The payback on AI infrastructure: about $740bn of capital expenditure a year, a token of the same quality 280 times cheaper, and whether multiples will compress to 10–12

02 · Timeline2017–2026: how the hyperscalers went from servers to gigawatts

Until 2022 the capital expenditure of the big cloud companies grew together with their business: more customers, more servers. After the launch of ChatGPT the logic changed. Companies began to build for demand that is not yet covered by end-user revenue, and to count capacity not in servers but in gigawatts. The timeline below is assembled from the same reports and sources as the charts; the filter at the top shows construction, token prices, revenue, debt and stock-market valuations separately.

03 · TokensA token of fixed quality gets 10–20 times cheaper a year, while consumption grows at a similar pace

A token is a fragment of text, the unit in which a model measures input and output; in English it is roughly three-quarters of a word. Prices are usually quoted per million tokens. The Stanford AI Index 2025 recorded a more than 280-fold fall in the cost of using a GPT-3.5-level system between November 2022 and October 2024 3 H. According to Epoch AI’s table, a model scoring at least 64.8 on the MMLU test (GPT-3.5 level) cost $20 per million tokens in November 2022 and $0.07 in October 2024 (Gemini-1.5-Flash-8B) 4 M. For the GPT-4 level, the lowest price fell over 23 months from $37.50 to $0.18, about 208-fold 4 M. Epoch AI calculates that the price of reaching a fixed level of quality falls by 9 to 900 times a year, depending on the test; for the GPT-4 level on graduate-level science questions, by about 40 times a year. The fastest fall came in the past year; whether it will last is unknown 4 M. The venture capital firm a16z called this “LLMflation”: a tenfold fall in cost every year 41 M. From the end points of Epoch’s table ($20 and $0.07 over 1.8 years), the rate works out at about 22 times a year (our calculation) L.

But this figure describes a fixed level of quality, not what people actually buy. It compares GPT-3.5 in 2022 with a small model from another company in 2024 that scores at least as well on the old test. Today customers pay for the strongest models, and those prices behave differently. GPT-5 was launched in August 2025 at $1.25 per million input tokens and $10 per million output tokens, and GPT-5.5 in April 2026 at $5 and $30; GPT-5.6 Sol, according to OpenAI’s changelog, launched in July at the same price, and from 21 August a promotional price of $4 and $20 has applied until at least 21 November 42 H. Anthropic cut the price of Opus to a third in November 2025, to $5 and $25, and in June 2026 launched a still more expensive tier, Fable 5, at $10 and $50; Opus 5.5 has cost $4 and $20 since September 43 H. OpenAI’s Pro versions cost 6–12 times more than the standard ones 42 H. In addition, from version 4.7 the new Claude models split the same text into about 30% more tokens, so the price per token understates the rise in the price per page 43 H. According to the BenchLM index, the prices of flagship models rose by 61% in the 12 months to October 2026, while those of mid-tier models fell by 45% 44 L.

The market price actually paid for a mix of models has been falling since May 2026, but before that it was rising. The Silicon Data index, weighted by usage volumes, was launched in December 2025; as reported by Bloomberg, it almost doubled by May, peaking at about $2 (our calculation) 45 L. On 3 July 2026 the index stood at $1.62 per million tokens, and on 25 October at $0.95 45,46 M. It depends on the pattern of usage, not only on price lists. According to press reports, at the turn of August and September it fell below $1 for the first time, because of cheap open Chinese models and OpenAI discounts 47 L. Meanwhile, as reported by Bloomberg in July 2026, optimists estimate that total spending on tokens roughly doubled over the year, although list prices have fallen by more than 90% since 2023 48 L. This is the Jevons paradox: when a resource gets cheaper, its consumption grows so fast that total spending rises. Satya Nadella, chief executive of Microsoft, wrote about it on 27 January 2025, the day Nvidia’s shares plunged because of China’s DeepSeek 49 M.

How many tokens are processed

Google provides the longest open series of volumes. In May 2024 the company processed 9.7trn tokens a month across all its products and its API, in May 2025 480trn, in October 2025 1.3 quadrillion, and in May 2026 3.2 quadrillion 8,50,9 H. That is roughly 330-fold in two years — the same order as the fall in the price of a GPT-3.5-level answer. Through its API, Google’s models processed 22bn tokens a minute in the second quarter of 2026, against 16bn a quarter earlier 51 H. In January–March 2025 Microsoft processed more than 100trn tokens, five times as many as a year earlier 52 H.

It is not only the number of queries that is growing but also the number of tokens per task. Reasoning models first “think” through hundreds or thousands of hidden tokens and only then answer. According to a study by the OpenRouter platform covering 100trn tokens, the share of reasoning models rose from almost zero to more than half over 2025, and the average request became about four times longer 53 M. According to Anthropic’s measurements in its own research system, an agent uses roughly four times as many tokens as a chat conversation, and a multi-agent system 15 times as many 54 H.

An important detail: consumption seems to be growing not so much because tokens have got cheaper as because models have learnt to do new things (our interpretation). In the same OpenRouter study, a 10% fall in a model’s price corresponded to a rise in usage of only 0.5–0.7% (a comparison between models on one platform, not a change over time; the platform is geared to independent developers) 53 M. Menlo Ventures, an investor in Anthropic, writes that on hard tasks customers are “fairly price-insensitive” and move to the best new model even when older ones get ten times cheaper 55 M. So for the excess-capacity hypothesis, what matters is not the price of a token but whether new uses keep appearing. So far they appear every quarter.

Why a token gets cheaper faster than hardware

Computation on the same hardware gets cheaper slowly. According to Epoch AI (data for 2010–2023, FP32 computation, before Blackwell), the number of operations per dollar on GPUs for AI doubled roughly every 2.1 years, that is, grew by ~39% a year 56 H. The AI Index puts the fall in hardware costs at 30% a year and the rise in energy efficiency at 40% 3 H. The rest of the fall in price comes from smaller models, better algorithms and competition: according to an estimate in a preprint by MIT researchers, algorithms alone give roughly a threefold cut in cost a year 57 M. Alphabet said that over 2025 it cut the unit cost of serving Gemini by 78% (the company did not specify the unit) 58 M. Nvidia promises that its new Rubin platform will cut the cost of a token by up to ten times compared with Blackwell for mixture-of-experts models (according to one independent analysis, by two to three times on other workloads); this is the manufacturer’s claim 59 L.

Hence the main arithmetic. If the number of tokens grows sevenfold a year and each token needs a third of the computation, the need for computation grows 2.3 times a year. Meanwhile hardware gets cheaper by only 1.4 times, so spending on it grows by roughly 1.7 times a year (our calculation) L. The calculator below lets you change these assumptions.

What data centres say

If cheaper tokens were reducing the need for capacity, it would show up in the leasing market. So far the opposite is visible. In the first half of 2026 vacancy in data centres in the main North American markets was 1.4% — a record low, though it was the same at the end of 2025; 7,481MW was under construction, and more than 80% had already been leased in advance 10 H. In July 2026 Microsoft said that customer demand continues to exceed available capacity 24 H, and Alphabet that the company is short of capacity and will rent it from third-party suppliers 60 M. The price of a one-year rental of an H100 GPU rose from $1.70 to $2.35 an hour between October 2025 and March 2026 61 M, and the Silicon Data index for neoclouds from $2.00 to $2.89 an hour between December 2025 and October 2026 (the FAQ on the same page gives $2.53) 62 M.

Signs of a glut have also appeared, but they were local. In March 2025 analysts at TD Cowen reported that Microsoft had walked away from about 2GW of capacity that was in the process of being leased 63 M. Yet in fiscal 2026 (to 30 June) Microsoft’s new finance leases came to $24.6bn, against $20.5bn a year earlier, and leases signed but not yet commenced, mostly for data centres, to $329.1bn 64 H. In China, according to local media reports, up to 80% of new computing capacity stood idle after a wave of more than 500 announced projects; among the reasons given was speculative building by local authorities 65 L.

The market came closest to testing the hypothesis on 27 January 2025. China’s DeepSeek said it had trained a strong model very cheaply, and Nvidia’s shares fell by 17% — the company lost about $590bn of market capitalisation, a one-day record 49,66 M. Analysts at SemiAnalysis quickly worked out that ~$6m was only the cost of the final stage of training, while DeepSeek’s servers had cost about $1.6bn 67 M. Within three weeks Nvidia’s shares had almost returned to their earlier level 68 M. And quarterly capital expenditure by the five biggest companies rose from $76bn in the fourth quarter of 2024 to $181.5bn in the second quarter of 2026 (Alphabet, Amazon, Meta, Microsoft and Oracle; Oracle’s quarter ends in May) 5 M. Building did not slow afterwards but accelerated; these data on their own do not prove a causal link with the cheapening of models.

04 · RevenueWho pays: the revenue of the two leading labs is about 10% of building costs

Hyperscalers do the building, but those who pay for computing are mainly the developers of models and their customers. The two biggest labs, OpenAI and Anthropic, already earn tens of billions of dollars a year, but they report differently, and headlines often mix up different things. Annualised revenue run-rate (run-rate) is revenue over a short recent period, scaled up to a year. Recognised revenue is what has reached the financial statements. Anthropic, for example, calculates its run-rate as consumption over the previous 28 days multiplied by 13, plus monthly subscriptions multiplied by 12 11 H. In addition, Anthropic counts sales through the AWS, Google and Microsoft marketplaces in full, on a gross basis 11 H, whereas OpenAI, according to the FT, counts only its own share 69 M. As a result, on 8 October 2026 OpenAI’s run-rate “shrank” in the press from almost $70bn to ~$50bn: all that changed was how sales through cloud partners are counted 69 M.

OpenAI: the biggest audience, the biggest losses

OpenAI does not publish public accounts. According to audited accounts obtained by Ed Zitron and checked by the Financial Times, the company booked revenue of $13.07bn in 2025 ($3.7bn in 2024), and its operating loss was $20.9bn — 160% of revenue; of $19.2bn of research costs, $10.6bn went to Microsoft 70 M. In the first half of 2026 revenue, according to the WSJ as relayed by one aggregator, was $12.4bn 12 L, and the operating loss, including share-based payments, was reportedly about $21.6bn 12 L. In February 2026 ChatGPT had 900m weekly users and 50m paying subscribers 71 M. According to an investor presentation reported in September by the FT and Reuters, OpenAI plans revenue of $36bn in 2026 and $350bn in 2030, and over 2026–2030 expects to spend about $856bn on computing and to have a cumulative negative free cash flow of $278bn 72 L.

Anthropic: the first lab with public accounts

On 26 October 2026 Anthropic published a prospectus for an initial public offering (Form S-1) — the first public financial statements of a frontier lab (the annual figures for 2025 are audited; those for the first half of 2026 are not). Revenue in 2025 was $4.59bn, against $386m in 2024; in the first half of 2026 it was $16.3bn, of which $11.6bn came in the second quarter 11 H. Spending on computing and infrastructure was 251% of revenue in the first quarter of 2025 and 48% in the second quarter of 2026: by the company’s explanation, revenue overtook a computing base bought 12–18 months in advance 11 M. According to the prospectus, in the second quarter of 2026 the company made an operating profit of $974m for the first time in the six-quarter series presented (from the first quarter of 2025); we have not checked this figure against the full text of the prospectus 11 M. About 90% of the net loss for the first half ($80.7bn) consists of a non-cash revaluation of convertible bonds, derivative liabilities and SAFE agreements ($72.4bn) 11 H.

Here is some evidence for the cheap-token hypothesis. In November 2025 Anthropic released Opus 4.5, a more powerful model at a third of the price of the previous Opus, and the quarterly growth of its revenue accelerated from 48% in the fourth quarter of 2025 to 132% in the first quarter of 2026 11 H. The prospectus mixes the effects of price, capabilities, new products and customer acquisition, so the contribution of the price cut alone cannot be isolated; but the cut clearly did not reduce revenue. For comparison: according to the prospectus, Anthropic’s revenue from customers on the Fortune 500 list averages less than 0.1% of their total spending, and for its ten biggest customers less than 1% 11 H; the company presents this as room for growth. According to anonymous sources cited by Bloomberg, at the end of July Anthropic’s annualised revenue run-rate reached $65bn (the latest official figure is $47bn in May; recognised revenue for the second quarter, annualised, is about $46bn, our calculation) 73 M.

How this compares with building costs

A simple comparison gives a ratio of about 13 to 1: ~$55bn of combined annualised run-rate of the two labs in spring 2026 (OpenAI ~$25bn at the end of February, Anthropic over $30bn at the start of April) against $725–760bn of planned capital expenditure 69,74 L. But the bases and dates are mixed here. If recognised revenue is compared with the actual capital expenditure of the big four over the same period, the share looks like this: 1.8% in 2024, 4.7% in 2025 and 9.7% in the first half of 2026 (our calculation, using cash capital expenditure excluding finance leases; Anthropic’s revenue is gross, OpenAI’s is net of the cloud partners’ share) L. If one takes the highest annualised run-rates (Anthropic $65bn, OpenAI ~$50bn) against the plan of $725–760bn, the result is about 1 to 6.5 L; the bases differ: Anthropic’s run-rate includes sales through cloud partners, OpenAI’s does not, and the dates differ (end of July and end of September). The share is growing fast, but even on the most generous count the labs’ revenue is a minority of spending.

The labs are not the only buyers. Microsoft said that its AI business had an annualised revenue run-rate of $37bn in March 2026 (+123% on the year), and that Azure exceeded $100bn for the first time in fiscal 2026 75,76 H. In the second quarter of 2026 Google Cloud’s revenue was $24.8bn (+82%, including the acquired Wiz), with an operating margin of 35.6% 2 H. AWS made $42.2bn (+37%) with a margin of 39.4%, but in the same quarter added $48.6bn of equipment — more than its entire quarterly revenue 30 H. Meta has no separate AI revenue: third-quarter revenue rose by 26%, ad impressions increased by 10% and the average price per ad by 14% 7 H.

The $2.5trn backlog and who stands behind it

The strongest argument of the optimists is signed contracts. Customer commitments to Microsoft as of June 2026 are $678bn, to Oracle as of August $664bn, Google Cloud’s backlog is $514bn, AWS’s $496bn and CoreWeave’s $104bn 64,77,2,30,78 H. In all, about $2.5trn — roughly 4.3 times the capital expenditure of Alphabet, Amazon, Meta, Microsoft and Oracle over the past 12 months ($566bn) and three times the big four’s plan for 2026 (our calculation). The figures are not uniform, and part of the commitments of OpenAI and Anthropic may be counted in the backlogs of several companies at once 32 L.

But this money is concentrated among a few customers and will arrive over years. By S&P’s estimate, about half of Oracle’s backlog is contracts with OpenAI 79 M, and only 13% of Oracle’s entire backlog falls in the next 12 months 77 H. Part of the suppliers’ revenue is money they have themselves put into customers. In the second quarter of 2026 Amazon booked a $50.5bn gain from revaluing its stake in Anthropic, and Anthropic committed to spend more than $100bn at AWS over ten years 30 H. Microsoft received $24.1bn of revenue from OpenAI in fiscal 2026 and at the same time owns about 25% of OpenAI 64 H. At Nvidia one direct customer (which may be a server maker or a distributor) accounted for 16% of revenue in the quarter to 26 July 2026 80 H.

How much revenue is needed

Estimates of the revenue needed move in step with capital expenditure: they are models that derive the required revenue from building costs. David Cahn of the venture firm Sequoia calculated in September 2023 that AI needed ~$200bn of revenue to pay back the building then under way; in June 2024 the figure was $600bn 81,82 M, and in July 2026 about $3trn 83 L. The consultancy Bain wrote in 2025 of $2trn of new annual revenue by 2030, and in 2026 of $6trn of total annual AI revenue by 2031, of which existing applications could supply $1.2–1.8trn 84,85 M. JPMorgan estimated that a 10% return on investment requires ~$650bn of revenue a year — about $35 a month from every iPhone owner 86 M. Goldman Sachs, as relayed by the press, regards ~$300bn a year as the break-even threshold for hyperscalers 87 L. For comparison: Epoch AI, drawing on an estimate from Exponential View, put all generative-AI revenue in June 2026 at ~$175bn a year 88 L.

Whether AI pays off for customers is still unknown. The 2025 MIT NANDA report, which drew on a review of more than 300 public initiatives, 52 interviews and a survey of 153 executives, claimed that 95% of organisations had received no measurable return from generative-AI pilot projects 89 M; critics note that interviews were conducted at only 52 organisations, and that every pilot that did not reach full deployment is counted as a failure 90 L. In McKinsey’s 2026 survey (from a summary in search results; we did not open the full text), 37% of respondents reported some impact of AI on operating profit, and only ~6% an impact of 5% or more 91 L. Meanwhile the share of American firms using AI in their work rose, according to the US Census Bureau, from 17.2% at the end of 2025 to 24.6% in October 2026 92 H.

05 · DepreciationHow spending turns into profit: servers are written off over 5–6 years, but the money has already been spent

Capital expenditure does not reduce profit at once. A company pays for a server today, but in its income statement it writes off the cost in instalments over the server's useful life — this is depreciation. While the building boom continues, depreciation lags behind spending. In fiscal 2026 Microsoft spent $115.9bn on equipment, while depreciation was $34.3bn; two years earlier it was $15.2bn 32,64 H. In the first half of 2026 Alphabet spent 5.9 times more on construction than it wrote off 2 H. In other words, today's profits — and today's low P/E ratios — do not yet carry the main weight of the spending already done.

A useful life that has been extended six years running

How many years a server counts as useful is decided by the company itself, and since 2020 all the big hyperscalers have lengthened this period. Microsoft raised it from 3 to 4 years in fiscal 2021 and to 6 years in 2023; the first decision added $2.7bn to operating profit, the second $3.7bn 93,94 H. Alphabet moved to 6 years in 2023, which cut depreciation by $3.9bn 95 H. Meta has used 5.5 years since 2025 (−$2.92bn of depreciation) 96 H, Oracle — 6 years from fiscal 2025 97 H, CoreWeave — 6 years 98 H. Amazon is the only one to have moved some servers back from 6 to 5 years: from 2025, “because of the pace of development in AI”, which added $1.4bn of depreciation 99 H. From fiscal 2027 Microsoft will also depreciate data-centre buildings over 25 years rather than 15 24 H.

In November 2025 the investor Michael Burry claimed that, because of the longer lives, hyperscalers would understate depreciation in 2026–2028 by about $176bn, and that Oracle's and Meta's profits would be overstated by 27% and 21% by 2028 100 M. He did not publish his calculation 101 L. Footnote Brief, using its own model, puts the understatement of depreciation for 2026–2028 anywhere from $80–100bn (if all servers last 5 years) to $280–320bn (if AI hardware lasts 3 years); this is the author's assumption, not company data 102 L. Nvidia replied to analysts that its customers depreciate its chips over 4–6 years, and the company's chief financial officer said that A100s, released six years ago, are still fully utilised 103,104 L. Microsoft's chief financial officer Amy Hood said that many of the company's GPUs are already contracted for most or all of their useful life 105 L. The problem lies elsewhere: Nvidia now releases a new generation every year — Blackwell in 2024–2025, Vera Rubin in 2026, Rubin Ultra in 2027, Feynman in 2028 106 M. Critics believe that each new generation devalues the previous one faster than happened with servers for conventional cloud services. The rental market does not show this yet: six-year-old A100s are still rented out, and the price of the H100 has risen since December 2025 107,62 M.

What an hour of GPU time costs

The best market test of useful life is the rental price. The list price of an H100 at the hyperscalers stood at $7–9 an hour in 2023–2024 and fell to about $6.3 in the second half of 2025 108 M; on 1 June 2025 AWS cut the H100 price by 44% 109 H. In the Silicon Data index for neoclouds the price fell to $2.00 (9 December 2025) but has been rising again since: on 26 October 2026 the index stands at $2.89, 44% higher (in the FAQ on the same page it is $2.53); the Ornn index of actual deals is $3.03 62,110 M. In a Silicon Data blog post the median for the neocloud segment in the second half of 2025 was about $3.3, and for marketplaces about $1.95: these are series with different definitions 108 M. Renting an A100, released in 2020, still costs $1.3–1.6 an hour (the Silicon Data index page shows two values: $1.56 and $1.29) 107 M. According to a 2024 model by Latent Space analysts, an H100 earns more than 20% a year at $4.50 an hour, just over 10% at $2.85 and makes a loss below $1.65 111 L.

Hyperscalers' margins on AI are lower than on conventional cloud services. Microsoft's cloud gross margin fell from 71% in fiscal 2024 to 66% in 2026, and the company explicitly attributes this to AI infrastructure 64 H. According to documents obtained by The Information, in the quarter to August 2025 Oracle earned a gross margin of only 14% on renting out Nvidia servers 112 M; at a meeting with analysts in October 2025 Oracle said that management's target is a 30–40% gross margin over the full term of the contracts, including the initial loss-making phase 113 L. CoreWeave, which only rents out GPUs, had an adjusted EBITDA margin of 57% in the first half of 2026 but, after depreciation ($2.5bn) and interest ($1.2bn), a net loss of $1.4bn 114 H.

Where a dollar of capex goes

Epoch AI modelled what a 1GW AI data centre in the US is made of: about $37.9bn of one-off costs, of which 56% is servers, 30% the building, cooling and electrical systems, and 13% networking 115 M. Operation costs about $0.9bn a year, mostly electricity 115 M. If servers last 5 years and the building 14, the full annual cost of one gigawatt, including a return on capital, is $8.5bn; if servers last 3 years, it is $12–13bn 115 M. Nvidia gives a higher figure — $50–60bn per gigawatt (a statement of August 2025; Bernstein estimates about $35bn). In October 2026 Jensen Huang said that Nvidia's own content per gigawatt already exceeds $40bn. These are claims by a supplier, not reported figures 116,117 L.

This structure is the key to the question of whether hardware will become a cheap commodity after the boom. Buildings, grid connections and fibre will last for decades, as railway tracks once did. Servers will not: they will have to be replaced every few years, and every replacement is new capital expenditure.

06 · FinancingFrom their own pockets to debt: bonds, leases, private credit and circular deals

As long as capital expenditure was smaller than operating cash flow, hyperscalers built with their own money and still bought back shares. In 2023 the operating cash flow of the five companies covered capex 2.4 times, in 2025 1.5 times (in the first quarter of 2025 also 1.5 times), and in the first quarter of 2026 only 1.06 times (our calculation from Epoch AI data) 5 L. The big four's free cash flow peaked in 2024 ($230bn) and fell in 2025 to $204bn, although operating cash flow rose by 27% 22 M. In 2026 negative free cash flow appeared at Alphabet (−$5.9bn in the second quarter) and Meta (−$2.1bn in the third quarter), while at Amazon it was negative over the 12 months to 30 June (−$7.6bn by the company's definition) 1,31,30 H.

Bonds, shares and leases

The difference began to be borrowed. In 2025 Amazon, Alphabet, Meta, Microsoft and Oracle together sold $121bn of US dollar bonds, against an average of $28bn a year in 2020–2024 118 M. The biggest deals: Oracle $18bn in September, Meta $30bn in October, Alphabet $17.5bn and Amazon $15bn in November 118 M. In 2026, up to 3 September, the same companies together with SpaceX had already borrowed more than $180bn 119 M; in May Meta sold $25bn of bonds, and its debt rose to $84bn 7 H. They borrow for the long term: 41% of these companies' bonds issued in 2025–2026 have a maturity of more than 15 years, and Alphabet sold sterling bonds maturing in 2126 119 M. The equipment this money pays for is written off over 5–6 years.

Besides debt, companies sell shares. In June 2026 Alphabet raised $49.6bn through new shares and mandatory convertible preferred stock 2 H. In a single quarter Oracle sold $19.9bn of shares 77 H. And the largest part of the obligations is not visible in capital expenditure at all: these are leases for data centres that have been signed but have not yet commenced. According to the latest reports (as of dates from 30 June to 30 September 2026) they add up to $1.23trn 7 L.

Part of the construction is moved off the balance sheet through special-purpose vehicles (SPVs). Meta's Hyperion campus in Louisiana is financed by $27.3bn of bonds issued by a joint venture controlled by the Blue Owl fund. Meta leases the campus in four-year blocks and guarantees the owners the residual value over 16 years if it does not renew the lease 120 M. At 30 September 2026 Meta's maximum possible loss under such structures was $68.4bn, of which $41.5bn is residual value guarantees 7 H. In August 2026 Nvidia guaranteed up to $105bn of lease and power payments, under 20-year contracts, for a 4.25GW campus in Ohio belonging to an OpenAI partner 80 H.

Where debt is already expensive

The risk is spread unevenly. Highly rated hyperscalers borrow cheaply. Oracle, which is building for OpenAI's contracts, no longer does. On 9 July 2026 S&P cut the company's rating to BBB-, one notch above “junk”, and forecasts a free cash flow deficit of almost $42bn for fiscal 2027 79 M. The price of five-year credit default swap (CDS) protection on Oracle reached 288 basis points on 15 October 121 M. CoreWeave, the largest independent renter of GPUs, had $35.1bn of debt at 30 June and paid $640m of interest in the quarter, against $267m a year earlier 78 H. Its bonds due 2032 traded in October at about 78 cents on the dollar with a yield of more than 16%, and the price of CDS protection exceeded 1,000 basis points 121 M.

Banks expect debt to become the main source. In 2025 Morgan Stanley put the data-centre “financing gap” to 2028 at about $1.5trn — money that companies' own funds will not cover 122 M. In mid-2026 JPMorgan raised its estimate of AI capital expenditure to 2030 from $5.1trn to $5.5trn and expects debt to finance $4.1trn of it 123 M.

Circular deals

A separate risk is deals in which a supplier finances a customer and the customer buys from the supplier. According to Trending Topics, the IMF's Global Financial Stability Report (October 2026) paints the following picture: in 14 public deals since 2023, hyperscalers have put about $300bn of equity into AI companies and extended about $650bn of debt, while the AI companies have committed to spend $1.4trn with them; we have not checked the report itself. The IMF calls such interdependence “wrong-way risk”: the investment loses value just when the revenue linked to it dries up 124 M. The best-known headlines, meanwhile, overstate the real money: “up to $100bn” from Nvidia for OpenAI was only a letter of intent, and in the March 2026 round Nvidia put in $30bn 125,126 M; of Amazon's $50bn, as of 30 June 2026 it had invested $28.7bn, and a further $21.3bn remained a commitment 30 H; according to Bloomberg Law, $35bn of the $50bn was conditional (an OpenAI stock-market listing or a technical milestone), but in Amazon's 10-Q part of this tranche is already shown as invested 127 M.

History has seen this before. In 2000 Lucent, Nortel and Cisco lent $25–30bn to telecoms start-ups so that they would buy their equipment. Within a year the share of problem loans rose from 2.6% to 60% at Lucent and from 25.5% to 80% at Nortel 128 M.

The first stresses have already occurred. On 28 April 2026, after reports that OpenAI had missed its internal targets for user and sales growth, CoreWeave's bonds fell to a record low and Oracle's CDS protection became more expensive 129 M. In October 2026 the IMF named a correction in AI as one of the likely near-term triggers of a sharp deterioration in financial conditions worldwide 124 M. The Bank of England had already warned of the risk of a sharp correction in October 2025 130 M.

A side effect: defence AI

For Ukraine this topic matters indirectly: the money flowing into AI also flows into defence technology. Venture funding of defence start-ups worldwide held at $2.8–3.8bn a year in 2022–2024, reached a record $9.6bn in 2025 and, in 2026 up to 2 June, exceeded $14.6bn 131 M. Most of it went to a few companies: in May Anduril raised $5bn at a valuation of $61bn, twice that of a year earlier 132 M; Shield AI and Saronic raised a further $3.75bn 131 M; in July Helsing raised $1.8bn, the largest round in the history of European defence 133 M. Meanwhile, in the first half of 2026 AI companies received 86% of all venture money in the US 134 M.

Ukrainian defence start-ups are small by comparison: according to Brave1, disclosed investment in them in 2025 exceeded $105m — about 1% of the global total 135 M. The most notable event was Swarmer's listing on Nasdaq in March 2026: the shares rose by almost 520% on the first day, and market capitalisation exceeded $650m 136 M. Such listings depend on the market's appetite for AI. If valuations of AI companies fall, the first to suffer will be the late rounds of large defence companies valued in the tens of billions, and the window for new listings. Ukrainian manufacturers that live off state orders and grants will feel it indirectly — through dearer and scarcer private capital (our estimate) L.

07 · MultiplesWill P/E ratios fall to 10–12: at Nvidia compression has already happened through earnings growth, at Oracle through a falling share price

P/E is the share price divided by earnings per share. Forward P/E uses the earnings forecast for the next year (for indices and sectors, the next 12 months); trailing P/E uses actual earnings for the past 12 months. The hypothesis’s starting point, “50–70”, does not match any large AI company as of October 2026: Palantir trades even higher, at 85. According to Finviz on the morning of 26 October, forward P/E was 23.5 for Amazon, 22.4 for Microsoft, 20.9 for Alphabet, 18.9 for Meta, 18.3 for TSMC, 16.9 for Broadcom, 13.6 for Nvidia and 11.5 for Oracle 13 M. Another service, StockAnalysis, gives 17.8 for Nvidia, probably because of a different forecast horizon 14 M. AMD (32), Intel (about 40) and Palantir (85) trade noticeably higher 13,17 M. For the S&P 500 as a whole, forward P/E is 19.0: almost the same as the 10-year average (19.1) and slightly below the five-year average (19.8); for the information technology sector it is 20.4, below its own five-year average (25.5) 15 H.

Trailing P/Es are misleading at the moment, too. In the second quarter of 2026 Alphabet booked $98bn of profit from revaluing its stakes in other companies, and Amazon $53.4bn of other income, mostly from its investments in Anthropic (section 04 gives a narrower measure: $50.5bn from revaluing its stake, according to the 10-Q). This money does not come from the core business 15 H. Because of it, Alphabet’s trailing P/E (16.1) is lower than its forward P/E (20.9) 13 M.

Nor do the “conservative industries” trade at 10–12. The utilities sector in the S&P 500 has a forward P/E of 16.3, the railways Union Pacific and CSX 20.5 and 21.9, Coca-Cola and Procter & Gamble 26.1 and 21.6, JPMorgan 13.1, Exxon 14.8 15,17 M. Only telecoms companies fall below 10: Verizon 8.3, AT&T 8.1, Comcast 6.3 17 M. So a P/E of 10–12 is not the level of “boring industries” in general: below that range trade only highly indebted, slow-growing network industries (telecoms, 6.3–8.3), and within the range, among the companies named, there is today only Oracle.

How Nvidia compressed its P/E while its value rose tenfold

The best example of compression having already happened is Nvidia. From January 2023 to October 2026 its market capitalisation grew about 10.4 times (from $501bn to $5.2trn), while its trailing P/E fell from 114.7 to 27.2, because earnings over 12 months rose about 44 times. In January 2023 the P/E was inflated by abnormally low earnings (≈$4.4bn); from January 2022 it fell from 58.6, while earnings rose about 20 times 16 L. In the second quarter of fiscal 2027 (to 26 July 2026) Nvidia’s revenue was $96.2bn (+106% year on year), and guidance for the third quarter is $108bn 137 H. But Nvidia’s revenue depends to a large extent on its customers’ capital expenditure. If that spending stops growing, Nvidia’s earnings growth will stop, and its multiple will have to be re-rated (our estimate) L. Nvidia’s P/E of 27.2 is on a GAAP basis, and it, too, contains non-operating gains (our note) L.

When P/Es really did fall to 10

History knows compression of this size, and it happened to companies whose earnings were growing. In September 1999 Morgan Stanley valued Microsoft at 66 times expected annual earnings 138 H. At the end of 2012 Microsoft’s trailing P/E was 11.7, and in 2013 it was 11.4 139 M. According to Boston Partners (earnings basis not stated), the company’s P/E fell from 53 to 9 in 2000–2011, although its revenue more than tripled and its operating profit more than doubled; the shares regained their 1999 record (nominal, excluding dividends) only in October 2016, after 17 years 140,141 M. Cisco’s P/E at the end of fiscal 2000 was 166, and at the end of fiscal 2010 it was 17 (both on earnings for the past 12 months) 142 M; Cisco’s shares exceeded their March 2000 record only on 10 December 2025, after 25.7 years 143 M. The economist Jeremy Siegel found that shares that traded for years at more than 50 times annual earnings tended to lag the market over the next 25 years 144 M.

What made these companies’ P/Es low? Not cheaper products, but slower growth and a new role: from “the future” they became “infrastructure”. Regulated industries show another pattern: the role of debt. In June 1947 the least indebted US electric utilities traded at 16.4 times annual earnings, the middling ones at 13.1, and the most indebted at 9.0 145 H. Analysts at LPL Research in July 2026 put the question in just this way: are hyperscalers still software platforms or already infrastructure assets 146 M?

A live example of compression already exists, and it is Oracle. The company’s market capitalisation peaked at about $900bn in September 2025 and on 26 October 2026 stood at $383bn (about −57%, our calculation based on market capitalisation; the fall in the share price is larger); in the week to 26 June the shares fell by about 19%, the worst week since 2001 21 M. Its forward P/E is 11.5 13 M. CoreWeave’s shares trade about 70% below their peak of about $187 (our calculation based on data from The Daily Upside and Finviz; the peak may be intraday) 13 L. The market already distinguishes those who build from their own cash flow from those who build on debt.

What it would take for P/E to fall to 12

The arithmetic is simple. If Microsoft’s forward P/E falls from 22.4 to 12 and earnings do not change, the shares will lose 46% of their value. For the price not to fall, earnings must grow by about 13% a year for five years in a row (our calculation) L. Analysts so far expect more: earnings of S&P 500 companies are due to grow by 32.8% in 2026 and by 16% in 2027, but the pace is slowing: +28.6% in the fourth quarter of 2026, +19.3% in the first quarter of 2027 and only +2.8% in the second quarter of 2027 15 H. Moreover, this growth is concentrated: excluding Nvidia and Meta, earnings of the Magnificent Seven will grow by only 7.5% in the third quarter of 2026 15 H, while 42 AI companies have accounted for 67% of the S&P 500’s price rise and 58% of its earnings growth since January 2024 147 M. The seven largest companies made up 34.3% of the index’s market capitalisation at the end of 2025 148 M. The long-term gauge of market valuation, the Shiller CAPE, which is the price divided by average inflation-adjusted earnings over 10 years, stood on 23 October 2026 at 40.4, against an all-time high of 44.2 in December 1999 149 M.

Investors are nervous but not selling. In Bank of America’s survey of fund managers in October 2026, the net share of those holding more technology shares than usual reached 37%, the highest since August 2020, and 45% named buying chipmakers’ shares the most crowded trade in the market 150 M.

08 · HistoryRailways, the telegraph and fibre: they built ahead of demand, investors lost money, and the networks served the next generation

The hypothesis we started with rests on a historical analogy: AI infrastructure repeats the railways and fibre. Construction runs ahead of revenue, revaluation of assets hits market capitalisation, and the capacity becomes the cheap raw material of the next decade. History confirms this pattern, but with important qualifications.

Britain’s Railway Mania of the 1840s

In 1844 the British Parliament authorised 810 miles of new railways, in 1845 2,816 and in 1846 4,540; in 1846 alone 263 acts creating new railway companies were passed 151,152 H. Investment peaked in 1847 — according to a reconstruction by the researcher Andrew Odlyzko, at about 7% of the country’s GDP, more than twice military spending 18 L. Railway share prices roughly tripled between 1843 and 1845; after the Bank of England raised its discount rate on 16 October 1845, a long decline began — by the end of 1849 about 60% below the peak 153,18 M.

Construction lagged the money by one to two years: authorisations peaked in 1846, spending in 1847 and openings of new lines in 1848, at 1,182 miles 154 H. When the new tracks went into operation, revenue per mile fell by 27% (1845–1848) and returned to its 1845 level only in 1856 (£2,527 against £2,499) 18 L. Lines authorised in 1845 promised a 7% return on capital, while the new lines of 1844–1846 earned, by Odlyzko’s rough estimate, about 4.5% 18 L. The main mistake was demand forecasts, not technology 18 M.

By the end of 1850 the British had invested about £250m in railways — almost half of annual GDP — and lost, by Odlyzko’s estimate, about a third of it, £80m 18 L. But almost no line went bankrupt to the point of closing: the tracks kept working, and the country, most likely, gained 18 M. The network grew from 1,498 miles in 1840 to 6,621 miles in 1850 (according to Odlyzko’s table 2) 18 L. Cumulative investment in British railways reached £1bn only in 1890, and the profits that The Times had promised in the 1840s appeared only shortly before the First World War 18 M.

The US: record building, and between 15% and a quarter of the network in receivership

American railways were built in waves. In 1887 a record 12,876 miles of track were laid 155 H. The Panic of 1873 began with the failure of Jay Cooke, the bank that financed the Northern Pacific; by November 55 railways had failed 156 M. In 1893 alone more than 27,000 miles of railway, with a capitalisation of almost $2bn, went into receivership (by Stuart Daggett’s count, 1908) 157 M. By our estimate, that is about 15% of the network at the time L. Bradstreet’s magazine for 1893 put the share higher, at 24% of network length, but this estimate is disputed 158 L. The railways went on carrying freight. And from 1920 the law set a “fair” return for them — 5.5% on the value of property, later 5.75% — and the state took half of any surplus above 6%. The railways rarely reached this return, so they became a regulated industry with a capped return 159 M.

The telegraph, electricity and cars

The first reliable transatlantic cable, in 1866, charged $10 a word and ran at about 5% of capacity. The next year the tariff was halved, and daily revenue rose by only about 12% 160 M. Cheaper prices do not fill capacity at once. In the 1920s American power holding companies built ahead of demand: in 1929 peak-hour reserve capacity was 27.5% 161 M. Samuel Insull’s empire, which covered 39 states, went into receivership in April 1932 with losses of almost $800m 162 M; from 1929 to 1936, 53 power holding companies failed, while the power stations kept running 145 H. The number of car makers in the US fell from 88 in 1921 to 44 in 1927 163 M.

Fibre, 1996–2002: the closest analogy

The “cheaper means less is needed” hypothesis is best tested by fibre. In the five years after the Telecommunications Act of 1996, US telecoms companies invested more than $500bn in networks, mostly with borrowed money 164 M. Meanwhile the technology advanced twice as fast as Moore’s law: the capacity of a single fibre rose from 2.5Gbit/s in 1996 to ~100Gbit/s in 2000, doubling about every nine months 165 H. Yet consumption of telecoms services grew by only 7.4% a year 165 H. Internet traffic doubled every year, not every 100 days as investors then believed 166 M. This is where falling prices outran demand, and in 2001 transatlantic cables were lit to 11.6% of potential capacity 167 M.

The consequences are well known. Global Crossing, valued at $47bn at its peak, went bankrupt in January 2002 and was bought in 2011 for $3bn 168 M. On 21 July 2002 WorldCom filed what was then the largest bankruptcy in US history, with assets of more than $107bn 169 M. Nortel’s market capitalisation fell from C$398bn to less than C$5bn, and Lucent’s shares fell by 99% 170 M. Holders of telecoms bonds recovered a little over 20% 164 M. By Odlyzko’s estimate, the real excess of investment over trend was only ~$150bn, and the stock-market losses were many times larger 18 L.

And then the surplus really did become a raw material. The price of IP transit fell from $1,200 per Mbit/s a month in 1998 to $75 in 2005 and $5 in 2010 19 L; in 2026 the weighted median price in key US and European cities is $0.08 20 M. In 2005 Google was buying up unneeded “dark” fibre 171 L, YouTube appeared the same year, and Google bought it for $1.65bn in 2006 172 M; Netflix launched video streaming in January 2007 173 M. According to an iSuppli analyst, the fibre surplus was absorbed only about ten years later 174 L. The economist William Nordhaus calculated that of the social return on technology in the US in 1948–2001, the innovating companies captured only a meagre share — a few per cent (Tim Harford cites 3.7%; the figure has not been checked in the paper itself) 175 L.

How big the current boom is

Comparisons of booms as a share of GDP depend on definitions, so we show them separately. Stijn Van Nieuwerburgh, an economist at Columbia University, in a Brookings podcast summarising his paper, compared average US spending over the years of each boom (he himself warns that the estimates for different booms are not harmonised): railways in 1870–1890 — 2.2% of GDP, highways — 1.1%, telecoms and fibre — 1.1%, electrification — 0.5%; AI building in 2025–2032, by his forecast, averages 3.6% a year, $10.3trn in total 176 M. Torsten Slok, chief economist of Apollo, counts only hyperscalers’ capital expenditure: 0.6% of GDP in 2023, 1.4% in 2025 and, on analyst consensus, about 3% a year in 2027–2029 177 M.

The analyst Paul Kedrosky rejects the analogy itself: “this is not railways”, because AI data centres are short-lived assets that need their equipment constantly replaced 178 M. He is right about the servers. Rails served for a century, and fibre in conduits for decades. Of the $37.9bn that Epoch AI estimates a gigawatt of capacity costs, 56% goes on servers, which companies depreciate over 5–6 years (see section 05) 115 M. What may become the cheap raw material after the boom is the buildings, land, utility connections and cooling — about a third of the cost (31%). Networking equipment (another 13%) Epoch classes as IT equipment together with servers, so only that third is long-lived. But even old chips, as the A100 at $1.56 an hour shows, do not become free 107 M.

The economist Carlota Perez describes each technological revolution as an installation period — 20–30 years ending in a financial frenzy — followed by a turning point (crash and recession) and a deployment period of another 20–30 years, when the technology pays off across the whole economy 179 M. She regards the excess investment in telecoms as useful: the frenzy left infrastructure that everyone used 179 M. If AI is a new installation phase, the turning point is still ahead; but on Perez’s own scheme the turning point of the ICT revolution falls in 2000–2003 (our reading of her scheme), so AI can also be assigned to the deployment of that same revolution 179 L. The timeline below shows how this played out with earlier networks.

09 · Forecast2027–2040: will capex exceed $1trn a year, will an answer of fixed quality get another 1,000 times cheaper, and what it means for valuations

2027: more than a trillion

No big company has promised to cut spending in 2027. Alphabet’s chief financial officer said it would “increase significantly” 180 M; Microsoft expects growth in fiscal 2027 24 H; Meta already has $121.4bn of non-cancellable purchase commitments (cloud capacity, servers, networking, energy) payable in 2027, but has not given its own capex figure for 2027 7 H. Forecasts for 2027 range from more than $1.1trn (JPMorgan, hyperscalers) 123 M to $1.3trn (S&P Global Ratings, six companies including SpaceX; as relayed by CryptoBriefing, the original did not open) 181 L. Epoch AI’s mechanical extension of the trend gives $1.19trn for five companies, against operating cash flow of about $0.9trn 28 L. In April 2026, before the July increase in plans, Barclays expected the big four plus Oracle to spend almost 90% of their operating cash flow on capex in 2027 182 M.

Long-term forecasts only move up. In August 2026 Dell’Oro Group raised its forecast for global data-centre capex (all data centres, not just AI) to more than $3trn a year in 2030, nearly double its January figure; the words “a year” are our reading of the release 183 M. Between January and September 2026 Gartner raised its forecast for global AI spending this year from $2.53trn to $2.7trn 184 L. Forecasters seem to follow company plans rather than test demand, so a forecast is not evidence of payback (our estimate) L.

Will an answer get another 1,000 times cheaper by 2040?

For the price of an answer of fixed quality to fall another 1,000-fold in 14 years, it needs to get cheaper by only 1.64 times a year (our calculation) L. That is far slower than the record so far: Epoch AI measured declines of 9–900 times a year 4 M, and OpenAI’s chief executive, Sam Altman, claims that the cost of a given level of AI falls about tenfold every year 185 L. So a thousandfold fall by 2040 for old quality is, if anything, a cautious scenario. But Epoch itself warns that the fastest decline came in the most recent year and may not be repeated 4 M; a preprint by MIT researchers puts the typical pace at 5–10 times a year 57 M. And hardware gets cheaper by only about 1.4 times a year 56,3 H; over 14 years that is roughly 100–150-fold (our calculation) L. As section 03 shows, the price of flagship models has not fallen on average since 2025: OpenAI’s more than doubled, while Anthropic cut the price of Opus threefold but launched an even dearer tier. So the main question to 2040 is not the price of a token but how many tokens new tasks will need.

More than 3% of GDP

The hypothesis that AI spending will exceed 3% of GDP already has support in forecasts for the United States. According to a consensus cited by Apollo, hyperscalers’ capex will be about 3% of GDP a year in 2027–2029 177 M; Van Nieuwerburgh calculates 3.6% on average over 2025–2032 if data centres, energy and chips are included 176 M. For the world economy to 2040 we found no such forecast from a well-known institution L. For comparison: American railways averaged 2.2% of GDP over some 20 years of boom (1870–1890) 176 M, and British railway investment at the peak of 1847 was about 7% of GDP in Odlyzko’s reconstruction 18 L; an average over several years and the peak of a single year cannot be compared.

Estimates of the effect on the economy vary widely. Bridgewater calculates that AI capex will add 1.4 percentage points to US GDP growth in 2026 and 1.5 in 2027, but warns that a large part of this is the profits of chipmakers, which do not flow back into the economy 186 H. The Nobel laureate Daron Acemoglu estimates the gain in America’s total factor productivity from AI at no more than 0.66% over ten years 187 H; Goldman Sachs in 2023 put it at +7% of global GDP over ten years under a hypothetical scenario 188 M; these are different quantities and cannot be compared directly.

What the builders say

The bosses of the biggest companies admit the risk of oversupply but think that falling behind is worse. In October 2025 Jeff Bezos called AI “more of an industrial bubble than a financial one” that would still bring society enormous benefits 189 M. In November 2025 Sundar Pichai spoke of “elements of irrationality” and said that no company, Google included, would be immune if the bubble burst 190 M. In September 2025 Mark Zuckerberg called an AI bubble “quite possible”: to “misspend a couple of hundred billion dollars” would be regrettable, but he saw slowness as the bigger risk 191 M. In August 2025 Sam Altman said that investors as a whole are overexcited about AI 189 M. By Gartner’s estimate, more than $3.5trn of spending on AI infrastructure could be at risk of write-off by the end of 2030 192 L.

What to check in the reports of 28–29 October

Microsoft reports on 28 October and Amazon on 29 October 193,194 H; Alphabet, according to earnings calendars, also on 28 October 195 L; Meta reported on 21 October. The most important figures: guidance for capex in 2027 (for the four, below the 2026 plan of about $0.74trn would mean a fall; about $0.9trn a sharp slowdown in growth; $1.1–1.3trn, as in forecasts for wider lists of companies, a continuation of the trend); the ratio of capex to operating cash flow; growth of Azure, Google Cloud and AWS; depreciation; backlogs and the share of OpenAI and Anthropic in them. In April 2026 Noah Weisberger of BCA Research suggested that investors might give companies about a year (that is, until spring 2027) to show that spending turns into revenue and cash 182 M.

Our estimate: what is likely by 2030–2040

10 · LimitsHonesty and limits