Alibaba (BABA), Kingsoft Cloud (KC): Owned Chips Decide China AI Cloud Margins

Between 12 and 20 August 2026, Tencent Holdings (TCEHY), Kingsoft Cloud (KC) and Alibaba (BABA) each told quarterly earnings calls that whether a cloud operator owns cheap compute now decides whether selling AI model services makes money [1][2][3].

Model services are billed per call, and the cost is almost all one accelerator card

Chinese cloud operators have added a business line called model-as-a-service: rather than running a large model themselves, customers buy the model's output by usage, priced in tokens — the smallest pieces of text a model processes. Each call occupies an AI accelerator card, so the purchase price of that card sets almost the entire cost of the business.

Commercial accelerators have been in short supply in China. Alibaba's management cited an industry consensus that the shortage will not clear before 2030 at the earliest [3], and scarcity puts a premium into the price of a card. That splits the operators selling model services into three positions. Those that design their own chips do not pay the premium, and Alibaba and Baidu (BIDU) sit in that group [4]. Those that buy commercial cards to build their own fleets do pay it. Those that only resell someone else's compute, or only ship software, hand that profit to the layer above them [2]. Gross margin, and how long a card takes to pay for itself, differ by position.

A three-year payback is the first such number a company has shown outsiders

Alibaba gave the full arithmetic on 20 August. At its current gross margin on AI products, a server fitted with its own chips pays back in about three years; as the proprietary-chip substitution rate rises, that payback shortens to two and a half years or less, at which point the business can grow at more than 40% and still hold positive cash flow [3]. Management also said large-scale deployment of its own chips lets it avoid the high price premium on commercial GPUs, and that as of early August those chips had served more than 650 customers [3].

Tencent addressed the margin end on 12 August. Domestic token prices are indeed low, it said, but the domestic cost of manufacturing a token is also extremely low — much lower than widely estimated outside the company — so the gross margin on model services for paying users is already comparable to Tencent Cloud's overall gross margin [1]. Kingsoft Cloud addressed attribution on 19 August: only companies with low-cost compute make money across the chain, and a party that ships software alone gives up much of the profit to its compute provider [2]. All three point to the same variable, the cost per unit of compute, rather than model quality or token pricing.

Whoever pays the premium carries the longer payback

Control moves from model capability to the purchase price of the fleet. The higher the proprietary-chip substitution rate, the faster the growth a given capital budget can support, and the more the business funds itself. An operator that buys commercial cards builds the premium into its own payback period. The reseller and software-only layers have no payback to calculate at all, and upstream price increases reach them as cost or as a cut budget.

Two boundaries belong with that. Tencent can run the same arithmetic and chose differently: it said renting compute out today would recover depreciation almost immediately, but it is still allocating a large share of new compute to training its own models [1], so the payback is a management choice rather than an automatic outcome. Kingsoft Cloud said the upstream price increase has made state-owned enterprise and government customers repeatedly re-cut their budgets and delay decisions [2], so the same increase squeezes the buyer side. What to watch next is the proprietary-chip substitution rate, cloud gross margin, and whether a second company is willing to publish a payback period.

Companies exposed to this change

• Cambricon (688256.SH): It sells domestic AI accelerators, and a cloud operator that does not design its own chips can only escape the premium on commercial cards by buying this kind of part on the open market, which puts Cambricon on the supply side of that substitution path.
• Beijing Sinnet Technology (300383.SZ): It operates Amazon's cloud computing business inside China under licence [5], selling compute it does not own, which is the position that hands profit to the layer above.
• Kingsoft Office (688111.SH): It ships AI features in WPS without owning any compute, so every call is paid to a fleet owner, and the heavier the AI usage the more of that bill can flow out of its software gross margin.

Sources

[1] Drillr · Tencent Holdings (TCEHY) · 2026-08-12 · earnings call

[2] Drillr · Kingsoft Cloud (KC) · 2026-08-19 · earnings call

[3] Drillr · Alibaba Group (BABA) · 2026-08-20 · earnings call

"At our current level, of gross margin for AI products and under the assumption of a three-year payback period on CapEx. Theoretically, keeping our growth rate below 33% would already enable positive cash flow. However, that is not our strategic choice at this time. Given that AI remains in a very early stage, we're committed to aggressively investing in CapEx and proactively scaling up to drive our rapid business expansions. As our product gross margin improves and our proprietary chip substitution rate increases, our payback period will shorten to two and a half years or even less. And so under those circumstances, while pursuing growth of over 40%, we'll also be able to maintain positive cash flow."

[4] Drillr · Baidu (BIDU) · 2026-08-18 · earnings call

[5] Drillr · Beijing Sinnet Technology (300383.SZ) · 2026-08-21 · company relationship data

This is only here to help you spot industry changes and companies you may have missed - it is not a stock recommendation.