Viral Wire

ByteDance cuts 30% of AI projects as costs outpace revenue 2.3x

The mobile internet playbook of free user growth is breaking in the AI era, as each query carries a real compute cost—ByteDance's 30% project cut is the first major signal of a structural reset.

Deep Dive

ByteDance, the private giant behind TikTok and Douyin, has slashed 30% of its AI projects after inference costs soared to RMB 8 billion ($1.1 billion) in 2025, outpacing incremental AI revenue by 2.3x. The company is abandoning its 'spray-and-pray' strategy—launching dozens of experimental AI apps like Dreamina and various assistants—and moving its flagship Doubao assistant to a paid model. This move underscores a fundamental shift: the era of subsidizing AI for user growth is over, because each query incurs a non-zero compute cost that scales with usage, unlike the zero-marginal-cost economics of the mobile internet.

The landscape reveals that ByteDance is not alone in facing this tension, but its competitors have structural advantages. Baidu, with its ERNIE Bot, is monetizing advanced features after a free trial but can lean on decades of search and cloud revenue to subsidize AI losses. Alibaba's Tongyi Qianwen focuses on enterprise customers, avoiding the race for consumer scale and emphasizing ROI from the start. Tencent integrates Hunyuan AI into its WeChat ecosystem, where stickiness and diversified revenue streams absorb AI costs. ByteDance, lacking such legacy cash cows in China, is most exposed—its AI products are expected to be standalone growth drivers, not enhancements to existing profitable businesses.

The implications extend beyond ByteDance. This is not a temporary cost overrun but a structural unit-economics problem for the entire AI industry. The hidden risks are multi-layered: U.S. export controls on advanced chips (like Nvidia H100) force Chinese firms to rely on less efficient domestic alternatives, inflating per-query costs. Additionally, cutting 30% of projects eliminates potential 'dark horses' that could have become profitable later—a ruthless efficiency that may curb long-term innovation. And the underlying infrastructure costs—power, cooling, maintenance—will continue to rise, eroding any gains from paid subscriptions. The move to a paid model for Doubao risks driving users to free alternatives like Baidu or iFlytek, creating a prisoner's dilemma where no player can sustainably go paid alone.

The bottom line: ByteDance's cut is a wake-up call that the AI industry must move from scale-at-all-costs to sustainable economics. Companies that survive will be those with proprietary low-cost inference—through efficient model architectures or custom silicon—or those that embed AI into already-profitable products rather than standalone apps. The Chinese AI market, estimated at $3-5 billion in 2025, will see a consolidation wave as the spray-and-pray era ends.

Key Points
  • Inference costs are structural, not temporary: each query loses money until per-token economics improve, forcing a shift from free growth to paid models or subsidized integrations.
  • Paid models risk user churn but free is unsustainable; ByteDance's Doubao monetization is a high-stakes test for the entire Chinese consumer AI sector.
  • US export controls on advanced chips amplify the cost problem for Chinese firms by forcing reliance on less efficient domestic hardware, a hidden factor that will continue to pressure margins.

Why It Matters

ByteDance's 30% AI project cut signals the end of the user-subsidization era and the beginning of a hard reset in AI's unit economics.