31 AUG 2026 — Moody's finds Chinese hyperscaler capital spending more than doubling, from US$65bn in 2025 to a projected US$140bn in 2026 and US$165bn in 2027. Growing faster than the Americans, the Chinese firms will still fall further behind in absolute terms every year.
The arithmetic that makes this counter-intuitive
Six major American hyperscalers are projected to spend more than US$785bn in 2026 and to approach US$1tn in 2027, roughly six times the Chinese total.
The Chinese cohort grows at a higher rate, and it does so from a much smaller base. Moody's states the consequence directly: even though China's growth rate is faster, the absolute gap in total data centre capacity will still widen over the forecast period. On these numbers the gap goes from about US$645bn in 2026 to roughly US$835bn in 2027.
This is the part that gets reported backwards. A doubling is a more dramatic headline than a 27 per cent increase, and the doubling belongs to the side that is losing ground. Percentage growth and absolute position point in opposite directions here, and capacity is bought in dollars.
Who is spending what
Alibaba committed RMB 380bn, about US$57bn, over three years in an announcement dating to February 2025, and posted quarterly capital expenditure of RMB 67.7bn in the second quarter of 2026. ByteDance is reported at up to US$70bn for 2026. Tencent spent RMB 52.8bn in the quarter, up 176 per cent year on year, and Baidu RMB 11.4bn, nearly tripled.
Those growth rates are extraordinary and they are measured against a 2025 base that was low by American standards. Tencent's spending tripled, and the absolute increase is smaller than a single quarter's growth at one of the larger US buyers.
Money is not the binding constraint
Moody's identifies access to advanced hardware as China's biggest constraint, with export restrictions pushing firms toward domestic alternatives that still lag Nvidia's technology and its ecosystem.
That reframes what the spending figures mean. A capital expenditure comparison implicitly assumes both sides are buying the same goods at the same prices, and they are not. If a domestic accelerator delivers materially less useful throughput per unit of capital, then the effective gap is wider than the dollar gap, and the dollar gap is already six to one.
The word ecosystem carries weight here. The hardware deficit is measurable and closing slowly. The software deficit — CUDA, the kernels, the frameworks that assume it, and the accumulated engineering practice of a decade — is the part that does not close by building a better chip, and it is the reason a domestic part at parity on paper still underperforms in a real training run.
The finding that applies to both sides
Moody's expects both groups may turn free-cash-flow negative in 2026 and 2027, with cash reserves sufficient to absorb 12 to 24 months of elevated capital expenditure without material deterioration in credit quality.
The analysis's most consequential sentence is not about China at all. It says the largest and most profitable technology companies in the world are spending more than they generate, and that the buffer is measured in one to two years rather than in a decade.
A ratings agency's 12-to-24-month horizon is not a prediction of distress. It marks the point at which balance sheets matter less than whether the returns on this spending have arrived. Nothing in the analysis says they will not. It says the timer is now visible, on both sides of the export-control line.
Why this matters from Southeast Asia
A meaningful share of both cohorts' capacity is being built in the region, which turns the spending forecast into a local construction forecast.
American hyperscalers are building in Singapore, Johor and Jakarta, and Chinese providers have been expanding across the same markets. If the aggregate figure approaching US$1tn is even roughly right, the constraint on how much of it lands here is not capital. It is power and grid connection, which is why Malaysia's pipeline is announcements rather than connections and why Johor has been turning projects away.
The free-cash-flow observation has a regional edge too. A hyperscaler with a one-to-two-year buffer is a hyperscaler that will eventually scrutinise which regions justify continued build-out, and markets competing on incentives rather than on power availability are the ones with the weakest case when that scrutiny arrives.
How to read the next set of numbers
Quarterly capital expenditure announcements will keep producing large percentage figures from the Chinese firms, and those figures will keep being reported as catching up.
The check is simple: convert to absolute dollars and compare the change, not the rate. A 176 per cent rise on RMB 19bn is a smaller addition to global compute than a 15 per cent rise on US$80bn. And where a domestic accelerator is being bought instead of an Nvidia part, the dollars themselves overstate what was acquired.
