A new ranking from Chinese research firm Ifenxi puts hard, if imperfect, numbers on something the industry has mostly discussed anecdotally: how much money Chinese AI vendors actually make selling tokens. The firm annualized each vendor's August 2026 token revenue and ranked the nine companies with meaningful API businesses.
ByteDance leads at roughly 30 billion yuan a year through its Volcano Engine platform, about double Alibaba's 16 billion yuan and three times Zhipu's 11 billion. The next tier runs DeepSeek at 6 billion yuan, Moonshot's Kimi at 5.8 billion, MiniMax at 4.8 billion and Kuaishou's Kling video business at 4 billion. Tencent sits at 3.2 billion and Baidu at 2.5 billion. Together the nine annualize to about 83.3 billion yuan, or roughly $11.7 billion at the firm's 7.1 yuan-per-dollar convention.
The distribution is steeply tiered. Three vendors clear 10 billion yuan; the middle four cluster between 4 billion and 6 billion; Tencent and Baidu trail at the bottom. Ifenxi notes that the gap between first and last place exceeds a factor of ten, and that each tier has a visible gap to the one above it.
How each vendor earns its place differs, and that is where the ranking gets more useful than a leaderboard. Alibaba is described as the only player that monetizes across all four layers of chips, cloud, models and applications. Zhipu's climb is attributed to AI coding, where token-heavy agent workloads drive consumption. DeepSeek's revenue rests partly on API price increases, Kimi's on a single breakout model, and MiniMax and Kling on video generation sold into overseas markets.
The China numbers sit inside a much larger global market. Ifenxi estimates worldwide token revenue at $135 billion for 2026, up more than 460% from $24 billion in 2025, a jump it attributes to mature open-weight ecosystems driving inference demand and to a new layer of third-party inference providers that did not exist at scale a year ago.
The caveats are real. This is a single research firm's estimate, built on a defined methodology of annualizing one month and converting at a fixed exchange rate. The firm says it cross-checked figures against multiple sources, but nothing here is audited, and token revenue is not profit: the report itself flags margin pressure among third-party inference providers, where competing on price compresses the very revenue being counted.
What the exercise captures is a market becoming measurable. Twelve months ago, Chinese labs were competing on benchmark scores; now they are competing on annualized consumption, and the shape of that leaderboard, ByteDance far ahead, Alibaba structurally integrated, Zhipu riding coding agents, Baidu visibly behind, looks a lot like the concentration pattern that formed first in the United States.
Comments (0)
Log in to join the discussion
Log InNo comments yet