Following AI exposure across borders, and finding where the statistics go dark

Earlier this month I wrote about Stijn Van Nieuwerburgh’s NBER paper, “Financing the AI Buildout”. His numbers are hard to forget. A one-gigawatt AI campus costs about $41 billion, and a U.S. buildout of 188 gigawatts by 2032 would need nearly $9 trillion of investment, about 3.2 percent of GDP a year. As the hyperscalers’ capital spending outgrows their cash flow, the financing is moving into leases, joint ventures, project debt, private credit, securitization and special-purpose vehicles.

In “The AI Boom’s Invisible Balance Sheet”, I mapped part of that financing network from public filings. My conclusion was that the main problem is visibility: the people best placed to judge the risk cannot yet see the whole picture.

That post asked who owes whom inside the United States. This one asks the next question: who outside the United States holds the AI boom, and how much of that can official statistics actually see? I am writing the answer up as a working paper. What follows is a preview of its main results, which are preliminary and may change.

1. Why the question matters

A correction in AI valuations would not stop at the U.S. border. Foreign investors have piled into U.S. equities since late 2022, and the firms at the center of the AI boom now make up a large part of what they hold. If those valuations fell sharply, the losses would land on pension funds in Canada, insurers in Japan, sovereign funds in the Gulf and households across Europe.

Policy makers abroad therefore need to know three things: how much of the AI boom their residents hold, through which channels, and how much of that they can see in their own data rather than infer.

2. How I measured it

The exercise has four steps. All of it uses public data, at end-2025 prices.

Define the AI firms. I take 74 listed firms across the “AI value stack”, from chip designers and hardware makers to cloud providers, energy suppliers, builders and applications, following a recent classification by Chen, Gautam and Wu (2026). I add the project-finance vehicles from my network map. At end-2025, these firms were 37 percent of the U.S. stock market by value: $25.7 trillion out of $68.9 trillion.
Find the holders. About 80 economies report how much they hold of each other’s stocks and bonds in an international survey, the Portfolio Investment Positions survey (formerly CPIS). It tells you that Japan holds so many dollars of U.S. equity, but not which companies.
Fill the gap, then test the fill. The standard approach assumes investors hold the market: an economy with $100 of U.S. company shares holds about $37 of AI firms. I test that assumption against holdings that investors disclose company by company (section 4).
Look through the funds. Investment funds in Ireland and Luxembourg and vehicles in the Cayman Islands hold large amounts of AI shares on behalf of investors elsewhere. I pass those holdings through to the economies that own the fund shares, as far as the data allow.

The result is a split of every dollar into four parts: what is observed directly, what is modelled, what cannot be allocated to any economy, and what sits in economies that do not report at all.

3. The headline: $7.8–9.9 trillion, mostly in shares

At end-2025, cross-border holdings of AI firms’ securities came to $7.8–9.9 trillion. Most of it is investors outside the United States holding U.S. AI firms; about $0.7–1.0 trillion is U.S. investors holding AI firms based elsewhere, such as chipmakers in Asia and Europe. The total is close to Van Nieuwerburgh’s $9 trillion, but it measures something different: his figure is investment to be spent through 2032, while mine is the market value of cross-border holdings of the firms at one point in time.

The more important finding is the composition. Cross-border exposure to the AI boom is almost entirely equity. AI firms are 37 percent of the U.S. stock market but only 1.7 percent of U.S. debt securities. A correction would hit foreign holders through share prices, not through bond defaults.

That brings the invisible balance sheet back into view, this time from abroad. The credit that does finance the buildout, in project bonds sold as private placements, private credit and leases, barely shows up in cross-border statistics. Take Meta’s Hyperion campus in Louisiana: its $27.3 billion of project bonds can be traced only to the extent that U.S. registered funds report them, about a third of the issue. Who holds the rest, at home or abroad, is not visible in any public source.

Figure 1. Cross-border holdings of AI firms at end-2025, by what can be seen. Low and high estimates. Observed: holdings disclosed security by security. Modelled: estimated by proportional allocation, holder economy known. Unallocated: located in a financial center but not traced to an owner. Not covered: economies that did not report. Source: author’s calculations from public data.

4. Investors hold AI close to market weights, once you count correctly

The whole estimate leans on one assumption: that holders own AI firms in proportion to the market. So I tested it against investors who disclose their holdings company by company.

Over 500 foreign institutions. Large foreign institutions that file Form 13F with the U.S. Securities and Exchange Commission list their U.S.-listed shares. Holding half of them back, the “hold the market” estimate predicts their AI holdings with a value-weighted error of 14 percent.
Country totals. Grouped by country, the institution-level differences largely cancel out. Across 19 countries, holdings depart from market weights by 4.4 percent on average; the United Kingdom and Canada are at 0.97 and 0.99 (Figure 2).
Asset owners. A manager’s location is not where its clients live, so I also looked at pension funds, insurers, sovereign funds and central banks that file for themselves. Weighted by value, they hold AI firms at 1.08 times market weight: slightly overweight, but within the range the estimates carry.

Bar chart: AI holdings of foreign 13F filers relative to market weight, by country

Figure 2. AI holdings of foreign 13F filers relative to market weight, by country. A value of 1.0 means a country’s filers hold AI firms exactly in proportion to the market. Shaded band: the ±14 percent range applied to the estimates. Countries with at least five filers; country is where the manager is located. Source: SEC Form 13F data, fourth quarter of 2025; author’s calculations.

One correction mattered more than any test. The statistics count shares in U.S. investment funds, including money market and bond funds, as “equity”, alongside company shares. About a quarter of foreign holdings of U.S. equity is not company stock. Counting fund shares at their actual AI content lowers the AI share of foreign holdings of U.S. equity from 37 percent to about 32 percent, and the headline total by roughly a tenth.

The correction also resolves a puzzle. Federal Reserve researchers found that in mid-2022 the “Magnificent 7” were 16.2 percent of foreign holdings of U.S. equity, against 20.0 percent of the market, which suggests foreigners underweight the largest tech firms. But their denominator includes fund shares. If those holdings are company shares, the ratio against company shares alone is about 1.09, close to market weight. That is an illustration rather than a direct comparison, but it is consistent with what my tests find.

5. Where the exposure sits

Outside the United States, the largest holders are Canada, Japan, the United Kingdom and Germany, each with roughly $0.4–0.8 trillion. U.S. investors are large cross-border holders too, but of non-U.S. AI firms such as Asian and European chipmakers: about $0.7–1.0 trillion. Relative to the size of the economy, the picture changes: Norway’s exposure is about 72 percent of GDP, Hong Kong SAR’s 60–68 percent, and Sweden’s, Canada’s and Denmark’s roughly a quarter to a third.

Funds reshuffle the ranking. Once holdings in Irish and Luxembourg funds are passed through to their owners, Hong Kong SAR rises 33 places in the ranking by AI share of portfolio, and it does so under every variant I tested. Several euro area economies rise by 13 to 21 places (Figure 3), but those moves shrink when independent estimates of who holds Irish and Luxembourg fund shares replace the survey’s figures: Italy’s rise falls from 18 places to 6. Who owns the funds is exactly what the statistics do not show. Italy receives about 60 percent of its AI exposure through funds in financial centers rather than by holding the shares directly.

Figure 3. Looking through funds reshuffles the ranking. Rank of each economy by the AI share of its portfolio, before and after holdings in Irish, Luxembourg and Cayman funds are passed through to their owners; the 12 largest movers among 74 economies. Risers in navy. Fallers (grey) have little or no fund holdings: they move down because fund-heavy economies rise past them. Hong Kong SAR’s rise survives every variant tested; the euro area rises are smaller with independent fund-holder estimates. Source: author’s calculations.

6. A fifth of the exposure cannot be traced to anyone

This is the finding I think matters most. Only 7–9 percent of the total ($0.72 trillion) is observed security by security, from a sovereign fund’s published holdings, the 13F filings of about 45 pension funds, insurers, central banks and other asset owners, and U.S. holdings of Chinese AI firms’ U.S.-listed shares. Everything else is estimated.

Worse, 21–22 percent ($1.7–2.1 trillion) cannot be attributed to any economy at all (Figure 4):

$1.1–1.3 trillion sits in the portfolios of funds in Ireland, Luxembourg and the Cayman Islands whose owners the survey does not reach.
$0.44 trillion is the estimated U.S. share of Chinese AI firms incorporated in the Cayman Islands, mostly traded in Hong Kong, beyond what U.S. filings show; who actually holds it cannot be pinned down.
$0.25–0.32 trillion is held by economies that did not report for end-2025, such as the United Arab Emirates, the British Virgin Islands and Taiwan Province of China.

Figure 4. From holders to AI firms: where the exposure sits and what can be seen. Midpoints of the low and high estimates at end-2025 (about $8.9 trillion in total); positions, not flows. Left: who holds; middle: held directly or through funds in Ireland, Luxembourg and the Cayman Islands; right: the AI firms. Orange is exposure with no identifiable owner. The AI content of modelled links is estimated, so the diagram shows the structure of the estimate rather than observed holdings. *The United States counts only its holdings of non-U.S. AI firms. Source: author’s calculations from public data.

Part of that is a choice. The $0.44 trillion is a model-based estimate I chose not to assign to U.S. investors without evidence. Counting it as theirs would cut the untraceable share to about 17 percent, which is the part due purely to missing statistics.

Since holdings look close to market weights, the allocation rule is not the binding constraint. Attribution is. We can estimate how much AI exposure there is reasonably well; we cannot say who holds a fifth of it.

7. What a correction would cost

As an illustration, suppose AI shares fell 35 percent, close to the Nasdaq-100’s 33 percent fall in 2022, and AI bond spreads widened by 2 percentage points. First-round losses to cross-border holders would be $2.7–3.4 trillion, of which $0.6–0.7 trillion would fall on holders that cannot be identified. These are mark-to-market losses only; they ignore hedges, the gains made before end-2025 and any knock-on effects.

8. What would help

In my earlier post I argued for measurement before capital rules. The same logic applies across borders, and the gaps can be ranked by how much they would reveal:

Statistics on who holds fund shares in financial centers. This alone would allocate $1.1–1.3 trillion of the untraced exposure.
Wider reporting, with detail on holder sectors. Every economy that reports, and says whether its holders are pension funds, insurers, banks or households, closes part of the rest.
Security-level data that identify the final holder. This would turn the modelled part into observation. But data that stop at the custodian, the nominee or the fund identify securities, not owners; they move the problem rather than solve it.
Better reference data. Company identifiers that lag mergers, or link a subsidiary’s shares to the wrong parent, decide whether a holding is attributed to the right firm.

9. Caveats

The results are preliminary and use public data only. They are a snapshot at end-2025; AI valuations have moved a lot since, so current exposure is likely higher. The test of “holding the market” rests on large institutions’ U.S.-listed holdings; households and smaller investors are not tested. The list of AI firms follows one published classification, and a broader list would add some non-U.S. firms. It also counts the full value of diversified giants such as Apple, Microsoft and Amazon, so the totals measure holdings of these firms’ securities, not the part of their value that depends on AI. And the paper measures who holds the securities, not who ultimately bears the risk once hedges and guarantees are counted.

Bottom line

My earlier post found that the domestic financing of the AI buildout is hard to see from outside. Following the exposure across borders gives a similar answer with a twist. The standard way of estimating who holds what works reasonably well; what is missing is the ability to name the holder. A fifth of the cross-border exposure to the AI boom, roughly $2 trillion, belongs to no one in the statistics.

I am finishing a working paper with the full method, the tests and the data gaps, and I would welcome comments.