ByteDance pushes full-stack AI: Doubao, Seedance, own chips

The scale ByteDance is now trying to hold together in AI is the interesting part, not any one model release. Its Doubao chatbot reached 345 million monthly active users by March 2026, well ahead of Alibaba's Qwen at 166 million and DeepSeek at 127 million, per figures picked up in Financial Times reporting on the company. Sitting behind that distribution is a foundation model group called Seed, run since February 2025 by Wu Yonghui, a veteran researcher who spent 17 years at Google before crossing over.

The money makes the ambition more concrete. ByteDance has budgeted roughly 160 billion yuan (about $23 billion) for AI in 2026, and per Caixin's reporting more than half of that is earmarked for chip procurement, with additional pools for model upgrades and hiring. On the silicon side, ByteDance is reportedly aiming to mass-produce two in-house AI chips co-designed with TSMC in 2026, a move that would cut its exposure to Nvidia and to further U.S. export controls.

Why this matters if you follow the AI market outside China: no Western AI company runs the full stack under one roof the way ByteDance now does. OpenAI has models but rents distribution and buys compute. Google has models and compute but nothing at TikTok scale on the consumer side. ByteDance is trying to be all three simultaneously, and that vertical integration is the thing worth watching, not any single leaderboard result.

The honest caveat is that a lot of these specifics come from unnamed sources, and ByteDance has publicly denied some of the earlier TSMC chip reporting, so take the numbers as reported rather than settled. What the reporting does not give you is per-chip performance against Nvidia, or how much of that 160 billion yuan is still going to imported GPUs rather than the new domestic silicon. Those are the two data points that will actually tell you whether the vertical stack is real or aspirational.

If even part of it lands, the winners are enterprise buyers in China who get one vendor spanning models, video generation, and chips, plus domestic accelerator suppliers riding the shift away from Nvidia. The loser is the assumption that only U.S. firms can compound scale across the whole AI stack.