Alibaba and Tencent spent $18bn in a quarter, and some of it bought price, not compute

Wu Yongming put two numbers on the same conference call, and they only work together in one direction. Alibaba, he said, can recoup what it spends on compute inside Alibaba Cloud within two and a half to three years ^[6]. The report carrying that is dated 22 August and covers the June quarter ^[7], which puts payback between late 2028 and the middle of 2029 ^[5], comfortably inside the shortage he describes ^[5]. Scarcity is not the risk being managed in that model. It is the assumption the return rests on.

Against that horizon, the standing budget looks stale. Alibaba said last year it would put at least 380 billion yuan into AI and cloud infrastructure over three years ^[8]. Spread evenly, that is about 31.7 billion yuan a quarter ^[2]; the June quarter ran at roughly 2.1 times the pace ^[3], which spends the whole envelope in under six quarters ^[4]. "At least" was doing more work in that sentence than it appeared to be.

The composition is the part worth arguing about. Alibaba's own explanation names two causes at once and separates neither ^[4]. The year-over-year increase alone is close to 29 billion yuan of additional spend ^[1], and nobody outside the company knows how much silicon it bought. Where capex climbs partly because inputs cost more, the line stops being a read on capacity, and anyone treating cloud capex as a proxy for compute added is wrong by whatever the price term is.

That is what makes Wu's aside about chips the load-bearing one. He said profitability could improve significantly if in-house designs displace commercially procured parts in the data centres ^[9]. Component prices are the one input Alibaba cannot negotiate; the mix of its own silicon is.

Tencent is running the same trade with a declared ceiling, capital spending at the low end of 10% of revenue ^[10], and a return argument that does not depend on a hit product: Martin Lau said the company can turn a profit simply by leasing computing resources out, and that protection against losses is clear ^[11]. Both firms still sent about 58.5 billion yuan out the door on a free cash flow basis in the quarter ^[6]. Baidu and Kuaishou, at 6 billion to 8 billion yuan a quarter, are spending a tenth of that ^[14], and ByteDance is assumed comparable only because it never has to say ^[15]. A shortage that runs to the end of the decade does not raise costs evenly across the field; it settles which of these buyers is still bidding in 2028 ^[16].