The $800 Billion Bet — The Full Research Note

Together, Amazon, Alphabet, Microsoft, Meta and SpaceX expect to spend $803–828 billion on capital projects in calendar 2026, on our sum of each company’s guidance.^{1,2,3,4,5,6,7,8} In the third quarter alone, the four hyperscalers spent $201.9 billion, more than four times their rate at the start of 2024. For the first time in at least a decade, their combined capex exceeded their combined operating cash flow.^2,3,6,9

This is capital spending, overwhelmingly driven by AI, not pure AI spending. As Business Insider put it, “That figure covers broad capital spending, not just on AI.”¹ Amazon says its year-on-year increase “primarily reflects investments in artificial intelligence.”²

This note asks what would have to be true for roughly $800 billion of 2026 spending to pay off, and for whom.

01 — The Spend$800 Billion, and Climbing

Five companies plan to spend more than $800 billion this year.¹ For the four hyperscalers among them, that is roughly triple their 2024 level, and most of the spenders expect to spend more again in 2027.

Where the $800 Billion Comes From

The headline number is built from five sets of guidance. Amazon says, “We continue to expect 2026 cash capital expenditures to be approximately $220 billion,” up from “about $200 billion” in February.^2,10,13 Alphabet held its guidance at $195–205 billion.⁴ Microsoft expects about $170 billion in calendar 2026.⁵ Meta narrowed its outlook to $135–145 billion, up from $115–135 billion in January; it spent $72.2 billion in 2025.^6,12

Those four add up to $720–740 billion.^2,4,5,6 SpaceX, which merged with xAI in February, gives no annual figure, but its nine-month spending of $48.1 billion plus fourth-quarter guidance of $35–40 billion implies $83–88 billion.^7,8

No company guided the $803–828 billion total, and the companies define capex differently. Amazon’s figure is “cash capex” and Alphabet’s is purchases of property and equipment. Microsoft’s covers calendar 2026 and includes finance leases, Meta’s includes principal payments on finance leases, and SpaceX’s counts cash purchases and excludes finance leases.^2,3,5,6,7 The sum is a guide to scale, not a like-for-like accounting total.

Roughly Triple in Two Years

On a cash basis, the four hyperscalers spent $228 billion in 2024 and $376 billion in 2025. On their headline definitions, the totals are about $250 billion and $414 billion. On that basis, 2026 guidance is a rise of 74–79% in a year and roughly triple the 2024 level.^{3,6,9,10,11,12}

The quarterly figures show how fast it has accelerated. On a cash basis, the four spent $44.3 billion in the first quarter of 2024 and $118.6 billion in the last quarter of 2025. In the third quarter of 2026 they spent $201.9 billion, or $207.4 billion on headline definitions. That was up $36.9 billion on the second quarter and 356% since early 2024.^2,3,6,9,10

With SpaceX, the five spent about $227 billion in the quarter by Business Insider’s tally of headline figures.¹ SpaceX’s own capex was $19.6 billion, against $6.9 billion a year earlier. Of that, $16.4 billion (84%) went to its AI segment, which houses xAI, X and Cursor.⁷

Microsoft’s Cut Is Not a Cut

Microsoft’s figure has fallen twice this year, and neither fall means less building. In April it guided to about $190 billion, including “approximately $25 billion from the impact of higher component pricing.”²⁷ In our reading, that is a sign the build is inflating the cost of its own inputs.

In July the figure fell to about $175 billion after Microsoft extended the useful lives of its data centers and office buildings from 15 to 25 years. That change moves future data-center leases from finance leases, which count as capex, to operating leases, which do not. “Outside of this useful life impact,” Microsoft said, “our calendar year 2026 CapEx investment expectations remain unchanged.”¹¹

In October it fell to about $170 billion “due to the 3P capacity delays,” as third-party capacity slipped into 2027.⁵

Leases create a wider blind spot. GPUs and buildings taken on leases, such as SpaceX’s $40 billion GPU lease starting in 2027, do not show up in cash capex.^11,22 From 2027, capex totals will understate the capacity being added, so a flat capex line should not be read as less building.

Most Signal More for 2027

Four of the five point to higher spending next year.

• Amazon. Chief executive Andy Jassy said Amazon expects to spend “meaningfully more” in 2027, adding, “We can fund this plan.”¹³
• Alphabet. CFO Anat Ashkenazi said, “We continue to expect our CapEx to increase significantly in 2027.”⁴
• Microsoft. CFO Amy Hood expects capex to accelerate in the first half of 2027, put the October–December quarter at “over $50 billion” and said fiscal 2027 capex will “grow year-over-year.”^5,11
• Meta. CFO Susan Li said it would keep “investing aggressively” but had nothing “more specific or quantitative to share until we get to the Q4 call.”²⁸
• SpaceX. Elon Musk said, “I’m fairly confident of being able to bring an incremental 10 gigawatts of compute online in 2027.”⁸

Forecasts keep rising, too. In figures reported by Bloomberg, Goldman Sachs projects capex for the five largest US hyperscalers at $1.2 trillion in 2027 and $1.4 trillion in 2028. Its 2027 figure compares with a consensus of $1.1 trillion, and its group includes Oracle but not SpaceX.^16,26 The Bank of England notes the market forecast for 2028 hyperscaler capex is now over $1 trillion, up from under $600 billion at its December 2025 report.²⁹

02 — The Cash SqueezeWhen Capex Outruns Cash

For a decade, Big Tech paid for data centers out of the cash its businesses threw off, with plenty left for buybacks. That changed in 2026. The build now costs more than the businesses bring in, and the accounting cost will keep rising for years after the cash is spent.

Spending Now Exceeds Cash Coming In

In the third quarter, capex exceeded operating cash flow, the cash generated by day-to-day business, at four of the five big spenders. Alphabet’s capex was 121% of its operating cash flow and Meta’s 106%; Amazon’s was 182% and SpaceX’s 203%. Microsoft was the exception at 78% (68% on a cash basis).

Measured against revenue, capex ran at 38–54% at the four hyperscalers and 134% at SpaceX, while Oracle’s fiscal 2026 capex was 83% of its revenue.^{2,3,6,7,9,30,31}

Combined free cash flow at the four hyperscalers was –$26.4 billion in the third quarter, down from +$61.7 billion a year earlier. That swing of $88.2 billion gave the group its first negative quarter in at least a decade.^2,3,6,9

The squeeze is uneven across the companies.

• Amazon’s trailing-twelve-month free cash flow was –$42.4 billion on its own definition, against +$14.8 billion a year earlier.²
• Alphabet’s was –$9.1 billion in the quarter, partly because inventory of its TPU chips rose from $2.4 billion to $13.3 billion in nine months.³
• Meta’s was –$2.1 billion on its own definition, against +$10.6 billion a year earlier, though its nine-month total is still +$11.0 billion.⁶
• Microsoft is the counterweight, with free cash flow of $18.6 billion, down 27%, and Hood expects it to “remain free cash flow positive in FY27.”^5,9,11
• SpaceX publishes no quarterly cash-flow statement, but by our calculation its free cash flow was about –$10.0 billion in the third quarter and about –$35.0 billion over nine months.^7,30

The Bill That Arrives Later

Capex hits cash at once but reaches profits gradually, through depreciation, the accounting charge that spreads an asset’s cost over its useful life. A delivery firm does the same with a new van, booking part of its cost in each year the van is on the road.

Depreciation and amortization at the four rose 25–49% year on year, to about $48 billion in the third quarter (definitions differ by company).^2,3,6,9 Because each year’s spending adds a layer, the charge will keep climbing even if capex flattens. At Microsoft, most of the money goes into short-lived assets: “Roughly three quarters of our capex was for short-lived assets, primarily CPUs and GPUs,” Hood said.⁵

Companies are now moving useful lives in both directions. With effect from January 2025, Amazon shortened the life of “a subset of our servers and networking equipment from six years to five years.” It cited the “increased pace of technology development,” and the change cut 2025 operating income by about $700 million.³³ Microsoft lengthened building lives, with “minimal benefit to FY27 operating income.”¹¹

GPU Rental Margins Can Be Thin

Even chips that hold their value may not be lucrative to rent. According to internal documents reported by The Information, Oracle earned about $125 million of gross profit renting Nvidia-based servers in the three months to August 2025. That was about 14% of $900 million of revenue.³⁷ That was an early-ramp quarter, but it shows how thin raw GPU rental can be next to a mature cloud business.

03 — The FinancingFrom Cash to Credit

By our calculation, the four hyperscalers’ cash from operations covered about 87% of their capex in the third quarter, so the rest had to come from somewhere else.^2,3,6,9 In 2026 it came from bonds, new shares and leases, all at once. That spreads the risk well beyond the companies’ own shareholders.

Borrowing

In the nine months to September, long-term debt at Amazon, Alphabet and Meta rose by $208 billion, from $170.9 billion to $379.1 billion. Amazon’s alone went from $65.6 billion to $173.1 billion.^2,3,6

Even Microsoft has joined in. In late October it came to the bond market for about $25 billion, its first US investment-grade sale in more than nine years. According to Reuters, the sale drew more than $100 billion of orders.³⁸ Hood had said on the 28 October earnings call, “we will access additional sources of funding when we believe it is attractive to do so.”⁵

Demand for the debt is strong but cooling. Apollo data cited by Forbes show hyperscaler bond order books falling from about five times covered in February to under two times by July.³⁹ Morgan Stanley estimated in 2025 that about $1.5 trillion of the $2.9 trillion of global data-center capex expected in 2025–28 would have to come from outside capital. About $800 billion of that would come from private credit.¹⁸

Issuing Shares

The sharper change is in equity. Alphabet has gone from buying back its shares to selling new ones, and Meta has followed.

Alphabet bought back no stock in the first nine months of 2026, against $39.8 billion a year earlier. Instead, it raised about $61.9 billion of equity: $42.9 billion of common stock and $19.1 billion of mandatory convertible preferred stock. The common stock included $12.4 billion sold in the third quarter through its at-the-market program, which sells new shares gradually at market prices.³

On 21 October Mark Zuckerberg said, “We don’t have any specific financing plans to announce at this time.”²⁸ Five days later, Meta sold about $37.5 billion of mandatory convertible preferred and common stock. It also plans an at-the-market program of about $60 billion once its new non-voting Class C shares start trading, which is expected in the first quarter of 2027. The stated use is “general corporate purposes.”^21,28

SpaceX raised $75 billion in its June IPO, about $86 billion with the over-allotment, then sold a debut $25 billion bond.⁷

Leases and Joint Ventures

The third route keeps spending off the capex line. In a lease, someone else owns the building or chips and the hyperscaler pays rent; in a joint venture, outside investors own most of the project.

Company disclosures on different definitions and dates; do not add the rows together.

Oracle’s $288 billion of uncommenced leases is more than twice its notes payable.⁴⁰ Its fiscal 2026 capex of $55.7 billion left it with free cash flow of –$23.7 billion.³¹ OpenAI reportedly signed a five-year computing contract with Oracle worth roughly $300 billion, starting in 2027; neither company has confirmed the amount.⁴² On 2 November Oracle filed to sell $10 billion of common stock and, separately, $10 billion of senior notes.⁴³

Stress Shows Up at the Edges

The strain is clearest at Oracle and the neoclouds. On 9 July, S&P cut Oracle to BBB–, the lowest investment-grade notch, with a stable outlook. It reportedly forecasts a free-operating-cash-flow deficit of about $42 billion in fiscal 2027.⁴⁴

CoreWeave, a “neocloud” that rents out GPU capacity, is further along. Its five-year credit default swaps, the market price of insuring its debt, rose above 1,000 basis points in October, which Bloomberg Intelligence said priced about a 60% probability of default.⁴⁵ Under a September 2025 order running to April 2032, Nvidia is obliged to buy CoreWeave’s unsold capacity, subject to termination rights. The order, with an initial value of $6.3 billion, is a capacity-purchase agreement rather than a debt guarantee.⁴⁶

Circular Money

Much of this financing loops back on itself, as suppliers invest in customers who buy from them. The IMF reviewed 14 such “circular” transactions since 2023 and found “about $300 billion in equity investments and $650 billion in debt, in exchange for $1.4 trillion in future revenue commitments.”⁴⁷

04 — The DemandReal, Fast-Growing and Concentrated

The case for the spending rests on numbers the companies report every quarter, and those numbers are strong. The weak points are that much of the demand traces back to two AI labs and that broad productivity gains are not yet proven.

Cloud Growth Is Speeding Up

• AWS, Amazon’s cloud arm, grew 39% to $46.0 billion in the third quarter, its fifth straight quarter of faster growth, at an operating margin of 38.7%. Jassy projects growth “of over 40% both in 2027 and in 2028.”²
• Microsoft’s Azure grew 48%.^5,9
• Google Cloud revenue roughly doubled (+103%) to $30.7 billion, and its operating margin rose to 36.2% from 23.7%. That revenue includes sales of TPU hardware systems, which Alphabet says were “a meaningful contributor to revenue growth,” so the growth rate is not comparable with earlier quarters.³

According to Bloomberg, Goldman calculates that cloud revenue growth at Amazon, Alphabet, Microsoft and Oracle accelerated to 48% in the second quarter of 2026 from 25% in 2024.^16,26 “Customer demand continues to exceed available supply,” Hood said.⁵

Nvidia’s results point the same way. Its data-center revenue was $89.0 billion in the quarter to 26 July, up 117%, at a 75.0% gross margin. For the quarter to October, it guided to $108 billion (±2%) of total revenue, with results due on 17 November.^24,48 Annualized, that data-center revenue of about $356 billion equals 43–44% of the five companies’ 2026 capex by our calculation, though Nvidia sells worldwide.⁴⁸

Huge Backlogs, Slow Conversion

Backlog, or remaining performance obligations (RPO), is revenue customers have contracted for but not yet received, much like a builder’s order book. The four below each exceed half a trillion dollars but turn into revenue at very different speeds.

Microsoft’s commercial RPO was $677 billion at the end of September, up 73%, with “roughly 30%” due in the next 12 months.^5,9 AWS’s backlog was $578 billion, “growing at a triple-digit rate year over year.”¹³ Google Cloud’s was $561 billion, with “just over 55%” due within 24 months.⁴ Oracle’s RPO was $664 billion, but only about 13% is expected to become revenue within 12 months.⁴⁰

Backlog also reveals concentration. About 45% of Microsoft’s commercial RPO came from OpenAI at the end of 2025. At the end of June 2026, RPO excluding OpenAI was up 25%, against 84% in total.^11,14 OpenAI accounts for about half of Oracle’s backlog, according to S&P.¹⁵

The Labs: Fast Growth, Heavy Losses

OpenAI’s annualized revenue was nearing $50 billion at the end of September, according to the FT as reported by Bloomberg. That is below a widely reported figure of about $70 billion, which reflected investors’ adjustments, including how sales through cloud partners are counted.¹⁹ According to the FT, OpenAI projects negative free cash flow of $278 billion over 2026–30 and revenue of about $350 billion in 2030, up from about $36 billion in 2026.⁴⁹

Anthropic’s S-1 shows first-half 2026 revenue of $16.3 billion, up from $1.2 billion a year earlier. It also shows a GAAP net loss of $80.7 billion, of which $72.4 billion was a fair-value charge on convertible notes and other instruments. Adjusted operating income was +$1.6 billion.⁵⁰ Anthropic told investors its third-quarter revenue was $16.8 billion, about $67 billion annualized, on a gross basis that includes sales via cloud partners.²⁰

The Next Web’s coverage of the S-1 lists purchase commitments of at least $114.5 billion to Google, $110 billion to Amazon and $44.9 billion to Microsoft. The filing itself leaves the total blank.⁵¹

Paid Products Are Still Small

Microsoft has “over 30 million paid Microsoft 365 Copilot seats,” worth at most about $10.8 billion a year at the $30-a-month list price, by our calculation.¹¹ Usage is growing faster than revenue. Google’s first-party model APIs process “nearly 32 billion tokens per minute,” about 4.6 times a year earlier.^52,53 One reason is price. According to Epoch AI, inference prices for equal capability have fallen 9–900 times a year depending on the benchmark, so revenue has to chase volume.⁵⁴

Wide Adoption, Mixed Productivity

On Ramp, a business-spending platform whose customers skew toward technology, 55.9% of US businesses paid for an AI product in September, up 12.1 points in a year.⁵⁵ The Census Bureau, by contrast, found that 19.8% of US businesses used AI in any business function in the two weeks to 3 May 2026. Among firms with 250 or more employees, the share was 37%.⁵⁶ Among workers, 40.7% used generative AI for their job in November 2025, but only 12% used it daily.⁵⁷

Productivity evidence ranges from strong to null. In a field study of 5,179 customer-support agents, an AI assistant raised issues resolved per hour by 14% on average and about 34% for novices.⁵⁹

But in METR’s 2025 randomized trial, experienced developers took 19% longer with AI while believing they were faster. Its 2026 update found point estimates of 18% and 4% faster, neither statistically significant, and METR calls the data an unreliable signal.^60,61 Danish registry data show “precise null effects on earnings and recorded hours.”⁶² A widely cited but methodologically disputed MIT-affiliated report claimed only 5% of integrated AI pilots delivered measurable P&L impact.⁵⁸

Ten-year estimates range from Daron Acemoglu’s total-factor-productivity gain of “no more than 0.66%” to Goldman economist Joseph Briggs’s 9% rise in productivity.^63,64

Some of the payoff is inside the spenders themselves. Meta’s revenue grew 26% to $64.6 billion while its headcount fell 15%, although operating income in its apps business fell, partly because of a $9.9 billion legal charge.⁶ Gains like these, in advertising and internal efficiency, never appear as “AI revenue,” which is why the payback math that follows does not capture them.

05 — The MathWhat Has to Be True

Demand is real. The question is whether it is big enough, soon enough, to pay for the assets being bought. By our calculation, the 2026 build alone needs about $275 billion of revenue a year, and the bar rises with each year’s spending.^16,17,18

In 2024 Goldman Sachs’s Jim Covello put the skeptic’s version in one line: “What $1tn problem will AI solve?”⁶⁴

Four Published Yardsticks

Several institutions have sized the revenue AI needs. They differ in scope, and none publishes its full inputs.

• Goldman Sachs estimates AI must generate about $300 billion of revenue a year to cover the spending of the five largest US hyperscalers (Amazon, Alphabet, Microsoft, Oracle, Meta). It also estimates that users would need to spend roughly $1 trillion a year on AI applications for solid returns, against a global software market of about $1.5 trillion in 2026.^16,26
• JPMorgan (November 2025): a 10% return on more than $5 trillion of global AI and data-center investment through 2030 needs about $650 billion of revenue a year in perpetuity. That equals $34.72 a month from every iPhone user.⁶⁵
• Bain (September 2025): $2 trillion of annual revenue is needed by 2030 to fund the compute AI demand requires, and “even with AI-related savings, the world is still $800 billion short.”⁶⁶
• Sequoia’s David Cahn asks “AI’s $X question” each year: $200 billion (2023), $600 billion (2024), $840 billion (2025) and $1.5 trillion (July 2026). Each is the lifetime end-customer revenue needed to justify one year of capex, so it is not comparable with annual figures.⁶⁷

On the other side, Morgan Stanley sees a $1.1 trillion generative-AI revenue opportunity in 2028.¹⁸ The yardsticks agree that the bar is hundreds of billions a year, not tens.

Our Calculation, Step by Step

Because none of those models is public in full, we built a transparent one. It asks how much revenue one year of spending, the 2026 “vintage,” must earn each year to pay for itself. It shows what has to be true, not what will happen.

1. Size: $800 billion, Goldman’s estimate for the five largest US hyperscalers in 2026, close to our $803–828 billion for a different five.^16,26
2. Split: 60% hardware (chips, servers, networking), in line with McKinsey’s 60% and Morgan Stanley’s 55%.^17,18 That is $480 billion of hardware and $320 billion of buildings, power and cooling.
3. Lives: hardware recovered over 5 years, buildings over 20.
4. Return: 10% pre-tax, via a capital-recovery factor of 0.264 for 5 years and 0.117 for 20. The factor is the level annual payment that repays an asset and 10% interest, like a mortgage.
5. Annual capital charge: $480 billion × 0.264 ≈ $126.6 billion, plus $320 billion × 0.117 ≈ $37.6 billion, about $164 billion a year.
6. Operating costs: power, staff, maintenance and networking take an assumed 40% of revenue, so the capital charge must come from the other 60%. Revenue needed = $164 billion ÷ 0.6 ≈ $274 billion a year, or about $275 billion, for five years.
7. Floor: with no return at all, hardware costs $96 billion a year ($480 billion ÷ 5) and buildings $16 billion ($320 billion ÷ 20). That is $112 billion in all, which needs about $187 billion of revenue just to break even.

A defensible range is $250–330 billion a year from the 2026 vintage alone. That brackets Goldman’s roughly $300 billion, although Goldman’s figure is a break-even estimate and its method is not public.^16,26 The 40% cost share is a judgment call. AWS’s operating margin of about 39% after depreciation implies that a mature cloud business has cash costs well under 60% of revenue. Oracle’s early-ramp GPU rental (section 02), however, shows how thin raw rental can be.^2,37

What Moves the Answer

Two assumptions matter most: how long the chips earn money and how much it costs to run them. At 40% costs, a 3-year hardware life (Burry’s case) raises the requirement to $384 billion a year; a 6-year life lowers it to $246 billion. The full grid runs from $211 billion (6 years, 30% costs) to $461 billion (3 years, 50%). If hardware is 70% of the spend rather than 60%, the base case rises to $293 billion.^17,18,34

The Vintages Stack Up

Each year’s spending creates its own five-year requirement, and they pile up, like a household that takes out a new five-year loan each year while still repaying the earlier ones. Our base case needs about $0.34 of annual revenue per dollar of capex (range $0.26–0.58). We apply that ratio to the 2025 vintage of about $414 billion and a 2027 vintage of $1.1–1.2 trillion.

On that basis, by 2028 the assets bought in 2025–27 need about $0.8 trillion a year, within a range of about $0.6–1.4 trillion. That is about $140 billion from 2025, $275 billion from 2026 and $380–410 billion from 2027. This is our calculation; the 2027 vintage uses Goldman’s estimate and the consensus for a group that includes Oracle, so the perimeters are not identical.^{3,6,9,10,11,12,16,26}

The comparison cuts both ways. Hyperscalers receive only the part of lab revenue spent on compute, but they also sell AI to many other customers. Our reading is that visible AI revenue is well short of the bar today but growing fast enough that the gap could close, if that growth holds for several more years.

Who Gets Paid, and When

“Paying off” means different things at different layers.

• Nvidia is already paid, with data-center revenue running at about $356 billion a year.⁴⁸
• The diversified hyperscalers collect the revenue in our model, and their core businesses can absorb a shortfall for a while.⁹
• The specialists, Oracle and the neoclouds, are betting far more relative to their size and depend on outside money.^31,45
• The labs must hit plans such as OpenAI’s, in which revenue rises roughly tenfold from 2026 to 2030.⁴⁹
• Their creditors and lessors carry the losses if the bar is missed.

Thresholds marked as Trimline calculations are ours; other figures are as reported by the sources cited. “Where it stands” is a snapshot, not a forecast.

Three Ways It Could Go

These scenarios are our framing, not forecasts.

Bull: “AWS again.” Revenue at the infrastructure layer passes the roughly $275 billion-a-year bar within the chips’ life. Backlogs convert, GPU rents hold and capex growth slows along a path like Goldman’s, so free cash flow recovers. The spenders and Nvidia win. Watch signals 1, 3, 7 and 8 below for confirmation.

Base: “Railways.” Demand grows strongly, but falling prices and competition cap revenue, and the bar is met late for the 2026 vintage. Diversified hyperscalers earn adequate but lower returns; leveraged specialists such as Oracle and the neoclouds struggle. Users and the wider economy win.

Bear: “Fiber, but faster.” Lab fundraising stalls, concentrated commitments are cut or renegotiated, GPU rents reset and write-downs follow; capex is cut in 2027–28. Losses travel into credit markets through the external financing gap and the circular deals the IMF has flagged.^18,47 Users eventually gain from cheap capacity; equity holders, private credit and leveraged neoclouds lose. It is “faster” because chips, unlike cables in the ground, last only a few years.

06 — The PrecedentsThe Assets Outlast Their First Backers

History cannot say whether AI is a bubble. It does show who tends to win and lose when a build-out outruns its revenue. In railways, electricity and fiber, the assets proved useful in the end, but many of the people who paid for them first lost money.

Railways

British railway investment reached about £44 million in 1847, 7.3% of GDP and half of all UK investment that year.⁷³ By the end of 1850, investors had put about £250 million into railways, “almost half of their GDP.” That was a cumulative total, not one year’s spending, and Andrew Odlyzko estimates they lost about a third of it.⁷³

The forecasts were the problem, not the technology. Railway revenue in 1852 was £15 million, a quarter of the £60 million promoters had expected. By 1905, however, it was £109.4 million, or 6.0% of GDP.⁷³ In the United States, railroad investment averaged about 2.24% of GDP in 1870–90, according to Stijn Van Nieuwerburgh of Stanford.⁷⁴

Electricity

Electrification shows how long benefits can take. Electric motors supplied under 5% of factory mechanical drive in 1899 and just over half by the early 1920s. Paul David found that motor adoption accounts for about half of the 5-point acceleration in US manufacturing total-factor productivity growth in 1919–29, once factories had been redesigned around the new power.⁷⁵

Spending on electric light and power averaged only 0.50% of GDP in 1905–25, but the financing still ended badly for some. Samuel Insull’s utility empire, which served 1.8 million customers in 39 states, went into receivership in April 1932, and its investors lost nearly $800 million.^74,76

Fiber and the Dot-Com Bust

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. US telecom investment peaked at 1.39% of GDP in 2000.^77,78 Demand fell far short: only about 2–5% of US fiber was lit in 2002, according to the FDIC.⁷⁹ Telecom shareholders lost roughly $2 trillion by late 2002, and defaulted high-yield telecom bonds recovered about 16 cents on the dollar in 2001 and 2002.^77,80

Even the suppliers took decades to recover. Cisco’s shares did not close above their March 2000 record until 10 December 2025, more than 25 years later.⁸¹

Where Today’s Boom Fits

By scale, AI spending is large but not unprecedented. By our calculation, guided 2026 capex equals roughly 2.2–2.5% of US GDP, depending on whether SpaceX is included. That is an upper bound, because the spending includes sites abroad, land and imported chips.^78,82 That is about a third of Britain’s 1847 railway peak. Van Nieuwerburgh’s scenario puts the US AI build-out at an average of 3.63% of GDP a year in 2025–32, above railroads, highways, telecom, canals and electrification.⁷⁴

On a broader measure, though, the boom has passed the dot-com peak. US investment in information-processing equipment and software reached a record 5.35% of GDP in the third quarter of 2026, above the 4.46% of the fourth quarter of 2000.⁷⁸ This is broad IT, not AI alone, and an advance estimate that will be revised.⁸² Equipment alone (2.83%) is just below its 2000 high (2.91%).⁷⁸

The Big Difference Is Asset Life

The sharpest contrast with past booms is shelf life. Rail lines, power grids and fiber could sit underused for years and still be there when demand arrived. Most of today’s money goes into chips expected to last about five years. If demand comes late, the hardware may be obsolete first. That is why the payback clock in section 05 matters so much.

There is a bullish precedent, from one of today’s spenders. AWS produced 67% of Amazon’s operating income in 2015 on 7.4% of its sales, and 57% in 2025. Years of heavy capex built Amazon’s main profit engine. The warning is scale. Amazon’s 2025 capex was about 2.9 times AWS’s operating income, and its 2026 plan is about 4.8 times, by our calculation.^2,10,85 And unlike railway promoters or telecom start-ups, today’s biggest spenders have large, profitable core businesses that can carry AI losses for a while.

The economist Carlota Perez, who studies technological revolutions, put the pattern bluntly: “I have not seen a golden age happening without a crash.”⁸⁶

07 — The Physical BillPower, Parts and Pushback

The build also needs electricity, turbines, transformers, memory and local permission. All of these are scarce or getting dearer, and some of the cost lands on people who will never use a data center.

The Biggest US Grid Hits Its Cap

The clearest signal comes from PJM, which runs the biggest US grid. It holds auctions to buy “capacity,” a promise from power plants to be available years ahead. In effect, the grid pays in advance to make sure enough plants will be there when demand peaks.

The price was $28.92 per megawatt-day for 2024/25 and $269.92 for 2025/26. It then cleared at the price cap three auctions running, at $329.17 for 2026/27, $333.44 for 2027/28 and $325.00 for 2028/29. Without the cap, PJM’s simulations put the last two at $529.80 and $554.72.^68,88

The last two auctions also fell short of the reliability requirement, by 6,623 MW and 6,831 MW. On PJM’s “cleared megawatts × price” measure, which is not the cost to consumers, the 2028/29 auction came to $16.4 billion, against $29.7 billion without the cap. For 2027/28, “nearly 5,100 MW” of the 5,250 MW rise in forecast load was attributable to data centers.^68,88

PJM’s independent market monitor estimates that data-center load raised capacity costs by about $23 billion across three auctions. The increase was $9.3 billion for the 2025/26 delivery year, $7.3 billion for 2026/27 and $6.5 billion for 2027/28. About $6.2 billion of the 2027/28 effect comes from data centers “that haven’t been built.”²³

PJM’s Stu Bresler said ratepayers would see little change in their bills from the 2027/28 auction, because the price was flat at the cap.⁸⁹ IEEFA, a clean-energy-oriented think tank, estimates the 2025/26 auction added about $16 a month to average Ohio bills and $18 in western Maryland. Capacity charges are roughly 10–20% of a typical bill.^90,91

The national picture is more mixed. Lawrence Berkeley National Laboratory and the Brattle Group find that through 2025, states with the fastest load growth generally saw real average prices fall; PJM is the exception. Gas, distribution and storm costs drove most 2024–25 increases.⁹² In September, Energy Secretary Chris Wright called data centers “a massive force to reduce the price of electricity.”⁹¹

Turbines and Memory

GE Vernova’s gas-turbine backlog plus slot reservations reached 122 GW, against output of about 20 GW a year. By our calculation, its firm backlog alone, 64 GW, is about three years of production.^69,93 According to TrendForce, conventional DRAM contract prices rose about 93–98% in a single quarter, the first of 2026.⁷⁰ Supply is catching up in places. Long-term agreements are now capping memory price rises, and TrendForce expected server DRAM contract prices to rise a more modest 13–18% in the third quarter.⁹⁴

Pushback

Local opposition is growing. Data Center Watch, a tracker run by research firm 10a Labs, counted $130 billion of projects blocked or delayed in the first quarter of 2026. It counted $68 billion in the second quarter, and the two counts overlap.^71,72 In a Pew survey, 50% of Americans say data centers are mostly bad for home energy costs, up from 38% in January.⁹⁶

Ballotpedia counted 34 local data-center ballot measures on the 3 November ballot in eight states, 20 of them in Ohio. Unofficial early results show at least 14 Ohio municipalities passing bans or restrictions.^25,97

08 — The DebateSame Numbers, Different Clocks

Bulls and bears mostly agree on the numbers. They differ on how fast revenue will arrive, and who absorbs the losses if it doesn’t.

The Bull Case

Bulls point to growth that is accelerating at scale, in the cloud and Nvidia figures in section 04 and in Anthropic’s first-half revenue, up roughly fourteenfold.⁵⁰ “We’re not investing approximately $200 billion in capex in 2026 on a hunch,” Jassy wrote in April.⁹⁸ Satya Nadella said in October, “I’ve never been more confident in the long-term returns for our investments.”⁵

The market has also shrugged off shocks that some expected to end the boom. In January 2025, DeepSeek’s R1 model wiped about $590 billion off Nvidia, the largest one-day loss of market value for any company. Yet when DeepSeek V4 launched in April 2026, Nvidia closed at a record.^99,100

Voices outside tech see substance too. Jamie Dimon said, “You can’t look at AI as a bubble, though some of these things may be in the bubble.”¹⁰¹ Morgan Stanley’s Mike Wilson said that “this is an earnings story, not a multiple expansion one.”¹⁰²

The Bear Case

The bears start with cash. As sections 02 and 03 showed, free cash flow at the four hyperscalers has turned negative and debt at three of them is up $208 billion in nine months.^2,3,6 CoreWeave’s credit default swaps implied about a 60% chance of default in October, by Bloomberg Intelligence’s estimate.⁴⁵ OpenAI projects $278 billion of negative free cash flow over 2026–30.⁴⁹ Evidence of broad productivity gains is thin.

Then there are valuations. The ten largest companies make up about 39–41% of the S&P 500, a level S&P Dow Jones Indices says was “not seen since the mid-1960s.” The Shiller CAPE, which compares prices with ten years of average earnings, was 40.1 on 30 October, second only to its December 1999 peak.^102,103,104

On 11 October, Michael Burry wrote that “almost my entire short position is expressed through SMH and NDX puts,” and “I believe we are within 0-9 months of a significant top.” His positions are self-reported.¹⁰⁵

The Builders’ Own Doubts

The builders concede a great deal. Sam Altman has talked about “random, new neocloud[s]” pledging compute without revenue to pay for it. He described “the first signs of what feels to me like unsustainable silliness,” adding, “I’m not worried about our compute buildout plans. I am worried about the world’s compute buildout plans.”¹⁰⁶

Sundar Pichai said, “I think it’s both rational and there are elements of irrationality in moments like these,” and “I think no company is going to be immune, including us.”¹⁰⁷ Nadella set the condition at Davos in January: “For this not to be a bubble, by definition it requires that the benefits of this are much more evenly spread.”¹⁰⁸

Yet the builders say building too little is the bigger risk. Mark Zuckerberg said in 2025, “If we end up misspending a couple of hundred billion dollars, I think that that is going to be very unfortunate, obviously.” But, he added, “I actually think the risk is higher on the other side.”¹⁰⁹ In our reading, racing is rational for each company, because falling behind could cost more than overspending. Across the industry, the same logic can add up to overbuilding.

Markets and Regulators

Investors have punished some heavy spenders, though not uniformly. Alphabet closed down 12.6% on 29 October, its steepest one-day fall since listing in 2004, according to Investing.com data.¹¹⁰ Meta closed down 7.0% on 22 October, the day after its results. The fall reflected capex, its new non-voting share class and a ratings warning.¹¹¹ Others were barely marked down: Microsoft closed 0.2% lower on 29 October, and Amazon, down nearly 6% after hours, closed only about 1% lower on 30 October.^112,113,114

Regulators flag a correction risk without calling AI a bubble. In July, the Bank of England said, “The risk of a sharp correction in equity markets remains high.” It noted that AI companies are about half of the S&P 500, up from about a quarter in 2022.²⁹

In the Fed’s May survey, 50% of market contacts cited AI as a salient financial-stability risk, up from 30% in autumn 2025.¹¹⁵ Separately, in the IMF’s general adverse stress test, banks holding 22% of global banking assets would fall below regulatory capital limits, up from 18% a year earlier.⁴⁷

Signals Worth Tracking

1. Nvidia’s results on 17 November, against guidance of $108 billion (±2%) of total revenue.^24,48
2. Anthropic’s IPO, reportedly targeted for the week of 16 November: pricing, trading and whether the amended S-1 discloses total commitments.^50,51
3. Fourth-quarter calls (late January to early February 2027) and the first 2027 capex numbers, including Meta’s promised “quantitative” outlook.^4,13,28
4. Microsoft’s acceleration: whether capex tops $50 billion in the October–December quarter and free cash flow stays positive.⁵
5. Lease-financed capacity: SpaceX’s Apollo GPU lease from 2027, Microsoft’s operating leases, and Meta’s and Oracle’s leases as they start.^22,40,41
6. Credit stress: Oracle and CoreWeave credit spreads, CoreWeave’s 2032 bonds and order-book coverage for hyperscaler bonds.^39,44,45
7. Backlog quality: Microsoft’s RPO and OpenAI’s share of it; how much of Oracle’s backlog converts within 12 months.^5,40
8. GPU rents: the Silicon Data H100 index and how newer chips are priced.³⁶
9. Adoption: the Census Bureau’s bi-weekly AI-use rate and Ramp’s monthly AI Index.^55,56
10. Power and politics: final results of the 3 November measures, PJM’s next auction and the future of its price cap.^68,97