Two decades ago, Loudoun County in Virginia, just outside Washington, was one of the most frenzied centers of America's housing bubble, before it was battered by the ebbing of global savings and the 2008 subprime defaults.

Twenty years later, the area is blanketed by sprawling data centers, with global capital from sources such as Canadian pension funds and Japan's SoftBank pouring in to build AI infrastructure, making it one of the world's densest hubs for internet connectivity and AI computing power.

As of fiscal 2026, more than 40% of the county's tax revenue — $1.2 billion — comes directly from data centers.

This extreme reliance on a single asset has revived fears among local officials of the eve of the 2008 crisis — Loudoun County has not only recently halted approvals for new data centers, but also set up a dedicated $120 million "fiscal stabilization fund" and required operators to post liquidation collateral, to prevent a repeat of the fiscal collapse of that era should the computing power bubble burst.

Loudoun County's anxiety is a microcosm of the broader global macroeconomic imbalance.

How international savings are being reallocated: from buying Treasuries to betting on US AI stocks

This time, the chain of global capital flows has undergone a key restructuring.

The channel has shifted from official to private: in the early 2000s, overseas savings from Asia and elsewhere were mainly managed by foreign central banks, which bought large amounts of US Treasuries and in effect pushed down global borrowing costs; today, vast sums are held more by overseas private funds chasing high returns, flowing directly into US equities and technology corporate debt.

Risk exposure is highly concentrated: analysis by Bloomberg Economics shows that as of the end of June 2026, overseas investors' exposure to core AI-related US equity sectors had reached about $4.9 trillion — roughly equivalent to Germany's annual GDP — with its share of their total US equity holdings surging from about 10% in 2019 to nearly 20%.

Hung Tran, a senior fellow at the Atlantic Council, commented: "This is like the Middle East petrodollars of the 1970s. Asian tech companies earn dollar profits and then recycle them back into the US market. Today's tech dollars have become the new version of petrodollars."

Frederic Neumann, chief Asia-Pacific economist at HSBC Holdings, noted: "Asian capital is precisely the fuel feeding America's AI boom. It is entirely built on fragile global demand and capital flow imbalances. Once AI investment stumbles, the US, the dollar and global markets will all face a painful adjustment."

Both "dot-com" and "real estate": the structural mutation of AI assets

Why can AI absorb such enormous cross-border capital? In its research report "American AI in the 'Marshmallow' Era," CICC put forward a core logic: the US AI industry simultaneously possesses hardcore technology and "real-estate-like" dual attributes.

On the technology dimension, AI's per-capita efficiency and ability to replace labor significantly surpass the internet companies of that era.

Data shows that since ChatGPT's release in 2022, nonfarm employment in the US telecommunications and computer industries has fallen 6%; during the 1995-2000 dot-com bubble, employment in that sector surged more than 60%. Measured by the ratio of employees to fixed assets as a gauge of labor density, the labor density index for Google, Microsoft and Meta plunged from 100 in 2021 to 37 in 2025.

But from a balance sheet perspective, AI looks more like a heavy-asset investment game.

CICC statistics show that as of the second quarter of 2026, fixed assets as a share of total assets at the five leading players — Microsoft, Alphabet, Amazon, Meta and Oracle — had averaged more than 40% (versus 37% for traditional heavy-asset industrial companies). At the peak of the dot-com bubble in 2000, tech companies such as Microsoft and Cisco had fixed assets averaging just 15% of the total.

Stijn Van Nieuwerburgh, an economist at Columbia Business School, calculates that by 2032 cumulative global investment in AI infrastructure will reach nearly $9 trillion. As a share of GDP, that would surpass the 19th-century railway boom and become the largest capital expenditure expansion in US history.

Massive fixed-asset investment gives AI companies a natural "collateral attribute" similar to real estate, opening up a credit financing logic entirely different from the dot-com era.

Debt and securitization surge: goodbye VC equity, a replica of pre-subprime financial engineering

Asset attributes determine financing models.

CICC's report notes that the 1990s dot-com bubble relied mainly on VC and stock IPOs (equity financing); today's AI investment, having accumulated vast data centers, land and equipment, is rapidly moving toward the debt expansion and structured finance model of the 2000s real estate bubble.

Data shows that in the second quarter of 2026, the ratio of net cash flow from financing to revenue for representative AI companies rose to 16.2%, of which net debt financing accounted for as much as 21.2%, while equity financing was actually in a state of net buybacks.

This is reflected in the establishment of three financing pillars:

Massive corporate bond issuance: Dealogic data shows that in just the first six months of 2026, the five AI giants issued $159 billion of bonds globally, compared with just $17 billion for all of 2024.

Off-balance-sheet debt and private credit expansion: Bloomberg statistics show the five giants have signed a cumulative nearly $3 trillion in long-term leases, purchase commitments and guarantees not reflected on their balance sheets. The Bank for International Settlements estimates that direct loans from large private credit institutions to AI-related companies have exceeded $200 billion, nearly 8% of their total, effectively replacing the lending role of traditional banks.

Computing power asset securitization (ABS): similar to the MBS of that era, computing power leases are being packaged as ABS assets. After CyrusOne, Sabey, Stack and QTS, Compass Datacenters issued $830 million of data center ABS in early 2026, with the underlying assets directly tied to cloud providers' long-term computing power leases and distributed to pension and insurance funds seeking fixed income.

Comparing with the laws of past bubble bursts: where are the cracks appearing?

Looking back at history, CICC's report notes that the final collapse of both the dot-com and real estate bubbles was accompanied by the simultaneous appearance of three phenomena: high macro and industry leverage, deteriorating fundamentals, and a systemic tightening of financing conditions.

At present, AI companies' interest coverage ratios remain high (about 82 times for Google, about 7 times for Oracle), and nominal industry earnings are still growing, so the conditions for a systemic bubble burst are not yet fully mature.

David Chao, global market strategist at Invesco, noted: "The technology is real, the applications are real, and many leading companies have extremely high profits. What I worry about is that investors' expectations will soon become impossible to satisfy."

But marginal cracks at the short-term funding level cannot be ignored.

Bond issuance costs are rising fast: in July 2026, the coupon on Meta's $12 billion bond issue reached 7.5%, 0.4 percentage points above its issuance cost at the end of 2025; yields on some high-yield bonds of computing power operator CoreWeave rose more than 0.5 percentage points after the Fed's September rate hike; and S&P recently downgraded Oracle to BBB-.

Government fiscal buffers are running out: Pierre-Olivier Gourinchas, former chief economist of the IMF, said bluntly: "Even a technological revolution is not a blank check with unlimited credit." He warned that during the dot-com bubble, government public debt as a share of GDP was only one-third to one-half of today's level; now, America's more than $40 trillion of debt and persistently rising Treasury yields have greatly weakened the fiscal space to respond to a crisis.

Risk of private institutions exiting: economist Gian Maria Milesi-Ferretti stressed that a large number of hedge funds are using Treasuries as collateral to add leverage in the AI game: "These institutions adjust positions extremely quickly. The risk now is not foreign central banks selling Treasuries, but these highly yield-seeking private institutions being forced to dump assets to cover losses elsewhere."

From the physical data centers whose approvals have been paused in Loudoun County to the layer upon layer of computing power ABS on Wall Street, this money game — blending technology valuations with a real estate debt model — is pulling global capital into an unprecedented network of resonance.