler depreciation expenses subtract about 5 percentage points—wiping out nearly half of that boost. And by 2028, Goldman expects AI investment to flip into a slight net drag on earnings.

What pushes earnings higher isn’t just the size of the investment—it’s the speed at which it keeps growing. So there’s no contradiction in AI investment’s contribution to earnings shrinking in the very year spending hits $1.2 trillion.

Another AI Hiding Inside the Profits

Inflating expenses isn’t the only way profits get puffed up. Goldman found that big tech companies booked roughly $150 billion in unrealized gains5 on their private-company equity stakes in Q2 alone this year — equivalent to 12% of S&P 500 EPS. Goldman expects more of these gains to show up in the second half, but far less next year. The report itself notes that these paper gains are temporarily inflating earnings.

So there are two strands of “AI” tucked inside this year’s S&P 500 profits: the portion where hyperscaler spending flows through as revenue for the supply chain, and the portion booked as private-company equity stakes appreciate in value. Both are legitimate profits by accounting standards. But neither represents money that companies and individuals actually using AI have paid for it.

Semiconductors carry one more variable: price. Demand is strong and supply is tight, pushing memory makers’ gross margins6 up to around 80% — more than double the historical average. Goldman found that a large share of recent semiconductor profit growth has come from this margin expansion, and expects the effect to fade going forward. As capacity comes online and the supply crunch eases, prices and margins could drift back toward normal levels.

This variable applies directly to Korean memory makers too. As I noted in SanDisk’s four-year pledge, memory makers stayed cautious about expanding capacity even as demand surged, while customers locked in volumes through long-term contracts. How long that 80% margin holds depends as much on when and how much the supply side ramps up equipment as it does on demand.

Between $300 Billion and $70 Billion

So what should we use to judge the AI investment cycle? Goldman’s strategy team put a number on it in their September 25th note. Their calculation: hyperscalers need roughly $300 billion in annual AI revenue within the next few years just to break even on this investment.

Let’s look at how far along we are. According to Goldman, hyperscaler cloud revenue has grown steeply this year — as of Q2, it’s running about $70 billion above the pre-generative-AI-boom trend on an annualized basis. Announced backlog exceeds $1.5 trillion. That’s enough to prove demand exists. But $70 billion isn’t a direct measurement of AI revenue — it’s cloud revenue deviating from trend — so it’s not measured on exactly the same ruler as the $300 billion figure. Even accounting for that, the gap is nearly fourfold, and backlog is a promise of future revenue, not money already collected.

The bar rises much higher if you’re looking past breakeven toward real profitability. Goldman estimates that for hyperscalers to earn adequate returns on investment, and for the AI application companies sitting above them to cover compute costs while still keeping a decent margin, AI users need to spend roughly $1 trillion a year on AI applications. Global software spending this year is about $1.5 trillion. That means the world would need to direct roughly two-thirds of everything it currently spends on all software toward AI applications alone.

HyperThis is where the number that matters shifts. Up to now, the AI investment debate has revolved around how much hyperscalers are spending. Profits flow to the supply chain from that spending, but ultimately that spending has to be recovered from the people actually using AI. Where companies pull their AI budgets from — and whether that budget replaces existing software spending or gets added on top of it — becomes the more important number going forward.

Worth noting too: Goldman’s base case isn’t a profit collapse but a slowdown. Goldman expects S&P 500 EPS to grow roughly 11% each in 2027 and 2028, with AI productivity effects increasingly filling in as the tailwind from AI investment itself weakens. Stock prices haven’t fully kept pace with this year’s earnings growth, so the forward P/E ratio has come down from 23x a year ago to 19x now. As for fears of an earnings bubble, Goldman’s answer is that current valuations show no signs of a bubble — but if today’s earnings don’t hold up, even an average multiple could turn out to be expensive.

Oswarld’s Lens

When I design projects that bring AI into air-gapped public-sector environments, the first thing I do is estimate what open-source and open-weight models will look like 6-12 months out, and I plan around that. I also find myself repeating that not everyone, and not every task, needs the most advanced frontier model. I think there’s a model that fits each job and each environment.

Set that experience next to the $1 trillion figure, and it becomes clear that the “money users have to pay,” as Goldman put it, might not all flow to hyperscaler cloud. Buyers are already factoring in the point at which cheaper, good-enough alternatives will show up. If the supply chain’s profits depend on how fast hyperscalers invest, hyperscalers’ returns depend on what model buyers decide to use for which task, and at what price. That’s why, when I look at the AI investment cycle, I pay as much attention to which models enterprise customers actually choose—and where their budgets come from—as I do to capex announcements.

Closing

Goldman’s $1.2 trillion forecast is a signal that AI infrastructure investment hasn’t peaked yet. But the same week, two other Goldman reports showed that the force this investment exerts in lifting U.S. corporate earnings is already near its own peak. Investment keeps rising even as its earnings-boosting power fades — and what has to fill that gap is spending from the companies actually using AI.

At the next round of Big Tech earnings, watch capex guidance side by side with cloud and AI revenue growth — that’s what will show you whether this gap is narrowing or widening.

This piece is not investment advice regarding any specific stock or asset. The figures cited are Goldman Sachs’s own estimates and projections.

💬 Is your company’s AI budget for next year going up or down from this year? If it’s going up, let me know in the comments — is it new money, or is it being shifted over from other software budgets?

📨 If you have a colleague who often talks AI investment and stock markets, pass this piece along.

Your take shapes the next issue

What resonated most in this issue, or where has your experience been different?

Any registered reader can comment for free.

References & Further Reading

Primary sources

Related issues worth reading

Illustrated portrait of Kwangseob Ahn (Oswarld)

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

📝 Glossary

Footnotes

  1. Hyperscaler: A cloud operator — like Google, Amazon, or Microsoft — that builds and runs massive data centers itself. Goldman’s forecast here is based on the top five U.S. hyperscalers. ↩

  2. Capex (capital expenditure): Money spent on assets used over multiple years, like buildings, equipment, and servers. Instead of being expensed all at once in the year it’s spent, it’s recorded as an asset. ↩

  3. EPS (Earnings Per Share): A company’s net income divided by its share count. S&P 500 EPS converts the combined earnings of index constituents into an index-level figure, useful for tracking the overall earnings trend of large U.S. public companies. ↩

  4. Depreciation: The accounting treatment of spreading the cost of an asset over its useful life, expensing a portion each year. The longer the useful life assumed, the smaller the expense recorded in any given year. ↩

  5. Unrealized gains: Gains recorded on the books when the value of a held stake rises, even though it hasn’t been sold. If the value falls instead, a loss is recorded. ↩

  6. Gross margin: The share of revenue left after subtracting the direct cost of producing what was sold. An 80% margin means that for every 100 won of sales, the cost was 20 won. ↩

Related issues

Issue #282

Apple Spells Out Its On-Device-vs-Server AI Decision Order

Apple's own docs test on-device models on real tasks first, escalating only when limits are hit.

Business
Issue #280

They Said 'Slow Down'—But Compute Kept Climbing

It was never about halting training—what's actually slowing is release timing, not capability.

Business
Issue #279

The White House Fact Sheet Renames AI 'Superintelligence'

A three-line US-China AI pact borrows a pre-summit proposal's name but skips its accident-response procedures.

Business