BeijingByteDance cut 30% of its AI products.
China’s ByteDance AI inference reportedly hit RMB 8 billion in 2025 — 2.3× incremental revenue — because every query incurs compute costs unlike mobile.
Tencent’s founder declared it had “switched ships”; analysts expect Alibaba and Baidu to consolidate AI portfolios next.
Sources: Forbes
WHY AI IS NOT MOBILE
A mobile app's marginal cost per user approaches zero — servers scale, ads pay. Generative AI inverts this: every query burns GPU cycles, electricity, and cooling. Revenue scales linearly with usage; so do costs. The flywheel that made TikTok print money does not exist here.
THE CONSOLIDATION PATTERN
Chinese internet giants ran a "hundred-model war" through 2023-24, with every major platform shipping its own LLM. The economics forced a cull: keep one flagship, sunset the rest. Tencent's Pony Ma signaled the pivot first; ByteDance's 30% product cut is the same logic executed at scale.
THE CHIP CONSTRAINT
US export controls since 2022 have blocked Nvidia's top-tier H100 and H200 chips from China. Domestic alternatives — Huawei Ascend, Cambricon — exist but trail in performance per watt. Inference at scale costs more in China than in the US for the same workload, which compresses margins before any pricing decision.
THE TENCENT PRECEDENT
Pony Ma's "switched ships" remark echoes Tencent's 2011 pivot to mobile, when it killed PC-era products to bet WeChat. That ruthlessness — sunset what doesn't compound, even if it's working — is the playbook. The cost of running an underperforming AI product is now high enough that keeping it is a strategic mistake.
THE GLOBAL ECHO
This is not a China-specific problem. OpenAI, Anthropic, and Google all run inference at a loss on their consumer tiers; the bet is that model efficiency improves faster than usage grows. China's consolidation is the same equation forced earlier by tighter capital and a chip ceiling — a preview of what happens when the subsidy ends.