The $800 Billion Bet

Together, Amazon, Alphabet, Microsoft, Meta and SpaceX expect to spend $803–828 billion on capital projects in calendar 2026.^{1,2,3,4,5,6,7} The four big cloud providers among them, known as hyperscalers, spent $201.9 billion in the third quarter alone, more than four times their rate at the start of 2024.^2,3,5,8 Their combined free cash flow also turned negative, for the first time in at least a decade.^2,3,5,8 Our reading is that the bet pays off only if AI revenue grows several-fold before the chips it buys wear out.

How we count. The total covers capital spending of every kind, though AI drives the overwhelming share of it. As Business Insider notes, “That figure covers broad capital spending, not just on AI.”¹

01 — The SpendTripled in Two Years, and Still Climbing

Spending has risen steeply for two years, and most of the companies expect to spend more again in 2027.

Capital expenditure, or capex, is spending on long-lasting assets such as buildings, servers and chips. On a cash basis, Amazon, Alphabet, Microsoft and Meta spent $228 billion on it in 2024 and $376 billion in 2025.^3,5,8,9,10 Their 2026 guidance adds up to $720–740 billion, roughly triple 2024.^1,2,4,5,11

Each company’s 2026 plan, on its own definition:

• Amazon. Expects “approximately $220 billion” of cash capex.²
• Alphabet. $195–205 billion, a range it held.¹¹
• Meta. $135–145 billion after narrowing its outlook, against $72.2 billion spent in 2025.^5,10
• Microsoft. About $170 billion in calendar 2026, including finance leases.^4,12 It has trimmed the figure twice, without cutting planned building (see chart note).
• SpaceX, which now includes xAI. No annual figure. Its nine-month spending of $48.1 billion plus fourth-quarter guidance of $35–40 billion implies $83–88 billion.^6,7

For 2027, Amazon’s Andy Jassy expects to spend “meaningfully more.”¹³ Microsoft expects spending to accelerate in the first half of the year,⁴ and Meta will “continue investing aggressively.”¹⁴ Elon Musk is “fairly confident of being able to bring an incremental 10 gigawatts of compute online in 2027.”⁷

Further out, Goldman Sachs expects the five largest US hyperscalers, a group that includes Oracle but not SpaceX, to spend $1.2 trillion in 2027 and $1.4 trillion in 2028, according to Bloomberg.¹⁵

02 — The CashWhen Spending Outruns Cash

The companies are now spending more than their businesses bring in, so they are borrowing and selling shares to cover the gap.

Free cash flow is the cash a business brings in from its operations, minus its capex. Put simply, it is what is left after paying for the build. For the four hyperscalers combined, it swung from +$61.7 billion in the third quarter of 2025 to –$26.4 billion a year later. That was the group’s first negative quarter in at least a decade.^2,3,5,8

Microsoft is the exception among the four, with free cash flow down 27% to $18.6 billion, and it expects to stay positive.^4,8

Other people’s money is filling the gap.

• Debt. Long-term debt at Amazon, Alphabet and Meta rose from $170.9 billion to $379.1 billion in nine months.^2,3,5
• Equity instead of buybacks. Alphabet bought back no stock in the first nine months of 2026, against $39.8 billion a year earlier, and raised about $61.9 billion of equity.³
• New shares. On 26 October, Meta sold about $37.5 billion of mandatory convertible preferred and common stock. It also announced plans for a roughly $60 billion at-the-market program, which sells new shares gradually, and cited “general corporate purposes.”^16,17 Five days earlier, Mark Zuckerberg had said Meta had no “specific financing plans to announce.”¹⁴
• Bonds. Microsoft came to the bond market for about $25 billion, its first US investment-grade sale in more than nine years, and drew more than $100 billion of orders, according to Reuters.¹⁸
• Leases. SpaceX has arranged a roughly $40 billion finance lease, led by Apollo and Goldman Sachs, for Nvidia chips delivered from 2027. Leased chips will not show up in cash capex,¹⁹ so a flat capex line would not necessarily mean less building.

03 — The MathWhat the Build Has to Earn

On our simple model, the 2026 build alone needs far more yearly revenue than the leading AI labs earn today, and how long the chips last matters most.

A simple payback model

Our model is illustrative, not a forecast. We start with the roughly $800 billion that Goldman Sachs expects the five largest US hyperscalers to spend in 2026.¹⁵ Following McKinsey’s estimate of the hardware share, we assume 60% goes on chips, servers and other short-lived hardware.²⁰ That hardware must pay for itself over five years, while the other 40% goes on buildings and power equipment that last 20 years.

Each year, the build has to repay its cost plus a 10% pre-tax return, much like a mortgage payment on the equipment. That capital charge works out to about $164 billion a year. If cash operating costs take 40% of revenue, the 2026 build needs about $275 billion of revenue a year for five years. Just breaking even needs about $187 billion.

Reasonable changes to the assumptions give a range of $250–330 billion a year, from the 2026 spending alone. Goldman, which has not published its method, puts the revenue needed to cover the spending at about $300 billion a year.¹⁵

The 40% cost share looks reasonable for a mature cloud business. AWS’s operating margin of 38.7%, which is measured after depreciation, implies cash costs well under 60% of revenue.² Margins on raw chip rental can be far thinner. Oracle earned a gross margin of about 14% renting Nvidia-based servers in an early quarter of 2025, according to documents reported by The Information.²¹

How today’s AI revenue compares

OpenAI’s annualized revenue was nearing $50 billion at the end of September, on a net basis.²² Anthropic told investors its third-quarter revenue was $16.8 billion, about $67 billion annualized, on a gross basis that includes sales through cloud partners.²³

Together, that is roughly $117 billion a year, or about $137 billion with OpenAI counted on a gross basis like Anthropic’s. The 2026 bar is 2–2.4 times as much, by our calculation, and much of the labs’ revenue is paid on to the hyperscalers anyway. Advertising gains at Meta and Google, and AI used inside the companies, are not captured here.

How long the chips last

The chips’ useful life is the biggest swing factor. At Microsoft, “roughly three quarters” of the latest quarter’s capex went on “short-lived assets.”⁴ If the hardware earns money for three years instead of five, the bar rises to about $384 billion a year. At six years, it falls to about $246 billion (our calculation, at 40% cash costs).

Depreciation is the yearly charge that spreads an asset’s cost over its life, much as a delivery van’s cost is spread over the years it is used. Michael Burry, the “Big Short” investor, alleges that hyperscalers will understate it by $176 billion over 2026–28 by stretching useful lives. CNBC could not confirm his figures.²⁴

Rental prices so far point the other way. The Silicon Data index for renting Nvidia’s H100 chip stood at $2.96 per GPU-hour on 3 November, up from $2.81 on 5 October.²⁵ AI output, though, keeps getting cheaper. For equal capability, Epoch AI finds the price of running AI models, known as inference, falling 9 to 900 times a year, depending on the benchmark.²⁶

04 — The PrecedentsThe Railways Could Wait. Chips Can’t.

Past infrastructure booms paid off for society but often not for the people who financed them. In our view, the main difference this time is asset life.

In 1847, investment in British railways reached about £44 million, or 7.3% of GDP. In 1852, the railways took in £15 million, a quarter of what promoters had expected. Andrew Odlyzko estimates that investors lost about a third of the roughly £250 million they had put in by 1850.²⁷

Fiber followed the same path. Telecom firms “poured more than $500 billion into laying fiber optic cable” in the five years after the 1996 Telecommunications Act, according to Robert Litan of Brookings. By 2002, “no more than two percent” of North American long-distance capacity was in use, and telecom shareholders had lost roughly $2 trillion.²⁸

Today’s boom is big too. US investment in information-processing equipment and software reached 5.35% of GDP in the third quarter of 2026, above the 4.46% peak at the end of 2000.²⁹ That measure, however, covers all IT, includes imported equipment and will be revised.

Rails and cable last for decades, so they could wait for demand to catch up. Chips with a working life of about five years cannot wait as long.

AWS itself offers a bullish precedent, producing 57% of Amazon’s operating income in 2025 at a 35.4% margin. Yet Amazon’s 2026 plan of about $220 billion is roughly 4.8 times AWS’s 2025 operating income of $45.6 billion, by our calculation.^2,9 As the economist Carlota Perez put it, “I have not seen a golden age happening without a crash.”³⁰

05 — The Verdict So FarReal Demand, Unproven Returns

Demand for AI computing is real and growing fast, but there is little proof yet that it will pay for the build.

The case for

AWS revenue grew 39% to $46.0 billion in the third quarter, its fifth straight quarter of faster growth.² Microsoft’s Azure grew 48%. The company’s commercial remaining performance obligation, its backlog of contracted revenue not yet booked, reached $677 billion.^4,8

Google Cloud revenue roughly doubled, rising 103% to $30.7 billion. That figure includes sales of TPU hardware systems, so the growth rate is not comparable with earlier quarters.³ “Customer demand continues to exceed available supply,” said Microsoft’s finance chief, Amy Hood.⁴

The case against

The bears point to concentrated customers and thin evidence of broad gains. About 45% of Microsoft’s commercial backlog came from OpenAI at the end of 2025, the latest split we have.³¹ OpenAI itself projects negative free cash flow of $278 billion over 2026–2030, according to the Financial Times.³²

Use is broad but shallow. In November 2025, 40.7% of US workers used generative AI for their jobs, but only 12% used it daily.³³ Markets have turned less forgiving. On 29 October, the day after its results, Alphabet closed down 12.6%, its steepest one-day fall since it listed in 2004, according to Investing.com data.³⁴

The physical bill is arriving as well. Grid operator PJM runs a capacity auction, in which the grid pays power plants in advance to be available years ahead. Its prices have cleared at the cap three times running, and the last two auctions fell short of the reliability target.^35,36 Data centers accounted for nearly 5,100 of the 5,250 megawatts of extra forecast demand for 2027/28.³⁵

The builders’ view

Even the builders concede the risk while arguing that holding back is riskier. Zuckerberg said in 2025 that misspending “a couple of hundred billion dollars” would be “very unfortunate,” adding, “I actually think the risk is higher on the other side.”³⁷ Microsoft’s Satya Nadella told Davos in January: “For this not to be a bubble, by definition it requires that the benefits of this are much more evenly spread.”³⁸