On 12 August, Tencent reported a second-quarter result that split the market. Revenue reached 204.785 billion yuan, up 11 per cent and past 200 billion for the first time, but capital expenditure hit 52.784 billion yuan, up 176 per cent year on year and 65 per cent quarter on quarter, far above the 32.1 billion the market expected. Free cash flow turned negative at 13.8 billion yuan. That 52.8 billion is what Tencent spent on data centres, GPU servers and intelligence-compute clusters.
Roughly calculated, Q2 capex was nearly triple the 19.1 billion of a year earlier; the first half reached 84.72 billion, already above the 79.2 billion for all of 2025. The negative free cash flow is not operational bleeding: Tencent notes large AI-related prepayments sit inside operating cash flow, and stripping those out, Q2 free cash flow was 37.6 billion. On that basis single-quarter AI compute prepayments hit 51.4 billion. Helped by 59.3 billion of capex payments and 41.6 billion of dividends, net cash fell from 146.86 billion at end-March to 58.191 billion at end-June, a 60 per cent drop in three months.
The four giants’ arms race
Tencent is not alone. ByteDance is the most aggressive: Bloomberg said on 27 May it is discussing raising 2026 capex to as high as 70 billion dollars, about 475.9 billion yuan, from profit reserves near 50 billion dollars, though a source puts the real range at 400 to 500 billion yuan and the 70 billion is a ceiling, not a baseline. Alibaba is the most transparent: in February CEO Eddie Wu promised over 380 billion yuan in cloud and AI infrastructure over three years, then said spend “will far exceed” that plan. Baidu’s capex is smaller but explosive, up to 5.839 billion in Q1, with AI revenue of 13.6 billion, first time beating traditional ads. Sell-side now models Tencent’s full-year capex at 250 billion, up from 170 billion at the start of the year, though that is forecast, not guidance.
“Not one card sits empty”
Demand is the core. CEO Martin Lau said prepayments and compute orders paid months ago “could be sold today at over 30 per cent profit”, a seller’s market. Wu was blunter: almost no card inside the servers sits idle. The China Academy of Information and Communications Technology says Q1 AI compute demand jumped 417 per cent year on year while supply grew only 128 per cent; by end-June national intelligent compute reached 2,185 EFLOPS, up about 37 per cent in six months. The Ministry says effective high-end training supply meets under 47 per cent of demand, a gap near 35 per cent. H100 lease prices hit 2.35 dollars an hour in March, up nearly 40 per cent, with new contracts slipping to early 2027.
Yet the rack rate is 71.4 per cent, meaning nearly 30 per cent of installed compute sits unused, and while high-end training is scarce, some mid-low clusters are underused. The seller’s market is in high-end GPU clusters, not compute overall.
The return account on AI spend
At first glance capex ate the profit. Q2 Non-IFRS operating profit was 75.636 billion, up 9 per cent; strip out new AI products, Yuanbao, CodeBuddy, WorkBuddy and WeChat’s AI, and it was 86.1 billion, up 19 per cent. The 10.5 billion gap is AI’s real bite on current profit. But AI also lifted ad growth 22 per cent and kept marketing services above 20 per cent; WorkBuddy leads the domestic market with over 20 million monthly visits. Lau splits capex into steady base business and a one-off AI-native surge he says should not be assumed every year. Goldman compares this stage to Meta a quarter or two ago: short-term pain, long-term ticket.
Risks under the feast
Three risks deserve a straight look. One, the free-cash-flow squeeze is structural, Tencent cash down 60 per cent, Alibaba operating loss and negative free cash flow, ByteDance profit down over 70 per cent. Two, headline ceilings are not actual spend, ByteDance’s 70 billion dollars and Tencent’s 250 billion are plans and forecasts that can move sharply. Three, scarcity and idle coexist, so returns on the compute chain will diverge, not lift evenly. For firms that find a real order inside the giants’ budgets, it is a historic chance. For those chasing the theme blindly, the prisoner’s dilemma, everyone dares not stop, everyone bets a rival breaks first, means industry returns may stay pressured. The feast has started, but the bill arrives on schedule.
Editor’s note: This is an adapted translation of the original Sohu report. It has been trimmed and restructured for readability for an international business audience.