> ## Content Index
> Fetch the complete content index at: https://www.wealthyparrot.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Who is actually paying for the AI boom?
- URL: https://www.wealthyparrot.com/who-is-actually-paying-for-the-ai-boom/
- Published: 2026-10-08T05:30:45.000Z
- Updated: 2026-10-08T05:30:45.000Z
- Author: Carmelo 
- Tags: Jumble

The Brontë sisters owned railway shares. So did Charles Darwin and the novelist William Thackeray, according to the historians William Quinn and John Turner. By one estimate, about a quarter of all members of Parliament did too. In the mid-1840s, half of respectable Britain was betting on the new technology.

Charlotte Brontë was the family sceptic. She expected the bubble to burst and told Emily and Anne to sell. She had good reason. In November 1845, *The Times* added up every railway scheme on the table. The capital needed came to close to **£600 million**. Britain's entire national income was about £550 million.

The railway share index peaked that summer. By April 1850 it had fallen 66%, according to Quinn and Turner's *Boom and Bust*.

And yet almost none of the lines that got built went bankrupt. By 1850 private investors had sunk about £250 million into railways, almost half of Britain's GDP. They [lost about a third of it](https://www-users.cse.umn.edu/~odlyzko/doc/hallucinations.pdf?ref=wealthyparrot.com). The trains kept running.

Something with the same shape is happening now. The five biggest cloud companies (Amazon, Microsoft, Alphabet, Meta and Oracle) are expected to spend more than **$690 billion** in their 2026 fiscal years, [according to FactSet](https://insight.factset.com/hyperscalers-tap-external-financing-as-ai-capex-outruns-cash-flow?ref=wealthyparrot.com). Most of it goes on AI data centres, chips and power. That's about 2.2% of the US economy, which the World Bank puts at $30.8 trillion for 2025.

Call it the *Field of Dreams* bet at planetary scale ("build it, and they will come"). Everyone is arguing about whether the customers will come. I think the more useful question is: who is actually paying for all this, and who is left holding the bill if the returns arrive later than hoped?

## The cash register is running out

For twenty years, big tech didn't need anyone's money. Its businesses threw off so much **operating cash flow** (the cash a company's day-to-day business produces) that they paid for everything themselves. There were still billions left over to buy back their own shares.

That era is ending. [A January 2026 study by the BIS](https://www.bis.org/publications/bulletin-120-financing-ai-boom-cash-flows-debt.pdf?ref=wealthyparrot.com) found that the leading AI firms' free cash flow (the cash left once new equipment is paid for) has recently fallen below their capital spending. The BIS is the Bank for International Settlements, the central bank for central banks. FactSet expects free cash flow near zero, or below it, in fiscal 2026 for all the big five except Alphabet and Microsoft.

Oracle shows it most clearly. Its [annual report](https://www.sec.gov/Archives/edgar/data/1341439/000119312526277521/orcl-20260531.htm?ref=wealthyparrot.com) for the year to May 2026 records $55.7 billion of capital spending against $32.0 billion of cash from operations. The gap has to come from somewhere, and there are four places to find it. Each one puts the risk on a different group of people.

## Layer one: selling more of the company

The cleanest option is to sell new shares, which I looked at from the buyer's side in [The IPO Trap](https://www.wealthyparrot.com/the-ipo-trap/). It's also the rarest, because it shrinks every existing owner's slice of the pie.

Alphabet did it anyway. In June 2026 it [announced an equity raise of **$84.75 billion**](https://abc.xyz/investor/news/news-details/2026/Alphabet-Announces-Upsize-and-Pricing-of-84-75-Billion-Equity-Capital-Raise-to-Expand-AI-Infrastructure--and-Compute-2026-QzN3D9yMAj/default.aspx?ref=wealthyparrot.com) to expand AI infrastructure. That included $10 billion sold directly to Berkshire Hathaway. FactSet calls it the largest equity capital transaction ever done by a listed company.

When a build-out is funded with shares, the shareholders carry the risk, and everyone knows exactly what they bought. That will matter later.

## Layer two: borrowing in the bond market

The second option is ordinary corporate debt, and here the numbers have moved fast. Bond issuance by the big cloud firms topped $100 billion in 2025, mostly long-dated, according to the [BIS Quarterly Review for March 2026](https://www.bis.org/publ/qtrpdf/r%5Fqt2603u.htm?ref=wealthyparrot.com).

FactSet measures it against the spending. New debt covered 9% of capital spending in fiscal 2024 and 32% in the twelve months to June 2026\. In two years, borrowed money went from a rounding error to a third of the bill.

In October 2025 Meta sold $30 billion of ordinary bonds, the biggest corporate bond deal of the year. That debt sits on Meta's balance sheet for anyone to see.

Borrowing isn't bad in itself. A data centre lasts decades, and matching a long asset with a long loan is what a sensible mortgage does.

What matters is who holds the bonds, mostly pension funds, insurers and bond ETFs. Bondholders don't share the upside if AI makes these companies twice as rich. They collect their interest and hope to be repaid, and as I showed in [Why not just buy 30-year bonds at 6%?](https://www.wealthyparrot.com/why-not-just-buy-30-year-bonds-at-6/), a long bond carries its own risks even when it is repaid.

## Layer three: private credit, the lender you've never met

**Private credit** is lending by investment funds instead of banks. The fund negotiates directly with the borrower and usually holds the loan until it's repaid. There's no public price, which makes it fast and flexible, and hard to see into.

It grew up largely financing private-equity deals like the ones I described in [Barbarians at the Gate](https://www.wealthyparrot.com/how-leveraged-buyouts-work/). Now it's one of the main engines of AI financing. The BIS puts outstanding private-credit loans to AI-related firms at over $200 billion, up from close to zero a decade ago.

Funds made over $40 billion of new AI loans in 2025 alone, against about $3 billion in 2010\. The BIS projects $300-600 billion by 2030.

> **Fun fact:** according to the BIS, private-credit loans to AI firms carry almost exactly the same **spread** as loans to any other company. The spread is the extra interest a lender charges over a safe benchmark rate, and here it's 6.2 percentage points against 6.1.

Stock markets price AI companies as if they'll be vastly more profitable in the future. Lenders price them like any ordinary business. Both can't be right, and the BIS authors spell out the two options. Either lenders are underestimating the risk "just as their exposures are growing significantly," or stock markets are overestimating the cash AI will generate.

Quinn and Turner describe a bubble as a fire that needs three things, and "the fuel for the bubble is money and credit." Private credit is a lot of new fuel arriving at once.

## Layer four: the debt that sits somewhere else

The fourth layer is the cleverest, and the one regulators watch most closely.

In October 2025, Meta [announced a joint venture](https://investor.atmeta.com/investor-news/press-release-details/2025/Meta-Announces-Joint-Venture-with-Funds-Managed-by-Blue-Owl-Capital-to-Develop-Hyperion-Data-Center/default.aspx?ref=wealthyparrot.com) to build Hyperion, a gigawatt-scale data-centre campus in rural Louisiana costing about $27 billion. Funds managed by Blue Owl Capital own 80% of it, and Meta owns 20%.

The money was raised through a **special purpose vehicle** (SPV), a separate company set up to hold one project and its debt. That vehicle, Beignet Investor LLC (yes, named after a pastry), sold [**$27.3 billion** of bonds](https://www.ifre.com/ifr-awards/2327933/financing-package-blue-owl-capitalbeignet-investors-us27.3bn-23.6-year-bond?ref=wealthyparrot.com) due in 2049\. Because Meta owns only a minority stake, that debt isn't Meta's borrowing. Meta just rents the campus.

Think of a restaurant that wants a bigger kitchen but not a bigger mortgage. It gets friends to set up a company that borrows the money, builds the kitchen and rents it back. The debt is real, it just lives at a different address.

> **Fun fact:** the bonds run until 2049, but Meta's lease has a four-year initial term, with options to extend. Investors are lending for 23 years to a building whose main tenant has firmly committed to four.

That sounds alarming, and the full picture is more balanced. Meta guaranteed the first 16 years, up to a cap and only under certain conditions. S&P rated the bonds A+, a high grade, precisely because the structure passes "substantial credit risk to Meta during both construction and operation phases," as IFR reports.

The bonds also **amortise**, repaying some of the loan every year instead of all of it at the end. And accounting rules such as [IFRS 16](https://www.ifrs.org/issued-standards/list-of-standards/ifrs-16-leases/?ref=wealthyparrot.com) make Meta record a liability for the lease it has signed. With only four years committed, though, that liability is a fraction of the $27 billion borrowed.

Hyperion isn't a one-off. In July 2026 a second Meta venture, with BlackRock's infrastructure funds, sold about [$12.5 billion of similar bonds](https://pitchbook.com/news/articles/investors-flock-to-12-5b-data-center-bond-deal-for-blackrock-sponsored-sopaipilla?ref=wealthyparrot.com) for a campus in El Paso. It paid a higher premium over US government bonds than Hyperion did, a sign that investors are starting to ask for more.

Oracle shows how big this layer can get. Its latest annual report discloses **$260 billion** of lease commitments that haven't started yet, "substantially all related to data center arrangements." None of it is on the balance sheet.

Companies reach for these structures when they carry a lot of debt and are short of their own cash, according to [a 2009 accounting study](https://doi.org/10.2308/accr.2009.84.6.1833?ref=wealthyparrot.com), which describes the AI build-out quite well. SPVs have legitimate uses, and Hyperion is disclosed and rated. But the BIS authors made the general point in one sentence: "leverage does not disappear by being out of sight." (That's finance-speak for borrowed money.)

## The accounting choice nobody can settle yet

How long did your last laptop feel fast? Big tech has to answer the same question about millions of servers, and the answer moves its profits by billions.

When a company buys a server, it doesn't count the whole cost as an expense on day one. It spreads the cost over the equipment's expected **useful life**, a process called depreciation. Assume a longer life and each year's expense shrinks, so reported profit rises.

Microsoft's [fiscal 2023 annual report](https://www.sec.gov/Archives/edgar/data/789019/000095017023035122/msft-20230630.htm?ref=wealthyparrot.com) explains that it stretched the assumed life of its servers and network equipment from four years to six. That added $3.7 billion to the year's operating income. Alphabet [made a similar move](https://www.sec.gov/Archives/edgar/data/1652044/000165204424000022/goog-20231231.htm?ref=wealthyparrot.com) to six years, cutting its 2023 depreciation by $3.9 billion.

Amazon went both ways. It lengthened server lives from five years to six in 2024, cutting depreciation by $3.2 billion. Then its [2024 annual report](https://www.sec.gov/Archives/edgar/data/1018724/000101872425000004/amzn-20241231.htm?ref=wealthyparrot.com) shortened some of them back to five. The reason given was "an increased pace of technology development, particularly in the area of artificial intelligence and machine learning."

![Changing how long servers are assumed to last moved annual profits by billions: Microsoft +$3.7bn to operating income (fiscal 2023, four to six years); Alphabet $3.9bn less depreciation (2023, to six years); Amazon $3.2bn less depreciation (2024, five to six years); Amazon about $0.7bn less operating income expected (2025, some servers back to five years)](https://www.wealthyparrot.com/content/images/2026/09/2026-09-27-who-is-paying-for-the-ai-boom-chart-2-2.png)

This matters for the financing. A data-centre building might stand for 40 years, while the chips inside may be outdated much sooner. When a lender accepts a data centre as collateral, the question is how much of its value is the durable shell and how much is hardware that loses value quickly. The BIS notes that investors have started asking about "the long-term value of collateral" in some of the largest deals.

> **Fun fact:** railway investors weren't the only Victorians caught out by fast-moving technology. Steamships built in 1870-73 were being sold "for half, or less than half, of their original cost" by 1875-76, as newer designs made them obsolete. The economist Carlota Perez quotes the account.

I don't know who's right about how long an AI chip earns its keep, and I'm wary of anyone who claims certainty. But the length of a lease, the life of a chip and the maturity of a bond are three different numbers. Someone is carrying the risk in the gaps.

## Everyone is investing in everyone

Follow the money in this part of the story and you'll need a whiteboard (and possibly a lie-down).

Microsoft owns about 27% of OpenAI. In the same agreement, OpenAI [committed to buy an extra $250 billion of Microsoft's Azure cloud services](https://blogs.microsoft.com/blog/2025/10/28/the-next-chapter-of-the-microsoft-openai-partnership/?ref=wealthyparrot.com). Weeks later, Microsoft and Nvidia [invested in OpenAI's direct rival, Anthropic](https://blogs.microsoft.com/blog/2025/11/18/microsoft-nvidia-and-anthropic-announce-strategic-partnerships/?ref=wealthyparrot.com), which promised to buy Azure computing in return.

Google competes with Anthropic through its own Gemini models, and has still [committed up to $40 billion to it](https://techcrunch.com/2026/04/24/google-to-invest-up-to-40b-in-anthropic-in-cash-and-compute/?ref=wealthyparrot.com). Anthropic runs much of its work on Google's chips. Amazon is [an investor too](https://www.aboutamazon.com/news/aws/amazon-invests-additional-4-billion-anthropic-ai?ref=wealthyparrot.com), and its cloud is one of Anthropic's primary providers.

![Everyone is investing in everyone. Investments: Microsoft owns about 27% of OpenAI (valued at about $135bn); Microsoft up to $5bn and Nvidia up to $10bn into Anthropic; Nvidia up to $100bn into OpenAI; Google $10bn into Anthropic plus up to $30bn more; Amazon $8bn into Anthropic. Purchase commitments: OpenAI $250bn of Microsoft Azure; Anthropic $30bn of Azure; Anthropic 5 GW of Google Cloud TPUs; Anthropic uses AWS as its primary cloud and training partner; OpenAI to deploy at least 10 GW of Nvidia systems. Backstop: Nvidia buys CoreWeave's unsold capacity, up to $6.3bn to 2032](https://www.wealthyparrot.com/content/images/2026/09/2026-09-27-who-is-paying-for-the-ai-boom-network.png)

In every gold rush, the steadiest money goes to whoever sells the shovels. Here that's Nvidia, which designs the chips most of this money buys (made on machines from a single Dutch supplier, [ASML](https://www.wealthyparrot.com/asml-euv-lithography-monopoly-explained/)).

Its [fiscal 2026 results](https://www.sec.gov/Archives/edgar/data/1045810/000104581026000019/q4fy26pr.htm?ref=wealthyparrot.com) show $193.7 billion of data-centre revenue. In the final quarter it kept about 75 cents of every sales dollar after the cost of making the chips, a **gross margin** most companies can only dream of. Then the shovel seller sends money back out to the miners. Nvidia [intends to invest up to $100 billion in OpenAI](https://nvidianews.nvidia.com/news/openai-and-nvidia-announce-strategic-partnership-to-deploy-10gw-of-nvidia-systems?ref=wealthyparrot.com) as OpenAI deploys Nvidia systems.

The BIS calls this **circular financing**. Suppliers and cloud firms invest in AI labs, and the labs then commit to buying from those same suppliers and clouds.

Picture a Monopoly game where the players lend each other cash to buy hotels on each other's streets. Money changes hands and the board looks busier, but no new cash has come in from the bank. The terms of the real deals are "poorly disclosed," says the BIS. No filing I could find reports how much of anyone's revenue comes from companies it has invested in.

We've seen a version of this before. In the late 1990s, telecom equipment makers lent money to their own customers, mostly young phone companies, so those customers could buy their equipment. When the customers collapsed in 2001, the loans went bad with them, as Newsweek reported the following year.

The difference today matters, and I want to be fair about it. Lucent's customers were start-ups with almost no revenue. OpenAI and Anthropic bring in tens of billions of dollars a year. And Google's stake in a rival is also a hedge, a bet that pays off if Anthropic beats Google's own Gemini, as well as a way to sell chips.

Strategic investment is normal business. The honest test is whether money from outside customers grows faster than the money going round in a circle, so let's look at the customers.

## The other side of the ledger

Some of the AI bill is paid by people like you and me, and that part is growing very fast.

In August, [Bloomberg reported](https://www.bloomberg.com/news/articles/2026-08-13/openai-s-revenue-run-rate-tops-40-billion-ahead-of-ipo?ref=wealthyparrot.com), citing people familiar with the matter, that OpenAI was on pace for more than **$40 billion** of annualized revenue. That's roughly double the "$20B+" its finance chief, Sarah Friar, [said it ended 2025 with](https://sherwood.news/business/openais-arr-reached-over-usd20-billion-in-2025-cfo-says/?ref=wealthyparrot.com) (hold on to that number). In February, OpenAI [said](https://openai.com/index/scaling-ai-for-everyone/?ref=wealthyparrot.com) ChatGPT had more than 900 million weekly users, over 50 million of them paying consumer subscribers.

Anthropic's annualized revenue reached $65 billion by the end of July, according to [Bloomberg](https://www.bloomberg.com/news/articles/2026-08-17/anthropic-revenue-run-rate-surpasses-65-billion-ahead-of-ipo?ref=wealthyparrot.com) and [TechCrunch](https://techcrunch.com/2026/08/17/anthropics-annualized-revenue-surges-to-65b/?ref=wealthyparrot.com). Whole new businesses sit on top of these models, like the voice company ElevenLabs, which [passed $500 million](https://elevenlabs.io/blog/500m-arr-and-new-investors?ref=wealthyparrot.com) of annual recurring revenue this year, and the coding platform Replit, at a reported $525 million annualized in April.

Three caveats keep this honest. First, "annualized" means the latest month multiplied by twelve, which flatters fast growers. It's a bit like judging your yearly pay from your best month. Against that $20 billion-plus run rate, the revenue OpenAI actually booked for 2025 was about $13 billion, according to audited accounts [reported in June](https://finance.yahoo.com/markets/stocks/articles/openai-2025-financials-leaked-38-121508294.html?ref=wealthyparrot.com).

Second, revenue isn't profit. OpenAI's own projections, as reported, show a loss of about [$14 billion in 2026](https://finance.yahoo.com/news/openais-own-forecast-predicts-14-150445813.html?ref=wealthyparrot.com).

Third, some of the revenue is investor money in disguise. [Many app companies pay the model makers more than their own customers pay them](https://techcrunch.com/2025/08/07/the-high-costs-and-thin-margins-threatening-ai-coding-startups/?ref=wealthyparrot.com). When a start-up funded by venture capital pays OpenAI, part of OpenAI's "customer revenue" is that venture capital, one step removed.

Even so, the two biggest labs together bring in roughly $105 billion a year at their current pace, against $690 billion of capital spending this year. Picture a café taking €105 a day after spending €690 on its espresso machine. That's fine if the machine lasts for years and the queue keeps growing, and a problem if it needs replacing soon.

In 2024, Sequoia's David Cahn estimated that the industry needed about [$600 billion of annual revenue](https://sequoiacap.com/article/ais-600b-question?ref=wealthyparrot.com) to justify its spending. Spending has grown faster than revenue since.

![The AI build-out in six numbers, in billions of dollars: big-five cloud capital spending, fiscal 2026 (expected), more than 690; Oracle lease commitments not yet on its balance sheet, 260; Nvidia data-centre revenue, fiscal 2026, 193.7; private credit lent to AI firms, outstanding, more than 200; OpenAI plus Anthropic annualized revenue, mid-2026, about 105; big-tech bonds issued in 2025, more than 100](https://www.wealthyparrot.com/content/images/2026/09/2026-09-27-who-is-paying-for-the-ai-boom-chart-1-2.png)

The value goes beyond revenue, too. [Erik Brynjolfsson and two colleagues studied 5,179 customer-support agents](https://doi.org/10.1093/qje/qjae044?ref=wealthyparrot.com) and found that an AI assistant raised issues resolved per hour by 14% on average. For novices the gain was 34%.

In medicine, rentosertib is a lung-disease drug whose target and molecule were both found with AI. It improved lung function in [a randomized trial](https://doi.org/10.1038/s41591-025-03743-2?ref=wealthyparrot.com) and [entered a final-stage trial](https://insilico.com/news/xmjsn4l091-insilico-initiates-phase-iii-clinical-tr?ref=wealthyparrot.com) in July 2026\. No AI-discovered drug has been approved yet, so that value is still a promise.

Real value, though, doesn't guarantee a return for whoever paid for it. [Two Chicago economists](https://doi.org/10.1257/aer.99.4.1451?ref=wealthyparrot.com), Ľuboš Pástor and Pietro Veronesi, put it in one line: "In the long run, new technology tends to benefit workers and consumers, not producers."

## What history says about who pays

We have run this experiment before. The pattern of who paid is consistent, though less tidy than it's usually told.

**The railways, 1840s.** Shareholders lost heavily, and the lines survived, though not efficiently. Quinn and Turner cite an estimate that about 7,000 of the roughly 20,000 miles running by 1914 were more than the country needed.

The losses didn't stay with investors either. Railway shares were sold **partly paid**, like buy-now-pay-later where the seller picks when "later" arrives.

When prices collapsed after 1845, speculators were "held to their contractual obligation to pay the subsequent share calls," as Chancellor records. Those calls drained bank deposits, helped push interest rates to around 10% by 1847, and fed that year's crisis. Commitments made in the boom came due in the bust.

**Telecom and fibre, 1990s.** The venture investor William Janeway cites an estimated $4 trillion of equity and debt raised for broadband networks before the bubble burst. Much of it was built far ahead of demand. Perez cites a 2001 *Financial Times* estimate that only 1-2% of the fibre buried under Europe and the US had been switched on.

> **Fun fact:** one of Janeway's own investments, the broadband start-up Covad, was offered $300 million of **junk bonds** (high-interest debt from companies with shaky credit). Its revenue for all of 1997 was $26,000\. Janeway's firm sold out before the crash, turning $6 million into just over $1 billion (nice timing). Covad later went through bankruptcy.

Then came the payoff, for someone else. [The Nasdaq quadrupled](https://doi.org/10.1257/aer.99.4.1451?ref=wealthyparrot.com) between 1996 and March 2000 and fell back to its 1996 level by October 2002, just before US productivity growth sped up from about 1% a year to 2.5%. The economist Brad DeLong summed it up in a line Janeway quotes: "Investors lost their money. We will now get to use their stuff."

![Three build-outs, one pattern: investors paid for the infrastructure, users got the benefit. Railways: 1845, about £600m of railway commitments, more than Britain's national income; 1845-1850, the railway share index falls 66%; 1905, railway revenues reach 6% of GDP. Telecom and fibre: 1990s, about $4tn raised for broadband networks; 2001, only 1-2% of buried fibre switched on; 2000-2002, about $2tn of telecom market value lost. AI and cloud: 2026, more than $690bn of big-five cloud capital spending, then a question mark: who ends up holding the bill?](https://www.wealthyparrot.com/content/images/2026/09/2026-09-27-who-is-paying-for-the-ai-boom-timeline.png)

## The flip side: this might not end badly

It's easy to write a scary article about big numbers, so here is the other side.

The BIS itself says the AI build-out "is not particularly large by historical standards." Spending on data centres and chip factories is about 1% of US GDP, roughly the size of the 2010s shale boom and half the 1990s IT boom, and the BIS judges the risks "moderate." [A July 2026 Federal Reserve note](https://www.federalreserve.gov/econres/notes/feds-notes/do-major-technology-advancements-lead-to-overinvestment-20260706.html) asks whether AI investment should be curbed pre-emptively, and answers "Not necessarily."

The word "bubble" is also less useful than it sounds. [Pástor and Veronesi](https://doi.org/10.1257/aer.99.4.1451?ref=wealthyparrot.com) show the rise-and-fall pattern is easy to spot in hindsight and impossible to predict in advance, even when every investor is rational.

The biggest difference from the telecom bust is who is paying. Most of today's spending comes from companies with enormous existing profits, not start-ups living on junk bonds. Demand for computing is real and paid for.

There is a risk in the other direction too. When the first wave of AI hype ended in the 1980s, Janeway points out, the field sank into an "AI winter" that lasted into the twenty-first century. A boom that ends badly can set a technology back as well as leave it behind. Ethan Mollick quotes Amara's law, the best summary I know: "We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run."

## Where you might already be holding some of this

Do you own a global or US index fund? Then you own shares in the companies doing the spending. That's layer one, and it's the contained kind of risk, because you share in both the upside and the losses. Several of these companies are among the largest in the index (a truly [global index](https://www.wealthyparrot.com/why-smart-investors-are-finally-going-global/) spreads that risk further).

If you own a broad bond fund, or a pension invested in one, you may hold some of the $100 billion-plus of big-tech bonds. That's layer two.

If you have money with an insurer or a pension fund, it may hold private credit or SPV bonds like Hyperion's. That's layers three and four. The BIS names insurers and private-credit vehicles as the channels a shock would travel through.

The same goes for a private credit fund or an "income" product with an attractive yield. By the BIS's count, AI-related companies went from under 1% of all private-credit loans to almost 8%.

Living in Europe doesn't put you at a distance from any of this. A world ETF sold in Europe (look for "UCITS" in the name, the EU's rulebook for retail funds) holds the same US tech giants as its American cousins. A euro bond fund holds big-tech debt only if those companies borrowed in euros, which the fund's holdings list will tell you. The EU has also been opening private markets to ordinary savers through ELTIFs (European Long-Term Investment Funds), and I couldn't find any published figure for how much of that money reaches AI.

None of this is a reason to panic or sell. Owning a small slice of a historic build-out is what diversified investing looks like.

## Practical takeaways

- **Know your concentration.** Look up the top 10 holdings of your main index fund. If the same handful of companies dominates it, that's your AI exposure, whether you chose it or not.
- **Ask what the "income" money lends to.** For a private-credit fund or an ELTIF, ask what it lends to, how often you can get your money out, and how loans are valued without a market price. Ask your pension provider what share of the fund is in private credit or private infrastructure. If it can't answer, that's information too.
- **Watch customer revenue, not announced investments.** Circular deals make headlines. What pays for the build-out in the end is money from customers outside the circle, and that's the number worth following.
- **Don't try to call the top, and diversify instead.** (I made the case in [Is It Too Late to Invest at All-Time Highs?](https://www.wealthyparrot.com/investing-at-all-time-highs/)) Quinn and Turner's advice for amateurs is to sit out bubbles, and the practical way is to avoid concentrating in the hottest names rather than selling the market and waiting.

## The railway you still ride

In 1905, half a century after the mania, Britain's railways brought in revenues equal to **6% of GDP**, according to Odlyzko. That was far more than the investors of the 1840s had expected. It just arrived fifty years late, long after most of them had sold at a loss.

The AI build-out may follow the same path, or it may pay off for the people funding it this time. Either way, the question for your own portfolio is which side of DeLong's line you're on: the investors who lose their money, or the people who get to use their stuff. A diversified investor gets to be a little of both.