The tabs stayed open all afternoon and none of them agreed. One said the chips had won. The next said the chips did not matter. A third said the market had already priced the whole argument and moved on.
The reader shut the laptop with a familiar feeling: every number was properly sourced, and the story still did not add up.
That is not a failure of the sources. It is a failure of category. Most headline figures about China's technology build-out describe capacity, shipments or downloads, while the questions readers actually care about are revenue, utilisation and capability. The numbers are answers to different questions.
Key takeaways
- Data-centre capacity is forecast above 24 GW by end-2026, against roughly 56 GW in the US (SemiAnalysis, 25 September 2026).
- Huawei targets 1.6 million Ascend dies in 2026, while memory output supports only 250,000–400,000 packages (SemiAnalysis).
- China averages 140 trillion tokens a day (National Data Administration), with no comparable revenue figure published.
- Huawei holds ~50% of China's accelerator market (Bernstein, cited by AP) yet roughly 1–5% of global AI compute (算力), per the Council on Foreign Relations.
- The four questions: capacity or utilisation, announced or shipped, cumulative or active, disclosed or audited.
Capacity is not compute
Start with the gigawatts. China's data-centre capacity is forecast to exceed 24 GW by the end of 2026, against about 56 GW in the United States, 14 GW in Europe, the Middle East and Africa, and 15 GW in Asia-Pacific excluding China (SemiAnalysis). It is a striking number and a weak one: a gigawatt measures what has been built, not what is being used.
An idle campus still draws power — China's compute-centre electricity demand is forecast at about 300 TWh in 2026 (Beijing Daily) — but it does not produce tokens, revenue or capability. The interesting comparison is not 24 against 56. It is 24 against however many gigawatts are actually serving traffic, and that figure is not published.
The first question: capacity or utilisation? If a source cannot answer it, the number describes a construction programme, not an industry.
Announced is not shipped
The second trap is the gap between a target and a package. Huawei's plan for 1.6 million Ascend dies in 2026 is nearly triple the previous year, and it is a plausible manufacturing target (Bernstein Research, cited by AP). But CXMT's 2026 output of about 2.2 million high-bandwidth memory stacks supports only 250,000–400,000 Ascend-class packages, while SMIC can make die for more than a million (SemiAnalysis).
Two true statements — a die target and a memory supply — produce a third number nobody announced. The gap is the number that matters.
The same logic applies to market share. Huawei's Ascend holds around 50% of China's AI accelerator market in 2026, up from about 40% (Bernstein, cited by AP). That is a share of a constrained domestic market. Measured against aggregate global AI compute, the Council on Foreign Relations estimates the most Huawei-favourable case at about 5% of Nvidia's output, with a median near 1%. Both figures are correct. Only one is about capability.
Cumulative is not active
Downloads are the softest currency in the sector. Alibaba's Qwen family passed 1 billion cumulative Hugging Face downloads in mid-2026 (Hugging Face counters; company statements), which is a real measure of developer attention and a poor measure of deployment. A download can be a production dependency, a benchmark run, a fine-tune that never shipped, or a mirror.
Nothing published converts that billion into paid inference. Treat it as evidence that a model is widely tried, not that it is widely used.
The same discipline applies to company counts. China has more than 6,600 AI companies, about 15% of the world total (China Economic Net), inside an industry CAICT valued at RMB 1.2 trillion (≈ US$169bn / HK$1.32tn) in 2025, up 40% year on year. A count of firms and a revenue estimate are different objects; neither tells you how many firms have revenue.
Disclosed is not audited
Most of the numbers that travel best are disclosed by the party that benefits from them. United Imaging Intelligence's chest agent reports 97.8% accuracy across roughly 100 abnormalities and more than 2.5 million cases (company disclosure). MIIT reports efficiency up 29%, defect rate down 47% and R&D cycle down 38% at 15 lighthouse factories (领航级智能工厂). These may all be accurate. None is an independent audit, and none ships with its baseline.
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the common error is not scepticism but category error: a capacity figure gets read as an output figure, and the two diverge most sharply at the top of a build-out.
The one-page checklist
- Capacity or utilisation? Gigawatts, factories and clusters describe supply. Ask what share is serving paid work.
- Announced or shipped? Die targets, orders and memoranda are intentions. Ask for packaged units and delivery dates.
- Cumulative or active? Downloads, registrations and MAU accumulate. Ask for the period and the paying subset.
- Disclosed or audited? Company accuracy claims and ministry aggregates are disclosures. Ask for the baseline, the definition and the auditor.
Four questions will not resolve every number. They will tell you which numbers are worth arguing about.
Honest limitations
These questions expose category errors but cannot fill the gaps they reveal: utilisation, vendor revenue and paid deployments are simply not published for most of the Chinese market, so the honest answer is often "unknown" rather than "overstated". The figures used here are a snapshot of public reporting as of October 2026, and company disclosures may be revised. Currency conversions use ¥7.1 to the dollar and can mislead across a year of movement. Nothing here establishes that any specific claim is false — only that it measures something narrower than the headline implies.
What readers can do now
- Operators: convert every supplier claim into units you can verify on site — installed seats, paid licences, delivered packages.
- Investors: build a category column beside every number in your model, and refuse to compare across categories.
- Researchers: publish the denominator alongside the figure; the missing denominator is where most of the confusion lives.