China Telecom's research arm puts inference at 80% of China's compute market by 2029

Training and inference both buy compute, and they land on different budget lines. A training run has an end date. Inference is metered by use, so it recurs for as long as anyone keeps using what you shipped; thenextweb calls it an operating expense, because it is what a model costs every time somebody uses it ^[13]. The China Telecom Research Institute puts that metered share at 80% of China's computing-power market by 2029 ^[1], and attributes the growth in volume to agent deployment ^[2].

The scale of the recurring side is already visible in Chinese spending. Technology companies there are on course for close to 600 billion yuan, about $89bn, on AI this year ^[3], which the report puts at more than a tenth of all investment in the country ^[4]. Divided over 365 days, that is roughly $244m a day ^[15]. Europe's committed EUR 1bn is about 3% of the EUR 30bn headline ^[16], and thenextweb notes that Chinese technology companies spend EUR 1bn about every five days ^[8].

The European schedule has moved before. Bidding slipped from May to July, the evaluation criteria were delayed more than once, and the field fell from about 70 interested companies to roughly ten expected bidders ^[9]. For up to seven sites, that is about 1.4 bidders per site ^[17]. On paper, thenextweb notes, the mid-2028 start date arrives a year before China's crossover, and that assumes nothing else slips ^[18].

Most of the public half of the programme depends on an EU budget for 2028 to 2035 that member states have not agreed ^[10]. The account does not say how gigafactory capacity would be split between training and inference. The forecast on the other side comes from the research arm of a state-owned carrier ^[11] and was carried by CCTV on Saturday ^[12]. It is an expectation for 2029.

For anyone planning a budget against this, spend that ends and spend that repeats per use behave differently. The first can go to a capital committee once and be finished. The second needs a unit price and a named owner who answers when the invoice grows in a month with no release to explain it. Europe's EUR 30bn buys gigafactories and the machines inside them ^[5]. The 80% in the forecast is what it costs to run models after they are built ^[1][13].