China's National Compute: 24GW and a Single Tenant

In 1847 London suspended its own credit rules

In 1846 British private capital financed thousands of miles of railway that still lacked passengers.

The mechanism was straightforward: capacity first, revenue later, debt in between. In October 1847 the government suspended the Bank Charter Act to keep commercial credit alive. The rails stayed where they were; the shareholders vanished.

In January 2002 Global Crossing took its books to court, having laid fibre for demand expected over the following decade. WorldCom collapsed in July of the same year.

A third precedent, closer to us: in October 2018 Sears filed for Chapter 11. Hundreds of American shopping malls lost, in one stroke, the tenant that brought them their footfall.

Three precedents are enough to call it a pattern. Capacity built ahead of time, debt financing it, demand concentrated in a very small number of signatures.

The real measure: more than 24GW already delivered

The debate over sovereign AI and national compute proceeds by model names. The contest is measured in gigawatts.

The Chinese data centre model published on 25 September 2026 maps more than 1,000 facilities across over 60 operators and arrives at a Chinese fleet above 24GW^[1]. Larger than EMEA, estimated at around 14GW. Larger than the rest of Asia, at around 15GW.

The United States remains ahead, with 56GW expected by the end of 2026. The Chinese figure excludes roughly 20GW of dated pipeline and a further 30GW of announced projects.

Published estimates of this capacity diverge by as much as 15 times. Two clichés have filled the vacuum: China is enormous, China is empty. The first holds up; the second deserves evidence.

A retail market that AI turned on its head

Structure matters more than scale.

These facilities grew up as retail colocation: racks and kilowatts leased to banks, telecom operators and e-commerce platforms, on short contracts with a thousand counterparties. Fragmented demand, diffuse risk, modest margins.

AI upended that model in a handful of quarters. Demand now arrives in wholesale blocks of tens of megawatts, with power and cooling requirements the legacy estate can barely carry.

Deliveries are running fast: a hundred megawatts in twelve months at a single operator. Whoever owns the property goes from a thousand small customers to two or three enormous ones. Revenue improves; the quality of the risk gets worse.

The tenant that escapes public accounts

According to the same work, ByteDance occupies roughly a fifth of delivered capacity in China, and leases almost all of it.

ByteDance remains a private company. Zero 10-Ks and zero quarterly reports. Zero covenants legible from the outside. The sector's most important customer is also its least documented.

Its Doubao app serves 345 million users a month, and is the Chinese ChatGPT.

GDS and VNET, the only two Chinese landlords listed in the United States, signed 1.3GW of wholesale orders in the first half of 2026. Those two combined captured barely a third of the orders placed by ByteDance and Alibaba between 2024 and today. The rest sits with operators that never listed and with the state carriers.

Anyone reading the two listed stocks is observing a third of the market, and believes they are seeing all of it.

The capex that switched off free cash flow

In the second quarter of 2026 the combined capex of Alibaba, Tencent and Baidu hit 20 billion dollars, more than double year on year.

For the first time in the available data, all three closed a quarter with negative free cash flow. It is the largest capex jump in the sector's history. The figure excludes ByteDance, which spends more than any of them.

The causal mechanism is this: capacity is paid for now, in cash and in debt, while AI revenue arrives later. Every quarter of negative flow shifts the financing towards banks, bonds and private credit.

The BIS and the IMF have been converging on the same point for months. The risk in the AI cycle lives in the financing structure, rather than in equity multiples. This is a regime change in credit, rather than a property cycle.

Here sovereignty is spelled in power, land and credit

The word sovereignty acquires physical content. The "Eastern Data Western Compute" programme moves computation towards the western provinces, where cheap electricity and land are.

The BBC reported from the ground, in Inner Mongolia, on the scale of these complexes. The report describes facilities rising in the desert^[2], next to their own power source.

American export controls hit advanced hardware. The foundations, the substations and the sheds went up anyway, while Washington was counting chips.

The chokepoint remains the semiconductor fab. The hole remains finance: capital routed through Hong Kong keeps reaching Chinese labs. A sanction stops a machine; stopping a wire transfer takes something else.

My position, and what would change it

This desk's thesis: the Chinese boom is a property market flipped by AI, rather than sovereignty planned from above.

The capacity exists, the scale is verified, state planning is real in the western provinces. The risk lives elsewhere: with whoever lent the money, and in the number of tenants signing the cheques. A fifth of national capacity in the hands of one private company is textbook concentration.

What would change this reading: multi-year take-or-pay contracts disclosed by the landlords, with at least three counterparties per site and a verifiable average utilisation above 80%. With that data the structure becomes stable income, rather than leverage in disguise.

The market has priced Chinese scale. It has ignored the contractual structure holding it up.

Three implications for capital

Each implication carries its own horizon, because duration changes the decision.

1. Family offices and sovereign funds (36 months): exposure to Chinese compute runs through property, the power grid and local debt, rather than through the internet stocks listed in New York.
2. Debt before equity (24 months): the loan curve for Chinese landlords signals the change of tempo well before share prices do.
3. Tenant concentration (12 to 36 months): ask for the figure site by site, or apply a discount to the asset value.

For a CEO the point is different. The strategic plan covers chip availability, and almost never covers the solvency of whoever owns the shed where the model runs.

For a chief risk officer, the scenario missing from the VAR models is the renegotiation of a wholesale contract by a private tenant that publishes zero accounts. For a CFO the narrative due for revision is a single sentence: China is behind on compute. In eighteen months that sentence ages badly in front of an investor reading the data facility by facility.

The prediction

By 31 December 2027 the combined capex of Alibaba, Tencent and Baidu will record a year-on-year contraction in at least two consecutive quarters, according to their public accounts.

Confidence: 60%. Horizon: 31 December 2027. Verification: the three companies' quarterly reports. Kill signal: combined capex at the three grows year on year in each of the four quarters of 2027.

Market indicator: GDS down within the horizon.

What to watch

Three indicators anticipate the outcome, and all three are public.

• The customer concentration notes in the GDS and VNET filings: a single tenant above 25% of revenue is the signal.
• Bond issuance and syndicated loans from Chinese landlords, read by maturity and by spread.
• The ratio between delivered capacity and capacity in use in the western provinces.

The first of those three data points is worth more than any aggregate estimate of national capacity. The divergence between capacity built and revenue under contract always resolves. The question is how.